Rename:
[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 *
a015262c 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.33 2003/10/01 22:53:44 dillon Exp $
8ad65e08
MD
32 */
33
34#include <sys/param.h>
35#include <sys/systm.h>
36#include <sys/kernel.h>
37#include <sys/proc.h>
38#include <sys/rtprio.h>
39#include <sys/queue.h>
f1d1c3fa 40#include <sys/thread2.h>
7d0bac62 41#include <sys/sysctl.h>
99df837e 42#include <sys/kthread.h>
f1d1c3fa 43#include <machine/cpu.h>
99df837e 44#include <sys/lock.h>
f1d1c3fa 45
7d0bac62
MD
46#include <vm/vm.h>
47#include <vm/vm_param.h>
48#include <vm/vm_kern.h>
49#include <vm/vm_object.h>
50#include <vm/vm_page.h>
51#include <vm/vm_map.h>
52#include <vm/vm_pager.h>
53#include <vm/vm_extern.h>
54#include <vm/vm_zone.h>
55
99df837e 56#include <machine/stdarg.h>
57c254db 57#include <machine/ipl.h>
96728c05
MD
58#ifdef SMP
59#include <machine/smp.h>
60#endif
99df837e 61
7d0bac62 62static int untimely_switch = 0;
4b5f931b 63SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
57c254db
MD
64#ifdef INVARIANTS
65static int token_debug = 0;
66SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
67#endif
4b5f931b
MD
68static quad_t switch_count = 0;
69SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
70static quad_t preempt_hit = 0;
71SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
72static quad_t preempt_miss = 0;
73SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b
MD
74static quad_t preempt_weird = 0;
75SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05
MD
76static quad_t ipiq_count = 0;
77SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
78static quad_t ipiq_fifofull = 0;
79SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 80
4b5f931b
MD
81/*
82 * These helper procedures handle the runq, they can only be called from
83 * within a critical section.
84 */
f1d1c3fa
MD
85static __inline
86void
87_lwkt_dequeue(thread_t td)
88{
89 if (td->td_flags & TDF_RUNQ) {
4b5f931b
MD
90 int nq = td->td_pri & TDPRI_MASK;
91 struct globaldata *gd = mycpu;
92
f1d1c3fa 93 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
94 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
95 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
96 }
97}
98
99static __inline
100void
101_lwkt_enqueue(thread_t td)
102{
103 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b
MD
104 int nq = td->td_pri & TDPRI_MASK;
105 struct globaldata *gd = mycpu;
106
f1d1c3fa 107 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
108 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
109 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
110 }
111}
8ad65e08 112
57c254db
MD
113static __inline
114int
115_lwkt_wantresched(thread_t ntd, thread_t cur)
116{
117 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
118}
119
8ad65e08
MD
120/*
121 * LWKTs operate on a per-cpu basis
122 *
73e4f7b9 123 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
124 */
125void
126lwkt_gdinit(struct globaldata *gd)
127{
4b5f931b
MD
128 int i;
129
130 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
131 TAILQ_INIT(&gd->gd_tdrunq[i]);
132 gd->gd_runqmask = 0;
73e4f7b9 133 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
134}
135
7d0bac62
MD
136/*
137 * Initialize a thread wait structure prior to first use.
138 *
139 * NOTE! called from low level boot code, we cannot do anything fancy!
140 */
141void
142lwkt_init_wait(lwkt_wait_t w)
143{
144 TAILQ_INIT(&w->wa_waitq);
145}
146
147/*
148 * Create a new thread. The thread must be associated with a process context
149 * or LWKT start address before it can be scheduled.
0cfcada1
MD
150 *
151 * If you intend to create a thread without a process context this function
152 * does everything except load the startup and switcher function.
7d0bac62
MD
153 */
154thread_t
ef0fdad1 155lwkt_alloc_thread(struct thread *td)
7d0bac62 156{
99df837e 157 void *stack;
ef0fdad1 158 int flags = 0;
7d0bac62 159
ef0fdad1 160 if (td == NULL) {
26a0694b 161 crit_enter();
ef0fdad1
MD
162 if (mycpu->gd_tdfreecount > 0) {
163 --mycpu->gd_tdfreecount;
164 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 165 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
166 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
167 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
168 crit_exit();
169 stack = td->td_kstack;
170 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
171 } else {
172 crit_exit();
173 td = zalloc(thread_zone);
174 td->td_kstack = NULL;
175 flags |= TDF_ALLOCATED_THREAD;
176 }
177 }
178 if ((stack = td->td_kstack) == NULL) {
99df837e 179 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 180 flags |= TDF_ALLOCATED_STACK;
99df837e 181 }
26a0694b 182 lwkt_init_thread(td, stack, flags, mycpu);
99df837e 183 return(td);
7d0bac62
MD
184}
185
186/*
187 * Initialize a preexisting thread structure. This function is used by
188 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
189 *
190 * NOTE! called from low level boot code, we cannot do anything fancy!
a72187e9 191 * Only the low level boot code will call this function with gd != mycpu.
f8c3996b
MD
192 *
193 * All threads start out in a critical section at a priority of
194 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
195 * appropriate.
7d0bac62
MD
196 */
197void
26a0694b 198lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 199{
99df837e
MD
200 bzero(td, sizeof(struct thread));
201 td->td_kstack = stack;
202 td->td_flags |= flags;
26a0694b 203 td->td_gd = gd;
f8c3996b 204 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
ece04fd0 205 lwkt_init_port(&td->td_msgport, td);
99df837e 206 pmap_init_thread(td);
73e4f7b9 207 crit_enter();
a72187e9 208 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
73e4f7b9
MD
209 crit_exit();
210}
211
212void
213lwkt_set_comm(thread_t td, const char *ctl, ...)
