Do not attempt to access kernel_map in free(). It's a bad idea, and doubly
[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 *
f8c3996b 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.30 2003/09/24 18:37:48 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
290 * the target thread (not the current thread).
8ad65e08 291 */
96728c05 292
8ad65e08
MD
293void
294lwkt_switch(void)
295{
4b5f931b 296 struct globaldata *gd;
f1d1c3fa 297 thread_t td = curthread;
8ad65e08 298 thread_t ntd;
8a8d5d85
MD
299#ifdef SMP
300 int mpheld;
301#endif
8ad65e08 302
96728c05
MD
303 if (mycpu->gd_intr_nesting_level &&
304 td->td_preempted == NULL && panicstr == NULL
305 ) {
03aa8d99 306 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 307 }
ef0fdad1 308
cb973d15
MD
309 /*
310 * Passive release (used to transition from user to kernel mode
311 * when we block or switch rather then when we enter the kernel).
312 * This function is NOT called if we are switching into a preemption
313 * or returning from a preemption. Typically this causes us to lose
314 * our P_CURPROC designation (if we have one) and become a true LWKT
315 * thread, and may also hand P_CURPROC to another process and schedule
316 * its thread.
317 */
318 if (td->td_release)
319 td->td_release(td);
320
f1d1c3fa 321 crit_enter();
4b5f931b 322 ++switch_count;
8a8d5d85
MD
323
324#ifdef SMP
325 /*
326 * td_mpcount cannot be used to determine if we currently hold the
327 * MP lock because get_mplock() will increment it prior to attempting
328 * to get the lock, and switch out if it can't. Look at the actual lock.
329 */
330 mpheld = MP_LOCK_HELD();
331#endif
99df837e
MD
332 if ((ntd = td->td_preempted) != NULL) {
333 /*
334 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
335 * thread. This occurs transparently, whether the preempted thread
336 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
337 * itself).
338 *
339 * We have to setup the MP lock for the original thread after backing
340 * out the adjustment that was made to curthread when the original
341 * was preempted.
99df837e 342 */
26a0694b 343 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 344#ifdef SMP
96728c05
MD
345 if (ntd->td_mpcount && mpheld == 0) {
346 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
347 td, ntd, td->td_mpcount, ntd->td_mpcount);
348 }
8a8d5d85
MD
349 if (ntd->td_mpcount) {
350 td->td_mpcount -= ntd->td_mpcount;
351 KKASSERT(td->td_mpcount >= 0);
352 }
353#endif
26a0694b 354 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 355 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 356 } else {
4b5f931b
MD
357 /*
358 * Priority queue / round-robin at each priority. Note that user
359 * processes run at a fixed, low priority and the user process
360 * scheduler deals with interactions between user processes
361 * by scheduling and descheduling them from the LWKT queue as
362 * necessary.
8a8d5d85
MD
363 *
364 * We have to adjust the MP lock for the target thread. If we
365 * need the MP lock and cannot obtain it we try to locate a
366 * thread that does not need the MP lock.
4b5f931b
MD
367 */
368 gd = mycpu;
4b5f931b
MD
369again:
370 if (gd->gd_runqmask) {
371 int nq = bsrl(gd->gd_runqmask);
372 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
373 gd->gd_runqmask &= ~(1 << nq);
374 goto again;
375 }
8a8d5d85
MD
376#ifdef SMP
377 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
378 /*
96728c05
MD
379 * Target needs MP lock and we couldn't get it, try
380 * to locate a thread which does not need the MP lock
3c23a41a 381 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
382 */
383 u_int32_t rqmask = gd->gd_runqmask;
384 while (rqmask) {
385 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
386 if (ntd->td_mpcount == 0)
387 break;
388 }
389 if (ntd)
390 break;
391 rqmask &= ~(1 << nq);
392 nq = bsrl(rqmask);
393 }
394 if (ntd == NULL) {
a2a5ad0d
MD
395 ntd = &gd->gd_idlethread;
396 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
397 } else {
398 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
399 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
400 }
401 } else {
402 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
403 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
404 }
405#else
4b5f931b
MD
406 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
407 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 408#endif
4b5f931b 409 } else {
3c23a41a
MD
410 /*
411 * Nothing to run but we may still need the BGL to deal with
412 * pending interrupts, spin in idle if so.
413 */
a2a5ad0d 414 ntd = &gd->gd_idlethread;
235957ed 415 if (gd->gd_reqflags)
3c23a41a 416 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 417 }
f1d1c3fa 418 }
26a0694b
MD
419 KASSERT(ntd->td_pri >= TDPRI_CRIT,
420 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
421
422 /*
423 * Do the actual switch. If the new target does not need the MP lock
424 * and we are holding it, release the MP lock. If the new target requires
425 * the MP lock we have already acquired it for the target.