214{
215 va_list va;
216
217 va_start(va, ctl);
218 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
219 va_end(va);
7d0bac62
MD
220}
221
99df837e 222void
73e4f7b9 223lwkt_hold(thread_t td)
99df837e 224{
73e4f7b9
MD
225 ++td->td_refs;
226}
227
228void
229lwkt_rele(thread_t td)
230{
231 KKASSERT(td->td_refs > 0);
232 --td->td_refs;
233}
234
235void
236lwkt_wait_free(thread_t td)
237{
238 while (td->td_refs)
377d4740 239 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
240}
241
242void
243lwkt_free_thread(thread_t td)
244{
245 struct globaldata *gd = mycpu;
246
d9eea1a5 247 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
248 ("lwkt_free_thread: did not exit! %p", td));
249
250 crit_enter();
73e4f7b9
MD
251 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
252 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
253 (td->td_flags & TDF_ALLOCATED_THREAD)
254 ) {
73e4f7b9
MD
255 ++gd->gd_tdfreecount;
256 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
257 crit_exit();
258 } else {
259 crit_exit();
260 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
261 kmem_free(kernel_map,
262 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 263 /* gd invalid */
99df837e
MD
264 td->td_kstack = NULL;
265 }
266 if (td->td_flags & TDF_ALLOCATED_THREAD)
267 zfree(thread_zone, td);
268 }
269}
270
271
8ad65e08
MD
272/*
273 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
274 * switch to the idlethread. Switching must occur within a critical
275 * section to avoid races with the scheduling queue.
276 *
277 * We always have full control over our cpu's run queue. Other cpus
278 * that wish to manipulate our queue must use the cpu_*msg() calls to
279 * talk to our cpu, so a critical section is all that is needed and
280 * the result is very, very fast thread switching.
281 *
96728c05
MD
282 * The LWKT scheduler uses a fixed priority model and round-robins at
283 * each priority level. User process scheduling is a totally
284 * different beast and LWKT priorities should not be confused with
285 * user process priorities.
f1d1c3fa 286 *
96728c05
MD
287 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
288 * cleans it up. Note that the td_switch() function cannot do anything that
289 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
290 * the target thread (not the current thread). It's nice to have a scheduler
291 * that does not need the MP lock to work because it allows us to do some
292 * really cool high-performance MP lock optimizations.
8ad65e08 293 */
96728c05 294
8ad65e08
MD
295void
296lwkt_switch(void)
297{
4b5f931b 298 struct globaldata *gd;
f1d1c3fa 299 thread_t td = curthread;
8ad65e08 300 thread_t ntd;
8a8d5d85
MD
301#ifdef SMP
302 int mpheld;
303#endif
8ad65e08 304
96728c05
MD
305 if (mycpu->gd_intr_nesting_level &&
306 td->td_preempted == NULL && panicstr == NULL
307 ) {
03aa8d99 308 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 309 }
ef0fdad1 310
cb973d15
MD
311 /*
312 * Passive release (used to transition from user to kernel mode
313 * when we block or switch rather then when we enter the kernel).
314 * This function is NOT called if we are switching into a preemption
315 * or returning from a preemption. Typically this causes us to lose
316 * our P_CURPROC designation (if we have one) and become a true LWKT
317 * thread, and may also hand P_CURPROC to another process and schedule
318 * its thread.
319 */
320 if (td->td_release)
321 td->td_release(td);
322
f1d1c3fa 323 crit_enter();
4b5f931b 324 ++switch_count;
8a8d5d85
MD
325
326#ifdef SMP
327 /*
328 * td_mpcount cannot be used to determine if we currently hold the
329 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
330 * to get the lock, and switch out if it can't. Our ownership of
331 * the actual lock will remain stable while we are in a critical section
332 * (but, of course, another cpu may own or release the lock so the
333 * actual value of mp_lock is not stable).
8a8d5d85
MD
334 */
335 mpheld = MP_LOCK_HELD();
336#endif
99df837e
MD
337 if ((ntd = td->td_preempted) != NULL) {
338 /*
339 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
340 * thread. This occurs transparently, whether the preempted thread
341 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
342 * itself).
343 *
344 * We have to setup the MP lock for the original thread after backing
345 * out the adjustment that was made to curthread when the original
346 * was preempted.
99df837e 347 */
26a0694b 348 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 349#ifdef SMP
96728c05
MD
350 if (ntd->td_mpcount && mpheld == 0) {
351 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
352 td, ntd, td->td_mpcount, ntd->td_mpcount);
353 }
8a8d5d85
MD
354 if (ntd->td_mpcount) {
355 td->td_mpcount -= ntd->td_mpcount;
356 KKASSERT(td->td_mpcount >= 0);
357 }
358#endif
26a0694b 359 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 360 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 361 } else {
4b5f931b
MD
362 /*
363 * Priority queue / round-robin at each priority. Note that user
364 * processes run at a fixed, low priority and the user process
365 * scheduler deals with interactions between user processes
366 * by scheduling and descheduling them from the LWKT queue as
367 * necessary.
8a8d5d85
MD
368 *
369 * We have to adjust the MP lock for the target thread. If we
370 * need the MP lock and cannot obtain it we try to locate a
371 * thread that does not need the MP lock.