426 */
427#ifdef SMP
428 if (ntd->td_mpcount == 0 ) {
429 if (MP_LOCK_HELD())
430 cpu_rel_mplock();
431 } else {
432 ASSERT_MP_LOCK_HELD();
433 }
434#endif
8a8d5d85 435 if (td != ntd) {
f1d1c3fa 436 td->td_switch(ntd);
8a8d5d85 437 }
96728c05 438
f1d1c3fa 439 crit_exit();
8ad65e08
MD
440}
441
cb973d15
MD
442/*
443 * Switch if another thread has a higher priority. Do not switch to other
444 * threads at the same priority.
445 */
446void
447lwkt_maybe_switch()
448{
449 struct globaldata *gd = mycpu;
450 struct thread *td = gd->gd_curthread;
451
452 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
453 lwkt_switch();
454 }
455}
456
b68b7282 457/*
96728c05
MD
458 * Request that the target thread preempt the current thread. Preemption
459 * only works under a specific set of conditions:
b68b7282 460 *
96728c05
MD
461 * - We are not preempting ourselves
462 * - The target thread is owned by the current cpu
463 * - We are not currently being preempted
464 * - The target is not currently being preempted
465 * - We are able to satisfy the target's MP lock requirements (if any).
466 *
467 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
468 * this is called via lwkt_schedule() through the td_preemptable callback.
469 * critpri is the managed critical priority that we should ignore in order
470 * to determine whether preemption is possible (aka usually just the crit
471 * priority of lwkt_schedule() itself).
b68b7282 472 *
26a0694b
MD
473 * XXX at the moment we run the target thread in a critical section during
474 * the preemption in order to prevent the target from taking interrupts
475 * that *WE* can't. Preemption is strictly limited to interrupt threads
476 * and interrupt-like threads, outside of a critical section, and the
477 * preempted source thread will be resumed the instant the target blocks
478 * whether or not the source is scheduled (i.e. preemption is supposed to
479 * be as transparent as possible).
4b5f931b 480 *
8a8d5d85
MD
481 * The target thread inherits our MP count (added to its own) for the
482 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
483 * MP lock during the preemption. Therefore, any preempting targets must be
484 * careful in regards to MP assertions. Note that the MP count may be
485 * out of sync with the physical mp_lock. If we preempt we have to preserve
486 * the expected situation.
b68b7282
MD
487 */
488void
96728c05 489lwkt_preempt(thread_t ntd, int critpri)
b68b7282 490{
73e4f7b9 491 thread_t td = curthread;
8a8d5d85
MD
492#ifdef SMP
493 int mpheld;
57c254db 494 int savecnt;
8a8d5d85 495#endif
b68b7282 496
26a0694b 497 /*
96728c05
MD
498 * The caller has put us in a critical section. We can only preempt
499 * if the caller of the caller was not in a critical section (basically
57c254db
MD
500 * a local interrupt), as determined by the 'critpri' parameter. If
501 * we are unable to preempt
96728c05
MD
502 *
503 * YYY The target thread must be in a critical section (else it must
504 * inherit our critical section? I dunno yet).
26a0694b
MD
505 */
506 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 507
cb973d15 508 need_resched();
57c254db
MD
509 if (!_lwkt_wantresched(ntd, td)) {
510 ++preempt_miss;
511 return;
512 }
96728c05
MD
513 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
514 ++preempt_miss;
515 return;
516 }
517#ifdef SMP
a72187e9 518 if (ntd->td_gd != mycpu) {
96728c05
MD
519 ++preempt_miss;
520 return;
521 }
522#endif
26a0694b
MD
523 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
524 ++preempt_weird;
525 return;
526 }
527 if (ntd->td_preempted) {
4b5f931b 528 ++preempt_hit;
26a0694b 529 return;
b68b7282 530 }
8a8d5d85 531#ifdef SMP
a2a5ad0d
MD
532 /*
533 * note: an interrupt might have occured just as we were transitioning
a5934754
MD
534 * to the MP lock. In this case td_mpcount will be pre-disposed but
535 * not actually synchronized with the actual state of the lock. We
536 * can use it to imply an MP lock requirement for the preemption but
537 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 538 */
a5934754 539 mpheld = MP_LOCK_HELD();
96728c05
MD
540 if (mpheld && td->td_mpcount == 0)
541 panic("lwkt_preempt(): held and no count");
542 savecnt = td->td_mpcount;
8a8d5d85
MD
543 ntd->td_mpcount += td->td_mpcount;
544 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
545 ntd->td_mpcount -= td->td_mpcount;
546 ++preempt_miss;
547 return;
548 }
549#endif
26a0694b
MD
550
551 ++preempt_hit;
552 ntd->td_preempted = td;
553 td->td_flags |= TDF_PREEMPT_LOCK;
554 td->td_switch(ntd);
555 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
556#ifdef SMP
557 KKASSERT(savecnt == td->td_mpcount);
558 if (mpheld == 0 && MP_LOCK_HELD())
559 cpu_rel_mplock();
560 else if (mpheld && !MP_LOCK_HELD())
561 panic("lwkt_preempt(): MP lock was not held through");
562#endif
26a0694b
MD
563 ntd->td_preempted = NULL;
564 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
565}
566
f1d1c3fa
MD
567/*
568 * Yield our thread while higher priority threads are pending. This is
569 * typically called when we leave a critical section but it can be safely
570 * called while we are in a critical section.