4b5f931b
MD
372 */
373 gd = mycpu;
4b5f931b
MD
374again:
375 if (gd->gd_runqmask) {
376 int nq = bsrl(gd->gd_runqmask);
377 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
378 gd->gd_runqmask &= ~(1 << nq);
379 goto again;
380 }
8a8d5d85
MD
381#ifdef SMP
382 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
383 /*
96728c05
MD
384 * Target needs MP lock and we couldn't get it, try
385 * to locate a thread which does not need the MP lock
3c23a41a 386 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
387 */
388 u_int32_t rqmask = gd->gd_runqmask;
389 while (rqmask) {
390 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
391 if (ntd->td_mpcount == 0)
392 break;
393 }
394 if (ntd)
395 break;
396 rqmask &= ~(1 << nq);
397 nq = bsrl(rqmask);
398 }
399 if (ntd == NULL) {
a2a5ad0d
MD
400 ntd = &gd->gd_idlethread;
401 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
402 } else {
403 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
404 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
405 }
406 } else {
407 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
408 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
409 }
410#else
4b5f931b
MD
411 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
412 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 413#endif
4b5f931b 414 } else {
3c23a41a
MD
415 /*
416 * Nothing to run but we may still need the BGL to deal with
417 * pending interrupts, spin in idle if so.
418 */
a2a5ad0d 419 ntd = &gd->gd_idlethread;
235957ed 420 if (gd->gd_reqflags)
3c23a41a 421 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 422 }
f1d1c3fa 423 }
26a0694b
MD
424 KASSERT(ntd->td_pri >= TDPRI_CRIT,
425 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
426
427 /*
428 * Do the actual switch. If the new target does not need the MP lock
429 * and we are holding it, release the MP lock. If the new target requires
430 * the MP lock we have already acquired it for the target.
431 */
432#ifdef SMP
433 if (ntd->td_mpcount == 0 ) {
434 if (MP_LOCK_HELD())
435 cpu_rel_mplock();
436 } else {
437 ASSERT_MP_LOCK_HELD();
438 }
439#endif
8a8d5d85 440 if (td != ntd) {
f1d1c3fa 441 td->td_switch(ntd);
8a8d5d85 442 }
96728c05 443
f1d1c3fa 444 crit_exit();
8ad65e08
MD
445}
446
cb973d15
MD
447/*
448 * Switch if another thread has a higher priority. Do not switch to other
449 * threads at the same priority.
450 */
451void
452lwkt_maybe_switch()
453{
454 struct globaldata *gd = mycpu;
455 struct thread *td = gd->gd_curthread;
456
457 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
458 lwkt_switch();
459 }
460}
461
b68b7282 462/*
96728c05
MD
463 * Request that the target thread preempt the current thread. Preemption
464 * only works under a specific set of conditions:
b68b7282 465 *
96728c05
MD
466 * - We are not preempting ourselves
467 * - The target thread is owned by the current cpu
468 * - We are not currently being preempted
469 * - The target is not currently being preempted
470 * - We are able to satisfy the target's MP lock requirements (if any).
471 *
472 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
473 * this is called via lwkt_schedule() through the td_preemptable callback.
474 * critpri is the managed critical priority that we should ignore in order
475 * to determine whether preemption is possible (aka usually just the crit
476 * priority of lwkt_schedule() itself).
b68b7282 477 *
26a0694b
MD
478 * XXX at the moment we run the target thread in a critical section during
479 * the preemption in order to prevent the target from taking interrupts
480 * that *WE* can't. Preemption is strictly limited to interrupt threads
481 * and interrupt-like threads, outside of a critical section, and the
482 * preempted source thread will be resumed the instant the target blocks
483 * whether or not the source is scheduled (i.e. preemption is supposed to
484 * be as transparent as possible).
4b5f931b 485 *
8a8d5d85
MD
486 * The target thread inherits our MP count (added to its own) for the
487 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
488 * MP lock during the preemption. Therefore, any preempting targets must be
489 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
490 * out of sync with the physical mp_lock, but we do not have to preserve
491 * the original ownership of the lock if it was out of synch (that is, we
492 * can leave it synchronized on return).
b68b7282
MD
493 */
494void
96728c05 495lwkt_preempt(thread_t ntd, int critpri)
b68b7282 496{
73e4f7b9 497 thread_t td = curthread;
8a8d5d85
MD
498#ifdef SMP
499 int mpheld;
57c254db 500 int savecnt;
8a8d5d85 501#endif
b68b7282 502
26a0694b 503 /*
96728c05
MD
504 * The caller has put us in a critical section. We can only preempt
505 * if the caller of the caller was not in a critical section (basically
57c254db
MD
506 * a local interrupt), as determined by the 'critpri' parameter. If
507 * we are unable to preempt
96728c05
MD
508 *
509 * YYY The target thread must be in a critical section (else it must
510 * inherit our critical section? I dunno yet).
26a0694b
MD
511 */
512 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 513
cb973d15 514 need_resched();
57c254db
MD
515 if (!_lwkt_wantresched(ntd, td)) {
516 ++preempt_miss;
517 return;
518 }
96728c05
MD
519 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
520 ++preempt_miss;
521 return;
522 }
523#ifdef SMP
a72187e9 524 if (ntd->td_gd != mycpu) {
96728c05
MD
525 ++preempt_miss;
526 return;
527 }
528#endif
26a0694b
MD
529 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
530 ++preempt_weird;
531 return;
532 }
533 if (ntd->td_preempted) {
4b5f931b 534 ++preempt_hit;
26a0694b 535 return;
b68b7282 536 }
8a8d5d85 537#ifdef SMP
a2a5ad0d
MD
538 /*
539 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
540 * to or from the MP lock. In this case td_mpcount will be pre-disposed
541 * (non-zero) but not actually synchronized with the actual state of the
542 * lock. We can use it to imply an MP lock requirement for the
543 * preemption but we cannot use it to test whether we hold the MP lock
544 * or not.
a2a5ad0d 545 */
96728c05 546 savecnt = td->td_mpcount;
71ef2f5c 547 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
548 ntd->td_mpcount += td->td_mpcount;
549 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
550 ntd->td_mpcount -= td->td_mpcount;
551 ++preempt_miss;
552 return;
553 }
554#endif
26a0694b
MD
555
556 ++preempt_hit;
557 ntd->td_preempted = td;
558 td->td_flags |= TDF_PREEMPT_LOCK;
559 td->td_switch(ntd);
560 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
561#ifdef SMP
562 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
563 mpheld = MP_LOCK_HELD();
564 if (mpheld && td->td_mpcount == 0)
96728c05 565 cpu_rel_mplock();
71ef2f5c 566 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
567 panic("lwkt_preempt(): MP lock was not held through");
568#endif
26a0694b
MD
569 ntd->td_preempted = NULL;
570 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
571}
572
f1d1c3fa
MD
573/*
574 * Yield our thread while higher priority threads are pending. This is
575 * typically called when we leave a critical section but it can be safely
576 * called while we are in a critical section.
577 *
578 * This function will not generally yield to equal priority threads but it
579 * can occur as a side effect. Note that lwkt_switch() is called from
580 * inside the critical section to pervent its own crit_exit() from reentering
581 * lwkt_yield_quick().
582 *
235957ed 583 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
584 * came along but was blocked and made pending.
585 *
f1d1c3fa
MD
586 * (self contained on a per cpu basis)
587 */
588void
589lwkt_yield_quick(void)
590{
7966cb69
MD
591 globaldata_t gd = mycpu;
592 thread_t td = gd->gd_curthread;
ef0fdad1 593
a2a5ad0d 594 /*
235957ed 595 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
596 * it with a non-zero cpl then we might not wind up calling splz after
597 * a task switch when the critical section is exited even though the
598 * new task could accept the interrupt. YYY alternative is to have
599 * lwkt_switch() just call splz unconditionally.
600 *
601 * XXX from crit_exit() only called after last crit section is released.
602 * If called directly will run splz() even if in a critical section.
603 */
7966cb69 604 if (gd->gd_reqflags)
f1d1c3fa 605 splz();
f1d1c3fa
MD
606
607 /*
608 * YYY enabling will cause wakeup() to task-switch, which really
609 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
610 * preemption and MP without actually doing preemption or MP, because a
611 * lot of code assumes that wakeup() does not block.
f1d1c3fa 612 */
7966cb69 613 if (untimely_switch && gd->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
614 crit_enter();
615 /*
616 * YYY temporary hacks until we disassociate the userland scheduler
617 * from the LWKT scheduler.
618 */
619 if (td->td_flags & TDF_RUNQ) {
620 lwkt_switch(); /* will not reenter yield function */
621 } else {
622 lwkt_schedule_self(); /* make sure we are scheduled */
623 lwkt_switch(); /* will not reenter yield function */
624 lwkt_deschedule_self(); /* make sure we are descheduled */
625 }
7966cb69 626 crit_exit_noyield(td);
f1d1c3fa 627 }
f1d1c3fa
MD
628}
629
8ad65e08 630/*
f1d1c3fa 631 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 632 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
633 * the crit_exit() call in lwkt_switch().
634 *
635 * (self contained on a per cpu basis)
8ad65e08
MD
636 */
637void
f1d1c3fa 638lwkt_yield(void)
8ad65e08 639{
f1d1c3fa
MD
640 lwkt_schedule_self();
641 lwkt_switch();
642}
643
644/*
645 * Schedule a thread to run. As the current thread we can always safely
646 * schedule ourselves, and a shortcut procedure is provided for that
647 * function.
648 *
649 * (non-blocking, self contained on a per cpu basis)
650 */
651void
652lwkt_schedule_self(void)
653{
654 thread_t td = curthread;
655
656 crit_enter();
657 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 658 _lwkt_enqueue(td);
26a0694b
MD
659 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
660 panic("SCHED SELF PANIC");
f1d1c3fa 661 crit_exit();
8ad65e08 662}
8ad65e08
MD
663
664/*
f1d1c3fa
MD
665 * Generic schedule. Possibly schedule threads belonging to other cpus and
666 * deal with threads that might be blocked on a wait queue.
667 *
96728c05 668 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
669 * Load balancing can be accomplished by requesting other sorts of actions
670 * for the thread in question.
8ad65e08
MD
671 */
672void
673lwkt_schedule(thread_t td)
674{
96728c05 675#ifdef INVARIANTS
26a0694b
MD
676 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
677 && td->td_proc->p_stat == SSLEEP
678 ) {
679 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
680 curthread,
681 curthread->td_proc ? curthread->td_proc->p_pid : -1,
682 curthread->td_proc ? curthread->td_proc->p_stat : -1,
683 td,
684 td->td_proc ? curthread->td_proc->p_pid : -1,
685 td->td_proc ? curthread->td_proc->p_stat : -1
686 );
687 panic("SCHED PANIC");
688 }
96728c05 689#endif
f1d1c3fa
MD
690 crit_enter();
691 if (td == curthread) {
692 _lwkt_enqueue(td);
693 } else {
694 lwkt_wait_t w;
695
696 /*
697 * If the thread is on a wait list we have to send our scheduling
698 * request to the owner of the wait structure. Otherwise we send
699 * the scheduling request to the cpu owning the thread. Races
700 * are ok, the target will forward the message as necessary (the
701 * message may chase the thread around before it finally gets
702 * acted upon).