571 *
572 * This function will not generally yield to equal priority threads but it
573 * can occur as a side effect. Note that lwkt_switch() is called from
574 * inside the critical section to pervent its own crit_exit() from reentering
575 * lwkt_yield_quick().
576 *
235957ed 577 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
578 * came along but was blocked and made pending.
579 *
f1d1c3fa
MD
580 * (self contained on a per cpu basis)
581 */
582void
583lwkt_yield_quick(void)
584{
7966cb69
MD
585 globaldata_t gd = mycpu;
586 thread_t td = gd->gd_curthread;
ef0fdad1 587
a2a5ad0d 588 /*
235957ed 589 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
590 * it with a non-zero cpl then we might not wind up calling splz after
591 * a task switch when the critical section is exited even though the
592 * new task could accept the interrupt. YYY alternative is to have
593 * lwkt_switch() just call splz unconditionally.
594 *
595 * XXX from crit_exit() only called after last crit section is released.
596 * If called directly will run splz() even if in a critical section.
597 */
7966cb69 598 if (gd->gd_reqflags)
f1d1c3fa 599 splz();
f1d1c3fa
MD
600
601 /*
602 * YYY enabling will cause wakeup() to task-switch, which really
603 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
604 * preemption and MP without actually doing preemption or MP, because a
605 * lot of code assumes that wakeup() does not block.
f1d1c3fa 606 */
7966cb69 607 if (untimely_switch && gd->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
608 crit_enter();
609 /*
610 * YYY temporary hacks until we disassociate the userland scheduler
611 * from the LWKT scheduler.
612 */
613 if (td->td_flags & TDF_RUNQ) {
614 lwkt_switch(); /* will not reenter yield function */
615 } else {
616 lwkt_schedule_self(); /* make sure we are scheduled */
617 lwkt_switch(); /* will not reenter yield function */
618 lwkt_deschedule_self(); /* make sure we are descheduled */
619 }
7966cb69 620 crit_exit_noyield(td);
f1d1c3fa 621 }
f1d1c3fa
MD
622}
623
8ad65e08 624/*
f1d1c3fa 625 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 626 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
627 * the crit_exit() call in lwkt_switch().
628 *
629 * (self contained on a per cpu basis)
8ad65e08
MD
630 */
631void
f1d1c3fa 632lwkt_yield(void)
8ad65e08 633{
f1d1c3fa
MD
634 lwkt_schedule_self();
635 lwkt_switch();
636}
637
638/*
639 * Schedule a thread to run. As the current thread we can always safely
640 * schedule ourselves, and a shortcut procedure is provided for that
641 * function.
642 *
643 * (non-blocking, self contained on a per cpu basis)
644 */
645void
646lwkt_schedule_self(void)
647{
648 thread_t td = curthread;
649
650 crit_enter();
651 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 652 _lwkt_enqueue(td);
26a0694b
MD
653 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
654 panic("SCHED SELF PANIC");
f1d1c3fa 655 crit_exit();
8ad65e08 656}
8ad65e08
MD
657
658/*
f1d1c3fa
MD
659 * Generic schedule. Possibly schedule threads belonging to other cpus and
660 * deal with threads that might be blocked on a wait queue.
661 *
96728c05 662 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
663 * Load balancing can be accomplished by requesting other sorts of actions
664 * for the thread in question.