703 *
704 * (remember, wait structures use stable storage)
705 */
706 if ((w = td->td_wait) != NULL) {
96728c05 707 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
708 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
709 --w->wa_count;
710 td->td_wait = NULL;
a72187e9 711 if (td->td_gd == mycpu) {
f1d1c3fa 712 _lwkt_enqueue(td);
57c254db 713 if (td->td_preemptable) {
96728c05 714 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
715 } else if (_lwkt_wantresched(td, curthread)) {
716 need_resched();
717 }
f1d1c3fa 718 } else {
a72187e9 719 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 720 }
96728c05 721 lwkt_reltoken(&w->wa_token);
f1d1c3fa 722 } else {
96728c05 723 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
724 }
725 } else {
726 /*
727 * If the wait structure is NULL and we own the thread, there
728 * is no race (since we are in a critical section). If we
729 * do not own the thread there might be a race but the
730 * target cpu will deal with it.
731 */
a72187e9 732 if (td->td_gd == mycpu) {
f1d1c3fa 733 _lwkt_enqueue(td);
57c254db 734 if (td->td_preemptable) {
96728c05 735 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
736 } else if (_lwkt_wantresched(td, curthread)) {
737 need_resched();
738 }
f1d1c3fa 739 } else {
a72187e9 740 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
741 }
742 }
8ad65e08 743 }
f1d1c3fa 744 crit_exit();
8ad65e08
MD
745}
746
d9eea1a5
MD
747/*
748 * Managed acquisition. This code assumes that the MP lock is held for
749 * the tdallq operation and that the thread has been descheduled from its
750 * original cpu. We also have to wait for the thread to be entirely switched
751 * out on its original cpu (this is usually fast enough that we never loop)
752 * since the LWKT system does not have to hold the MP lock while switching
753 * and the target may have released it before switching.
754 */
a2a5ad0d
MD
755void
756lwkt_acquire(thread_t td)
757{
758 struct globaldata *gd;
759
760 gd = td->td_gd;
761 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
762 while (td->td_flags & TDF_RUNNING) /* XXX spin */
763 ;
a2a5ad0d
MD
764 if (gd != mycpu) {
765 crit_enter();
766 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
767 gd = mycpu;
768 td->td_gd = gd;
a2a5ad0d
MD
769 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
770 crit_exit();
771 }
772}
773
8ad65e08 774/*
f1d1c3fa
MD
775 * Deschedule a thread.
776 *
777 * (non-blocking, self contained on a per cpu basis)
778 */
779void
780lwkt_deschedule_self(void)
781{
782 thread_t td = curthread;
783
784 crit_enter();
785 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
786 _lwkt_dequeue(td);
787 crit_exit();
788}
789
790/*
791 * Generic deschedule. Descheduling threads other then your own should be
792 * done only in carefully controlled circumstances. Descheduling is
793 * asynchronous.
794 *
795 * This function may block if the cpu has run out of messages.
8ad65e08
MD
796 */
797void
798lwkt_deschedule(thread_t td)
799{
f1d1c3fa
MD
800 crit_enter();
801 if (td == curthread) {
802 _lwkt_dequeue(td);
803 } else {
a72187e9 804 if (td->td_gd == mycpu) {
f1d1c3fa
MD
805 _lwkt_dequeue(td);
806 } else {
a72187e9 807 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
808 }
809 }
810 crit_exit();
811}
812
4b5f931b
MD
813/*
814 * Set the target thread's priority. This routine does not automatically
815 * switch to a higher priority thread, LWKT threads are not designed for
816 * continuous priority changes. Yield if you want to switch.
817 *
818 * We have to retain the critical section count which uses the high bits
26a0694b
MD
819 * of the td_pri field. The specified priority may also indicate zero or
820 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
821 */
822void
823lwkt_setpri(thread_t td, int pri)
824{
26a0694b 825 KKASSERT(pri >= 0);
a72187e9 826 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
827 crit_enter();
828 if (td->td_flags & TDF_RUNQ) {
829 _lwkt_dequeue(td);
830 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
831 _lwkt_enqueue(td);
832 } else {
833 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
834 }
835 crit_exit();
836}
837
838void
839lwkt_setpri_self(int pri)
840{
841 thread_t td = curthread;
842
4b5f931b
MD
843 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
844 crit_enter();
845 if (td->td_flags & TDF_RUNQ) {
846 _lwkt_dequeue(td);
847 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
848 _lwkt_enqueue(td);
849 } else {
850 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
851 }
852 crit_exit();
853}
854
855struct proc *
856lwkt_preempted_proc(void)
857{
73e4f7b9 858 thread_t td = curthread;
4b5f931b
MD
859 while (td->td_preempted)
860 td = td->td_preempted;
861 return(td->td_proc);
862}
863
ece04fd0
MD
864typedef struct lwkt_gettoken_req {
865 lwkt_token_t tok;
866 int cpu;
867} lwkt_gettoken_req;
868
869#if 0
4b5f931b 870
f1d1c3fa
MD
871/*
872 * This function deschedules the current thread and blocks on the specified
873 * wait queue. We obtain ownership of the wait queue in order to block
874 * on it. A generation number is used to interlock the wait queue in case
875 * it gets signalled while we are blocked waiting on the token.
876 *
877 * Note: alternatively we could dequeue our thread and then message the
878 * target cpu owning the wait queue. YYY implement as sysctl.
879 *
880 * Note: wait queue signals normally ping-pong the cpu as an optimization.