8ad65e08
MD
665 */
666void
667lwkt_schedule(thread_t td)
668{
96728c05 669#ifdef INVARIANTS
26a0694b
MD
670 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
671 && td->td_proc->p_stat == SSLEEP
672 ) {
673 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
674 curthread,
675 curthread->td_proc ? curthread->td_proc->p_pid : -1,
676 curthread->td_proc ? curthread->td_proc->p_stat : -1,
677 td,
678 td->td_proc ? curthread->td_proc->p_pid : -1,
679 td->td_proc ? curthread->td_proc->p_stat : -1
680 );
681 panic("SCHED PANIC");
682 }
96728c05 683#endif
f1d1c3fa
MD
684 crit_enter();
685 if (td == curthread) {
686 _lwkt_enqueue(td);
687 } else {
688 lwkt_wait_t w;
689
690 /*
691 * If the thread is on a wait list we have to send our scheduling
692 * request to the owner of the wait structure. Otherwise we send
693 * the scheduling request to the cpu owning the thread. Races
694 * are ok, the target will forward the message as necessary (the
695 * message may chase the thread around before it finally gets
696 * acted upon).
697 *
698 * (remember, wait structures use stable storage)
699 */
700 if ((w = td->td_wait) != NULL) {
96728c05 701 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
702 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
703 --w->wa_count;
704 td->td_wait = NULL;
a72187e9 705 if (td->td_gd == mycpu) {
f1d1c3fa 706 _lwkt_enqueue(td);
57c254db 707 if (td->td_preemptable) {
96728c05 708 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
709 } else if (_lwkt_wantresched(td, curthread)) {
710 need_resched();
711 }
f1d1c3fa 712 } else {
a72187e9 713 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 714 }
96728c05 715 lwkt_reltoken(&w->wa_token);
f1d1c3fa 716 } else {
96728c05 717 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
718 }
719 } else {
720 /*
721 * If the wait structure is NULL and we own the thread, there
722 * is no race (since we are in a critical section). If we
723 * do not own the thread there might be a race but the
724 * target cpu will deal with it.
725 */
a72187e9 726 if (td->td_gd == mycpu) {
f1d1c3fa 727 _lwkt_enqueue(td);
57c254db 728 if (td->td_preemptable) {
96728c05 729 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
730 } else if (_lwkt_wantresched(td, curthread)) {
731 need_resched();
732 }
f1d1c3fa 733 } else {
a72187e9 734 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
735 }
736 }
8ad65e08 737 }
f1d1c3fa 738 crit_exit();
8ad65e08
MD
739}
740
d9eea1a5
MD
741/*
742 * Managed acquisition. This code assumes that the MP lock is held for
743 * the tdallq operation and that the thread has been descheduled from its
744 * original cpu. We also have to wait for the thread to be entirely switched
745 * out on its original cpu (this is usually fast enough that we never loop)
746 * since the LWKT system does not have to hold the MP lock while switching
747 * and the target may have released it before switching.
748 */
a2a5ad0d
MD
749void
750lwkt_acquire(thread_t td)
751{
752 struct globaldata *gd;
753
754 gd = td->td_gd;
755 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
756 while (td->td_flags & TDF_RUNNING) /* XXX spin */
757 ;
a2a5ad0d
MD
758 if (gd != mycpu) {
759 crit_enter();
760 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
761 gd = mycpu;
762 td->td_gd = gd;
a2a5ad0d
MD
763 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
764 crit_exit();
765 }
766}
767
8ad65e08 768/*
f1d1c3fa
MD
769 * Deschedule a thread.
770 *
771 * (non-blocking, self contained on a per cpu basis)
772 */
773void
774lwkt_deschedule_self(void)
775{
776 thread_t td = curthread;
777
778 crit_enter();
779 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
780 _lwkt_dequeue(td);
781 crit_exit();
782}
783
784/*
785 * Generic deschedule. Descheduling threads other then your own should be
786 * done only in carefully controlled circumstances. Descheduling is
787 * asynchronous.
788 *
789 * This function may block if the cpu has run out of messages.
8ad65e08
MD
790 */
791void
792lwkt_deschedule(thread_t td)
793{
f1d1c3fa
MD
794 crit_enter();
795 if (td == curthread) {
796 _lwkt_dequeue(td);
797 } else {
a72187e9 798 if (td->td_gd == mycpu) {
f1d1c3fa
MD
799 _lwkt_dequeue(td);
800 } else {
a72187e9 801 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
802 }
803 }
804 crit_exit();
805}
806
4b5f931b
MD
807/*
808 * Set the target thread's priority. This routine does not automatically
809 * switch to a higher priority thread, LWKT threads are not designed for
810 * continuous priority changes. Yield if you want to switch.