881 */
96728c05 882
f1d1c3fa 883void
ae8050a4 884lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
885{
886 thread_t td = curthread;
f1d1c3fa 887
f1d1c3fa 888 lwkt_gettoken(&w->wa_token);
ae8050a4 889 if (w->wa_gen == *gen) {
f1d1c3fa
MD
890 _lwkt_dequeue(td);
891 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
892 ++w->wa_count;
893 td->td_wait = w;
ae8050a4 894 td->td_wmesg = wmesg;
ece04fd0 895again:
f1d1c3fa 896 lwkt_switch();
ece04fd0
MD
897 lwkt_regettoken(&w->wa_token);
898 if (td->td_wmesg != NULL) {
899 _lwkt_dequeue(td);
900 goto again;
901 }
8ad65e08 902 }
ae8050a4
MD
903 /* token might be lost, doesn't matter for gen update */
904 *gen = w->wa_gen;
f1d1c3fa
MD
905 lwkt_reltoken(&w->wa_token);
906}
907
908/*
909 * Signal a wait queue. We gain ownership of the wait queue in order to
910 * signal it. Once a thread is removed from the wait queue we have to
911 * deal with the cpu owning the thread.
912 *
913 * Note: alternatively we could message the target cpu owning the wait
914 * queue. YYY implement as sysctl.
915 */
916void
ece04fd0 917lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
918{
919 thread_t td;
920 int count;
921
922 lwkt_gettoken(&w->wa_token);
923 ++w->wa_gen;
ece04fd0
MD
924 if (count < 0)
925 count = w->wa_count;
f1d1c3fa
MD
926 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
927 --count;
928 --w->wa_count;
929 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
930 td->td_wait = NULL;
ae8050a4 931 td->td_wmesg = NULL;
a72187e9 932 if (td->td_gd == mycpu) {
f1d1c3fa
MD
933 _lwkt_enqueue(td);
934 } else {
a72187e9 935 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
936 }
937 lwkt_regettoken(&w->wa_token);
938 }
939 lwkt_reltoken(&w->wa_token);
940}
941
ece04fd0
MD
942#endif
943
f1d1c3fa 944/*
96728c05 945 * Acquire ownership of a token
f1d1c3fa 946 *
96728c05 947 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
948 * section side effects, depending on its purpose. These side effects
949 * guarentee that you will maintain ownership of the token as long as you
950 * do not block. If you block you may lose access to the token (but you
951 * must still release it even if you lose your access to it).
952 *
96728c05 953 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 954 * a token, but do we really only need to disable IPIs ?
96728c05
MD
955 *
956 * YYY certain tokens could be made to act like mutexes when performance
957 * would be better (e.g. t_cpu == -1). This is not yet implemented.
958 *
a2a5ad0d
MD
959 * YYY the tokens replace 4.x's simplelocks for the most part, but this
960 * means that 4.x does not expect a switch so for now we cannot switch
961 * when waiting for an IPI to be returned.
962 *
963 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
964 * it and then block. The IPI causes the token to be given away to the
965 * requesting cpu, unless it has already changed hands. Since only the
966 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 967 * This needs serious optimization.
f1d1c3fa 968 */
57c254db
MD
969
970#ifdef SMP
971
96728c05
MD
972static
973void
974lwkt_gettoken_remote(void *arg)
975{
976 lwkt_gettoken_req *req = arg;
977 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 978#ifdef INVARIANTS
a2a5ad0d
MD
979 if (token_debug)
980 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 981#endif
96728c05 982 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
983 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
984 /* else set reqcpu to point to current cpu for release */
96728c05
MD
985 }
986}
987
57c254db
MD
988#endif
989
8a8d5d85 990int
f1d1c3fa
MD
991lwkt_gettoken(lwkt_token_t tok)
992{
993 /*
994 * Prevent preemption so the token can't be taken away from us once
995 * we gain ownership of it. Use a synchronous request which might
996 * block. The request will be forwarded as necessary playing catchup
997 * to the token.
998 */
96728c05 999
f1d1c3fa 1000 crit_enter();
57c254db 1001#ifdef INVARIANTS
1d4c9574
MD
1002 if (curthread->td_pri > 1800) {
1003 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1004 tok, ((int **)&tok)[-1]);
1005 }
a2a5ad0d
MD
1006 if (curthread->td_pri > 2000) {
1007 curthread->td_pri = 1000;
1008 panic("too HIGH!");
57c254db
MD
1009 }
1010#endif
96728c05 1011#ifdef SMP
d0e06f83 1012 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1013 struct lwkt_gettoken_req req;
1014 int seq;
96728c05
MD
1015 int dcpu;
1016
1017 req.cpu = mycpu->gd_cpuid;
1018 req.tok = tok;
1019 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1020 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1021#ifdef INVARIANTS
a2a5ad0d
MD
1022 if (token_debug)
1023 printf("REQT%d ", dcpu);
634081ff 1024#endif
96728c05
MD
1025 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1026 lwkt_wait_ipiq(dcpu, seq);
634081ff 1027#ifdef INVARIANTS
a2a5ad0d
MD
1028 if (token_debug)
1029 printf("REQR%d ", tok->t_cpu);
634081ff 1030#endif
f1d1c3fa
MD
1031 }
1032#endif
1033 /*
1034 * leave us in a critical section on return. This will be undone
8a8d5d85 1035 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1036 */
8a8d5d85 1037 return(++tok->t_gen);
f1d1c3fa
MD
1038}
1039
96728c05
MD
1040/*
1041 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1042 * failure.