811 *
812 * We have to retain the critical section count which uses the high bits
26a0694b
MD
813 * of the td_pri field. The specified priority may also indicate zero or
814 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
815 */
816void
817lwkt_setpri(thread_t td, int pri)
818{
26a0694b 819 KKASSERT(pri >= 0);
a72187e9 820 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
821 crit_enter();
822 if (td->td_flags & TDF_RUNQ) {
823 _lwkt_dequeue(td);
824 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
825 _lwkt_enqueue(td);
826 } else {
827 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
828 }
829 crit_exit();
830}
831
832void
833lwkt_setpri_self(int pri)
834{
835 thread_t td = curthread;
836
4b5f931b
MD
837 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
838 crit_enter();
839 if (td->td_flags & TDF_RUNQ) {
840 _lwkt_dequeue(td);
841 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
842 _lwkt_enqueue(td);
843 } else {
844 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
845 }
846 crit_exit();
847}
848
849struct proc *
850lwkt_preempted_proc(void)
851{
73e4f7b9 852 thread_t td = curthread;
4b5f931b
MD
853 while (td->td_preempted)
854 td = td->td_preempted;
855 return(td->td_proc);
856}
857
ece04fd0
MD
858typedef struct lwkt_gettoken_req {
859 lwkt_token_t tok;
860 int cpu;
861} lwkt_gettoken_req;
862
863#if 0
4b5f931b 864
f1d1c3fa
MD
865/*
866 * This function deschedules the current thread and blocks on the specified
867 * wait queue. We obtain ownership of the wait queue in order to block
868 * on it. A generation number is used to interlock the wait queue in case
869 * it gets signalled while we are blocked waiting on the token.
870 *
871 * Note: alternatively we could dequeue our thread and then message the
872 * target cpu owning the wait queue. YYY implement as sysctl.
873 *
874 * Note: wait queue signals normally ping-pong the cpu as an optimization.
875 */
96728c05 876
f1d1c3fa 877void
ae8050a4 878lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
879{
880 thread_t td = curthread;
f1d1c3fa 881
f1d1c3fa 882 lwkt_gettoken(&w->wa_token);
ae8050a4 883 if (w->wa_gen == *gen) {
f1d1c3fa
MD
884 _lwkt_dequeue(td);
885 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
886 ++w->wa_count;
887 td->td_wait = w;
ae8050a4 888 td->td_wmesg = wmesg;
ece04fd0 889again:
f1d1c3fa 890 lwkt_switch();
ece04fd0
MD
891 lwkt_regettoken(&w->wa_token);
892 if (td->td_wmesg != NULL) {
893 _lwkt_dequeue(td);
894 goto again;
895 }
8ad65e08 896 }
ae8050a4
MD
897 /* token might be lost, doesn't matter for gen update */
898 *gen = w->wa_gen;
f1d1c3fa
MD
899 lwkt_reltoken(&w->wa_token);
900}
901
902/*
903 * Signal a wait queue. We gain ownership of the wait queue in order to
904 * signal it. Once a thread is removed from the wait queue we have to
905 * deal with the cpu owning the thread.
906 *
907 * Note: alternatively we could message the target cpu owning the wait
908 * queue. YYY implement as sysctl.
909 */
910void
ece04fd0 911lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
912{
913 thread_t td;
914 int count;
915
916 lwkt_gettoken(&w->wa_token);
917 ++w->wa_gen;
ece04fd0
MD
918 if (count < 0)
919 count = w->wa_count;
f1d1c3fa
MD
920 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
921 --count;
922 --w->wa_count;
923 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
924 td->td_wait = NULL;
ae8050a4 925 td->td_wmesg = NULL;
a72187e9 926 if (td->td_gd == mycpu) {
f1d1c3fa
MD
927 _lwkt_enqueue(td);
928 } else {
a72187e9 929 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
930 }
931 lwkt_regettoken(&w->wa_token);
932 }
933 lwkt_reltoken(&w->wa_token);
934}
935
ece04fd0
MD
936#endif
937
f1d1c3fa 938/*
96728c05 939 * Acquire ownership of a token
f1d1c3fa 940 *
96728c05 941 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
942 * section side effects, depending on its purpose. These side effects
943 * guarentee that you will maintain ownership of the token as long as you
944 * do not block. If you block you may lose access to the token (but you
945 * must still release it even if you lose your access to it).
946 *
96728c05 947 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 948 * a token, but do we really only need to disable IPIs ?
96728c05
MD
949 *
950 * YYY certain tokens could be made to act like mutexes when performance
951 * would be better (e.g. t_cpu == -1). This is not yet implemented.
952 *
a2a5ad0d
MD
953 * YYY the tokens replace 4.x's simplelocks for the most part, but this
954 * means that 4.x does not expect a switch so for now we cannot switch
955 * when waiting for an IPI to be returned.
956 *
957 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
958 * it and then block. The IPI causes the token to be given away to the
959 * requesting cpu, unless it has already changed hands. Since only the
960 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 961 * This needs serious optimization.
f1d1c3fa 962 */
57c254db
MD
963
964#ifdef SMP
965
96728c05
MD
966static
967void
968lwkt_gettoken_remote(void *arg)
969{
970 lwkt_gettoken_req *req = arg;
971 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 972#ifdef INVARIANTS
a2a5ad0d
MD
973 if (token_debug)
974 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 975#endif
96728c05 976 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
977 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
978 /* else set reqcpu to point to current cpu for release */
96728c05
MD
979 }
980}
981
57c254db
MD
982#endif
983
8a8d5d85 984int
f1d1c3fa
MD
985lwkt_gettoken(lwkt_token_t tok)
986{
987 /*
988 * Prevent preemption so the token can't be taken away from us once
989 * we gain ownership of it. Use a synchronous request which might
990 * block. The request will be forwarded as necessary playing catchup
991 * to the token.
992 */
96728c05 993
f1d1c3fa 994 crit_enter();
57c254db 995#ifdef INVARIANTS
a2a5ad0d
MD
996 if (curthread->td_pri > 2000) {
997 curthread->td_pri = 1000;
998 panic("too HIGH!");
57c254db
MD
999 }
1000#endif
96728c05 1001#ifdef SMP
d0e06f83 1002 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1003 struct lwkt_gettoken_req req;
1004 int seq;
96728c05
MD
1005 int dcpu;
1006
1007 req.cpu = mycpu->gd_cpuid;
1008 req.tok = tok;
1009 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1010 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1011#ifdef INVARIANTS
a2a5ad0d
MD
1012 if (token_debug)
1013 printf("REQT%d ", dcpu);
634081ff 1014#endif
96728c05
MD
1015 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1016 lwkt_wait_ipiq(dcpu, seq);
634081ff 1017#ifdef INVARIANTS
a2a5ad0d
MD
1018 if (token_debug)
1019 printf("REQR%d ", tok->t_cpu);
634081ff 1020#endif
f1d1c3fa
MD
1021 }
1022#endif
1023 /*
1024 * leave us in a critical section on return. This will be undone
8a8d5d85 1025 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1026 */
8a8d5d85 1027 return(++tok->t_gen);
f1d1c3fa
MD
1028}
1029
96728c05
MD
1030/*
1031 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1032 * failure.
1033 */
1034int
1035lwkt_trytoken(lwkt_token_t tok)
1036{
1037 crit_enter();
1038#ifdef SMP
1039 if (tok->t_cpu != mycpu->gd_cpuid) {
1040 return(0);
1041 }
1042#endif
1043 /* leave us in the critical section */
1044 ++tok->t_gen;
1045 return(1);
1046}
1047
f1d1c3fa
MD
1048/*
1049 * Release your ownership of a token. Releases must occur in reverse
1050 * order to aquisitions, eventually so priorities can be unwound properly
1051 * like SPLs. At the moment the actual implemention doesn't care.
1052 *
1053 * We can safely hand a token that we own to another cpu without notifying
1054 * it, but once we do we can't get it back without requesting it (unless
1055 * the other cpu hands it back to us before we check).
1056 *
1057 * We might have lost the token, so check that.
1058 */
1059void
1060lwkt_reltoken(lwkt_token_t tok)
1061{
d0e06f83 1062 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1063 tok->t_cpu = tok->t_reqcpu;
1064 }
1065 crit_exit();
1066}
1067
1068/*
8a8d5d85
MD
1069 * Reacquire a token that might have been lost and compare and update the
1070 * generation number. 0 is returned if the generation has not changed
1071 * (nobody else obtained the token while we were blocked, on this cpu or
1072 * any other cpu).
1073 *
1074 * This function returns with the token re-held whether the generation
1075 * number changed or not.
1076 */
1077int
1078lwkt_gentoken(lwkt_token_t tok, int *gen)
1079{
1080 if (lwkt_regettoken(tok) == *gen) {
1081 return(0);
1082 } else {
1083 *gen = tok->t_gen;
1084 return(-1);
1085 }
1086}
1087
1088
1089/*
96728c05 1090 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1091 * number of the token.