1043 */
1044int
1045lwkt_trytoken(lwkt_token_t tok)
1046{
1047 crit_enter();
1048#ifdef SMP
1049 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1050 crit_exit();
96728c05
MD
1051 return(0);
1052 }
1053#endif
1054 /* leave us in the critical section */
1055 ++tok->t_gen;
1056 return(1);
1057}
1058
f1d1c3fa
MD
1059/*
1060 * Release your ownership of a token. Releases must occur in reverse
1061 * order to aquisitions, eventually so priorities can be unwound properly
1062 * like SPLs. At the moment the actual implemention doesn't care.
1063 *
1064 * We can safely hand a token that we own to another cpu without notifying
1065 * it, but once we do we can't get it back without requesting it (unless
1066 * the other cpu hands it back to us before we check).
1067 *
1068 * We might have lost the token, so check that.
1069 */
1070void
1071lwkt_reltoken(lwkt_token_t tok)
1072{
d0e06f83 1073 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1074 tok->t_cpu = tok->t_reqcpu;
1075 }
1076 crit_exit();
1077}
1078
1079/*
8a8d5d85
MD
1080 * Reacquire a token that might have been lost and compare and update the
1081 * generation number. 0 is returned if the generation has not changed
1082 * (nobody else obtained the token while we were blocked, on this cpu or
1083 * any other cpu).
1084 *
1085 * This function returns with the token re-held whether the generation
1086 * number changed or not.
1087 */
1088int
1089lwkt_gentoken(lwkt_token_t tok, int *gen)
1090{
1091 if (lwkt_regettoken(tok) == *gen) {
1092 return(0);
1093 } else {
1094 *gen = tok->t_gen;
1095 return(-1);
1096 }
1097}
1098
1099
1100/*
96728c05 1101 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1102 * number of the token.
f1d1c3fa
MD
1103 */
1104int
1105lwkt_regettoken(lwkt_token_t tok)
1106{
96728c05 1107 /* assert we are in a critical section */
d0e06f83 1108 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1109#ifdef SMP
d0e06f83 1110 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1111 struct lwkt_gettoken_req req;
1112 int seq;
96728c05 1113 int dcpu;
57c254db 1114
96728c05
MD
1115 req.cpu = mycpu->gd_cpuid;
1116 req.tok = tok;
1117 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1118 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1119#ifdef INVARIANTS
cb973d15
MD
1120 if (token_debug)
1121 printf("REQT%d ", dcpu);
634081ff 1122#endif
96728c05
MD
1123 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1124 lwkt_wait_ipiq(dcpu, seq);
634081ff 1125#ifdef INVARIATNS
cb973d15
MD
1126 if (token_debug)
1127 printf("REQR%d ", tok->t_cpu);
634081ff 1128#endif
f1d1c3fa 1129 }
f1d1c3fa 1130#endif
96728c05
MD
1131 ++tok->t_gen;
1132 }
8a8d5d85 1133 return(tok->t_gen);
8ad65e08
MD
1134}
1135
72740893
MD
1136void
1137lwkt_inittoken(lwkt_token_t tok)
1138{
1139 /*
1140 * Zero structure and set cpu owner and reqcpu to cpu 0.
1141 */
1142 bzero(tok, sizeof(*tok));
1143}
1144
99df837e
MD
1145/*
1146 * Create a kernel process/thread/whatever. It shares it's address space
1147 * with proc0 - ie: kernel only.
1148 *
1149 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1150 *
1151 * The thread will be entered with the MP lock held.
99df837e
MD
1152 */
1153int
1154lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1155 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1156 const char *fmt, ...)
99df837e 1157{
73e4f7b9 1158 thread_t td;
99df837e
MD
1159 va_list ap;
1160
a2a5ad0d
MD
1161 td = lwkt_alloc_thread(template);
1162 if (tdp)
1163 *tdp = td;
99df837e 1164 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1165 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1166#ifdef SMP
1167 td->td_mpcount = 1;
1168#endif
99df837e
MD
1169
1170 /*
1171 * Set up arg0 for 'ps' etc
1172 */
1173 va_start(ap, fmt);
1174 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1175 va_end(ap);
1176
1177 /*
1178 * Schedule the thread to run
1179 */
ef0fdad1
MD
1180 if ((td->td_flags & TDF_STOPREQ) == 0)
1181 lwkt_schedule(td);
1182 else
1183 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1184 return 0;
1185}
1186
1187/*
1188 * Destroy an LWKT thread. Warning! This function is not called when
1189 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1190 * uses a different reaping mechanism.
1191 */
1192void
1193lwkt_exit(void)
1194{
1195 thread_t td = curthread;
1196
1197 if (td->td_flags & TDF_VERBOSE)
1198 printf("kthread %p %s has exited\n", td, td->td_comm);
1199 crit_enter();
1200 lwkt_deschedule_self();
1201 ++mycpu->gd_tdfreecount;
1202 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1203 cpu_thread_exit();
1204}
1205
1206/*
1207 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1208 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1209 */
1210int
1211kthread_create(void (*func)(void *), void *arg,
1212 struct thread **tdp, const char *fmt, ...)