f1d1c3fa
MD
1092 */
1093int
1094lwkt_regettoken(lwkt_token_t tok)
1095{
96728c05 1096 /* assert we are in a critical section */
d0e06f83 1097 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1098#ifdef SMP
d0e06f83 1099 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1100 struct lwkt_gettoken_req req;
1101 int seq;
96728c05 1102 int dcpu;
57c254db 1103
96728c05
MD
1104 req.cpu = mycpu->gd_cpuid;
1105 req.tok = tok;
1106 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1107 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1108#ifdef INVARIANTS
cb973d15
MD
1109 if (token_debug)
1110 printf("REQT%d ", dcpu);
634081ff 1111#endif
96728c05
MD
1112 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1113 lwkt_wait_ipiq(dcpu, seq);
634081ff 1114#ifdef INVARIATNS
cb973d15
MD
1115 if (token_debug)
1116 printf("REQR%d ", tok->t_cpu);
634081ff 1117#endif
f1d1c3fa 1118 }
f1d1c3fa 1119#endif
96728c05
MD
1120 ++tok->t_gen;
1121 }
8a8d5d85 1122 return(tok->t_gen);
8ad65e08
MD
1123}
1124
72740893
MD
1125void
1126lwkt_inittoken(lwkt_token_t tok)
1127{
1128 /*
1129 * Zero structure and set cpu owner and reqcpu to cpu 0.
1130 */
1131 bzero(tok, sizeof(*tok));
1132}
1133
99df837e
MD
1134/*
1135 * Create a kernel process/thread/whatever. It shares it's address space
1136 * with proc0 - ie: kernel only.
1137 *
1138 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1139 *
1140 * The thread will be entered with the MP lock held.
99df837e
MD
1141 */
1142int
1143lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1144 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1145 const char *fmt, ...)
99df837e 1146{
73e4f7b9 1147 thread_t td;
99df837e
MD
1148 va_list ap;
1149
a2a5ad0d
MD
1150 td = lwkt_alloc_thread(template);
1151 if (tdp)
1152 *tdp = td;
99df837e 1153 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1154 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1155#ifdef SMP
1156 td->td_mpcount = 1;
1157#endif
99df837e
MD
1158
1159 /*
1160 * Set up arg0 for 'ps' etc
1161 */
1162 va_start(ap, fmt);
1163 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1164 va_end(ap);
1165
1166 /*
1167 * Schedule the thread to run
1168 */
ef0fdad1
MD
1169 if ((td->td_flags & TDF_STOPREQ) == 0)
1170 lwkt_schedule(td);
1171 else
1172 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1173 return 0;
1174}
1175
1176/*
1177 * Destroy an LWKT thread. Warning! This function is not called when
1178 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1179 * uses a different reaping mechanism.
1180 */
1181void
1182lwkt_exit(void)
1183{
1184 thread_t td = curthread;
1185
1186 if (td->td_flags & TDF_VERBOSE)
1187 printf("kthread %p %s has exited\n", td, td->td_comm);
1188 crit_enter();
1189 lwkt_deschedule_self();
1190 ++mycpu->gd_tdfreecount;
1191 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1192 cpu_thread_exit();
1193}
1194
1195/*
1196 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1197 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1198 */
1199int
1200kthread_create(void (*func)(void *), void *arg,
1201 struct thread **tdp, const char *fmt, ...)
1202{
73e4f7b9 1203 thread_t td;
99df837e
MD
1204 va_list ap;
1205
a2a5ad0d
MD
1206 td = lwkt_alloc_thread(NULL);
1207 if (tdp)
1208 *tdp = td;
99df837e
MD
1209 cpu_set_thread_handler(td, kthread_exit, func, arg);
1210 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1211#ifdef SMP
1212 td->td_mpcount = 1;
1213#endif
99df837e
MD
1214
1215 /*
1216 * Set up arg0 for 'ps' etc
1217 */
1218 va_start(ap, fmt);
1219 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1220 va_end(ap);
1221
1222 /*
1223 * Schedule the thread to run
1224 */
1225 lwkt_schedule(td);
1226 return 0;
1227}
1228
26a0694b
MD
1229void
1230crit_panic(void)
1231{
73e4f7b9 1232 thread_t td = curthread;
26a0694b
MD
1233 int lpri = td->td_pri;
1234
1235 td->td_pri = 0;
1236 panic("td_pri is/would-go negative! %p %d", td, lpri);
1237}
1238
99df837e
MD
1239/*
1240 * Destroy an LWKT thread. Warning! This function is not called when
1241 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1242 * uses a different reaping mechanism.
1243 *
1244 * XXX duplicates lwkt_exit()
1245 */
1246void
1247kthread_exit(void)
1248{
1249 lwkt_exit();
1250}
1251
96728c05
MD
1252#ifdef SMP
1253
1254/*
1255 * Send a function execution request to another cpu. The request is queued
1256 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1257 * possible target cpu. The FIFO can be written.
1258 *
1259 * YYY If the FIFO fills up we have to enable interrupts and process the
1260 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1261 * Create a CPU_*() function to do this!
1262 *
1263 * Must be called from a critical section.
1264 */
1265int
1266lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1267{
1268 lwkt_ipiq_t ip;
1269 int windex;
a2a5ad0d 1270 struct globaldata *gd = mycpu;
96728c05 1271
a2a5ad0d 1272 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1273 func(arg);
1274 return(0);
1275 }
cb973d15 1276 crit_enter();
a2a5ad0d
MD
1277 ++gd->gd_intr_nesting_level;
1278#ifdef INVARIANTS
1279 if (gd->gd_intr_nesting_level > 20)
1280 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1281#endif
96728c05
MD
1282 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1283 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1284 ++ipiq_count;
a2a5ad0d 1285 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1286
1287 /*
1288 * We always drain before the FIFO becomes full so it should never
1289 * become full. We need to leave enough entries to deal with
1290 * reentrancy.
1291 */
1292 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1293 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1294 ip->ip_func[windex] = func;
1295 ip->ip_arg[windex] = arg;
1296 /* YYY memory barrier */
1297 ++ip->ip_windex;
96728c05
MD
1298 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1299 unsigned int eflags = read_eflags();
1300 cpu_enable_intr();
1301 ++ipiq_fifofull;
cb973d15 1302 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1303 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1304 lwkt_process_ipiq();
1305 }
1306 write_eflags(eflags);
1307 }
a2a5ad0d 1308 --gd->gd_intr_nesting_level;
96728c05 1309 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1310 crit_exit();
96728c05
MD
1311 return(ip->ip_windex);
1312}
1313
cb973d15
MD
1314/*
1315 * Send a message to several target cpus. Typically used for scheduling.
1316 */
1317void
1318lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1319{
1320 int cpuid;
1321
1322 while (mask) {
1323 cpuid = bsfl(mask);
1324 lwkt_send_ipiq(cpuid, func, arg);
1325 mask &= ~(1 << cpuid);
1326 }
1327}
1328
96728c05
MD
1329/*
1330 * Wait for the remote cpu to finish processing a function.
1331 *
1332 * YYY we have to enable interrupts and process the IPIQ while waiting
1333 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1334 * function to do this! YYY we really should 'block' here.
1335 *
1336 * Must be called from a critical section. Thsi routine may be called
1337 * from an interrupt (for example, if an interrupt wakes a foreign thread
1338 * up).
1339 */
1340void
1341lwkt_wait_ipiq(int dcpu, int seq)
1342{
1343 lwkt_ipiq_t ip;
a2a5ad0d 1344 int maxc = 100000000;
96728c05
MD
1345
1346 if (dcpu != mycpu->gd_cpuid) {
1347 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1348 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1349 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1350 unsigned int eflags = read_eflags();
1351 cpu_enable_intr();
cb973d15 1352 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1353 lwkt_process_ipiq();
a2a5ad0d 1354 if (--maxc == 0)
cb973d15 1355 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1356 if (maxc < -1000000)
1357 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1358 }
1359 write_eflags(eflags);
1360 }
1361 }
1362}
1363
1364/*
1365 * Called from IPI interrupt (like a fast interrupt), which has placed
1366 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1367 * May also be called from doreti or splz, or be reentrantly called
1368 * indirectly through the ip_func[] we run.
96728c05
MD
1369 */
1370void
1371lwkt_process_ipiq(void)
1372{
1373 int n;
1374 int cpuid = mycpu->gd_cpuid;
1375
1376 for (n = 0; n < ncpus; ++n) {
1377 lwkt_ipiq_t ip;
1378 int ri;
1379
1380 if (n == cpuid)
1381 continue;
1382 ip = globaldata_find(n)->gd_ipiq;
1383 if (ip == NULL)
1384 continue;
1385 ip = &ip[cpuid];
cb973d15
MD
1386
1387 /*
1388 * Note: xindex is only updated after we are sure the function has
1389 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1390 * function may send an IPI which may block/drain.
1391 */
96728c05
MD
1392 while (ip->ip_rindex != ip->ip_windex) {
1393 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1394 ++ip->ip_rindex;
cb973d15
MD
1395 ip->ip_func[ri](ip->ip_arg[ri]);
1396 /* YYY memory barrier */
1397 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1398 }
1399 }
1400}
1401
1402#else
1403
1404int
1405lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1406{
1407 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1408 return(0); /* NOT REACHED */
1409}
1410
1411void
1412lwkt_wait_ipiq(int dcpu, int seq)
1413{
1414 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1415}
1416
1417#endif