1213{
73e4f7b9 1214 thread_t td;
99df837e
MD
1215 va_list ap;
1216
a2a5ad0d
MD
1217 td = lwkt_alloc_thread(NULL);
1218 if (tdp)
1219 *tdp = td;
99df837e
MD
1220 cpu_set_thread_handler(td, kthread_exit, func, arg);
1221 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1222#ifdef SMP
1223 td->td_mpcount = 1;
1224#endif
99df837e
MD
1225
1226 /*
1227 * Set up arg0 for 'ps' etc
1228 */
1229 va_start(ap, fmt);
1230 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1231 va_end(ap);
1232
1233 /*
1234 * Schedule the thread to run
1235 */
1236 lwkt_schedule(td);
1237 return 0;
1238}
1239
26a0694b
MD
1240void
1241crit_panic(void)
1242{
73e4f7b9 1243 thread_t td = curthread;
26a0694b
MD
1244 int lpri = td->td_pri;
1245
1246 td->td_pri = 0;
1247 panic("td_pri is/would-go negative! %p %d", td, lpri);
1248}
1249
99df837e
MD
1250/*
1251 * Destroy an LWKT thread. Warning! This function is not called when
1252 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1253 * uses a different reaping mechanism.
1254 *
1255 * XXX duplicates lwkt_exit()
1256 */
1257void
1258kthread_exit(void)
1259{
1260 lwkt_exit();
1261}
1262
96728c05
MD
1263#ifdef SMP
1264
1265/*
1266 * Send a function execution request to another cpu. The request is queued
1267 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1268 * possible target cpu. The FIFO can be written.
1269 *
1270 * YYY If the FIFO fills up we have to enable interrupts and process the
1271 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1272 * Create a CPU_*() function to do this!
1273 *
1274 * Must be called from a critical section.
1275 */
1276int
1277lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1278{
1279 lwkt_ipiq_t ip;
1280 int windex;
a2a5ad0d 1281 struct globaldata *gd = mycpu;
96728c05 1282
a2a5ad0d 1283 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1284 func(arg);
1285 return(0);
1286 }
cb973d15 1287 crit_enter();
a2a5ad0d
MD
1288 ++gd->gd_intr_nesting_level;
1289#ifdef INVARIANTS
1290 if (gd->gd_intr_nesting_level > 20)
1291 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1292#endif
96728c05
MD
1293 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1294 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1295 ++ipiq_count;
a2a5ad0d 1296 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1297
1298 /*
1299 * We always drain before the FIFO becomes full so it should never
1300 * become full. We need to leave enough entries to deal with
1301 * reentrancy.
1302 */
1303 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1304 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1305 ip->ip_func[windex] = func;
1306 ip->ip_arg[windex] = arg;
1307 /* YYY memory barrier */
1308 ++ip->ip_windex;
96728c05
MD
1309 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1310 unsigned int eflags = read_eflags();
1311 cpu_enable_intr();
1312 ++ipiq_fifofull;
cb973d15 1313 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1314 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1315 lwkt_process_ipiq();
1316 }
1317 write_eflags(eflags);
1318 }
a2a5ad0d 1319 --gd->gd_intr_nesting_level;
96728c05 1320 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1321 crit_exit();
96728c05
MD
1322 return(ip->ip_windex);
1323}
1324
cb973d15
MD
1325/*
1326 * Send a message to several target cpus. Typically used for scheduling.
1327 */
1328void
1329lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1330{
1331 int cpuid;
1332
1333 while (mask) {
1334 cpuid = bsfl(mask);
1335 lwkt_send_ipiq(cpuid, func, arg);
1336 mask &= ~(1 << cpuid);
1337 }
1338}
1339
96728c05
MD
1340/*
1341 * Wait for the remote cpu to finish processing a function.
1342 *
1343 * YYY we have to enable interrupts and process the IPIQ while waiting
1344 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1345 * function to do this! YYY we really should 'block' here.
1346 *
1347 * Must be called from a critical section. Thsi routine may be called
1348 * from an interrupt (for example, if an interrupt wakes a foreign thread
1349 * up).
1350 */
1351void
1352lwkt_wait_ipiq(int dcpu, int seq)
1353{
1354 lwkt_ipiq_t ip;
a2a5ad0d 1355 int maxc = 100000000;
96728c05
MD
1356
1357 if (dcpu != mycpu->gd_cpuid) {
1358 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1359 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1360 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1361 unsigned int eflags = read_eflags();
1362 cpu_enable_intr();
cb973d15 1363 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1364 lwkt_process_ipiq();
a2a5ad0d 1365 if (--maxc == 0)
cb973d15 1366 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1367 if (maxc < -1000000)
1368 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1369 }
1370 write_eflags(eflags);
1371 }
1372 }
1373}
1374
1375/*
1376 * Called from IPI interrupt (like a fast interrupt), which has placed
1377 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1378 * May also be called from doreti or splz, or be reentrantly called
1379 * indirectly through the ip_func[] we run.
96728c05
MD
1380 */
1381void
1382lwkt_process_ipiq(void)
1383{
1384 int n;
1385 int cpuid = mycpu->gd_cpuid;
1386
1387 for (n = 0; n < ncpus; ++n) {
1388 lwkt_ipiq_t ip;
1389 int ri;
1390
1391 if (n == cpuid)
1392 continue;
1393 ip = globaldata_find(n)->gd_ipiq;
1394 if (ip == NULL)
1395 continue;
1396 ip = &ip[cpuid];
cb973d15
MD
1397
1398 /*
1399 * Note: xindex is only updated after we are sure the function has
1400 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1401 * function may send an IPI which may block/drain.
1402 */
96728c05
MD
1403 while (ip->ip_rindex != ip->ip_windex) {
1404 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1405 ++ip->ip_rindex;
cb973d15
MD
1406 ip->ip_func[ri](ip->ip_arg[ri]);
1407 /* YYY memory barrier */
1408 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1409 }
1410 }
1411}
1412
1413#else
1414
1415int
1416lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1417{
1418 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1419 return(0); /* NOT REACHED */
1420}
1421
1422void
1423lwkt_wait_ipiq(int dcpu, int seq)
1424{
1425 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1426}
1427
1428#endif