As part of the libcaps threading work a number of routines in lwkt_thread.c
[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 *
2d93b37a 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.45 2003/12/04 00:12:40 dillon Exp $
75cdbe6c
MD
27 */
28
29/*
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
f1d1c3fa 34 *
75cdbe6c
MD
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
8ad65e08
MD
39 */
40
05220613
MD
41#ifdef _KERNEL
42
8ad65e08
MD
43#include <sys/param.h>
44#include <sys/systm.h>
45#include <sys/kernel.h>
46#include <sys/proc.h>
47#include <sys/rtprio.h>
48#include <sys/queue.h>
f1d1c3fa 49#include <sys/thread2.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f1d1c3fa 54
7d0bac62
MD
55#include <vm/vm.h>
56#include <vm/vm_param.h>
57#include <vm/vm_kern.h>
58#include <vm/vm_object.h>
59#include <vm/vm_page.h>
60#include <vm/vm_map.h>
61#include <vm/vm_pager.h>
62#include <vm/vm_extern.h>
63#include <vm/vm_zone.h>
64
99df837e 65#include <machine/stdarg.h>
57c254db 66#include <machine/ipl.h>
96728c05 67#include <machine/smp.h>
99df837e 68
05220613
MD
69#define THREAD_STACK (UPAGES * PAGE_SIZE)
70
71#else
72
73#include <sys/stdint.h>
fb04f4fd 74#include <libcaps/thread.h>
05220613
MD
75#include <sys/thread.h>
76#include <sys/msgport.h>
77#include <sys/errno.h>
fb04f4fd 78#include <libcaps/globaldata.h>
05220613
MD
79#include <sys/thread2.h>
80#include <sys/msgport2.h>
81#include <stdlib.h>
c95cd171 82#include <machine/cpufunc.h>
05220613
MD
83
84#endif
85
7d0bac62 86static int untimely_switch = 0;
57c254db
MD
87#ifdef INVARIANTS
88static int token_debug = 0;
05220613
MD
89#endif
90static __int64_t switch_count = 0;
91static __int64_t preempt_hit = 0;
92static __int64_t preempt_miss = 0;
93static __int64_t preempt_weird = 0;
94static __int64_t ipiq_count = 0;
95static __int64_t ipiq_fifofull = 0;
96
97#ifdef _KERNEL
98
99SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
100#ifdef INVARIANTS
57c254db
MD
101SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
102#endif
4b5f931b 103SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 104SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 105SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 106SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05 107SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
96728c05 108SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 109
05220613
MD
110#endif
111
4b5f931b
MD
112/*
113 * These helper procedures handle the runq, they can only be called from
114 * within a critical section.
75cdbe6c
MD
115 *
116 * WARNING! Prior to SMP being brought up it is possible to enqueue and
117 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
118 * instead of 'mycpu' when referencing the globaldata structure. Once
119 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 120 */
f1d1c3fa
MD
121static __inline
122void
123_lwkt_dequeue(thread_t td)
124{
125 if (td->td_flags & TDF_RUNQ) {
4b5f931b 126 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 127 struct globaldata *gd = td->td_gd;
4b5f931b 128
f1d1c3fa 129 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
130 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
131 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
132 }
133}
134
135static __inline
136void
137_lwkt_enqueue(thread_t td)
138{
139 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 140 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 141 struct globaldata *gd = td->td_gd;
4b5f931b 142
f1d1c3fa 143 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
144 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
145 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
146 }
147}
8ad65e08 148
57c254db
MD
149static __inline
150int
151_lwkt_wantresched(thread_t ntd, thread_t cur)
152{
153 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
154}
155
2d93b37a
MD
156/* lwkt_gdinit() has a userland override */
157#ifdef _KERNEL
158
8ad65e08
MD
159/*
160 * LWKTs operate on a per-cpu basis
161 *
73e4f7b9 162 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
163 */
164void
165lwkt_gdinit(struct globaldata *gd)
166{
4b5f931b
MD
167 int i;
168
169 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
170 TAILQ_INIT(&gd->gd_tdrunq[i]);
171 gd->gd_runqmask = 0;
73e4f7b9 172 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
173}
174
2d93b37a
MD
175#endif /* _KERNEL */
176
7d0bac62
MD
177/*
178 * Initialize a thread wait structure prior to first use.
179 *
180 * NOTE! called from low level boot code, we cannot do anything fancy!
181 */
182void
183lwkt_init_wait(lwkt_wait_t w)
184{
185 TAILQ_INIT(&w->wa_waitq);
186}
187
188/*
189 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
190 * or LWKT start address before it can be scheduled. If the target cpu is
191 * -1 the thread will be created on the current cpu.
0cfcada1
MD
192 *
193 * If you intend to create a thread without a process context this function
194 * does everything except load the startup and switcher function.
7d0bac62
MD
195 */
196thread_t
75cdbe6c 197lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 198{
99df837e 199 void *stack;
ef0fdad1 200 int flags = 0;
7d0bac62 201
ef0fdad1 202 if (td == NULL) {
26a0694b 203 crit_enter();
ef0fdad1
MD
204 if (mycpu->gd_tdfreecount > 0) {
205 --mycpu->gd_tdfreecount;
206 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 207 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
208 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
209 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
210 crit_exit();
211 stack = td->td_kstack;
212 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
213 } else {
214 crit_exit();
05220613 215#ifdef _KERNEL
ef0fdad1 216 td = zalloc(thread_zone);
05220613
MD
217#else
218 td = malloc(sizeof(struct thread));
219#endif
ef0fdad1
MD
220 td->td_kstack = NULL;
221 flags |= TDF_ALLOCATED_THREAD;
222 }
223 }
224 if ((stack = td->td_kstack) == NULL) {
05220613
MD
225#ifdef _KERNEL
226 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
227#else
fb04f4fd 228 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 229#endif
ef0fdad1 230 flags |= TDF_ALLOCATED_STACK;
99df837e 231 }
75cdbe6c
MD
232 if (cpu < 0)
233 lwkt_init_thread(td, stack, flags, mycpu);
234 else
235 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 236 return(td);
7d0bac62
MD
237}
238
239/*
240 * Initialize a preexisting thread structure. This function is used by
241 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
242 *
f8c3996b
MD
243 * All threads start out in a critical section at a priority of
244 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
245 * appropriate. This function may send an IPI message when the
246 * requested cpu is not the current cpu and consequently gd_tdallq may
247 * not be initialized synchronously from the point of view of the originating
248 * cpu.
249 *
250 * NOTE! we have to be careful in regards to creating threads for other cpus
251 * if SMP has not yet been activated.
7d0bac62 252 */
75cdbe6c
MD
253static void
254lwkt_init_thread_remote(void *arg)
255{
256 thread_t td = arg;
257
258 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
259}
260
2d93b37a
MD
261/* lwkt_init_thread has a userland override */
262#ifdef _KERNEL
263
7d0bac62 264void
26a0694b 265lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 266{
99df837e
MD
267 bzero(td, sizeof(struct thread));
268 td->td_kstack = stack;
269 td->td_flags |= flags;
26a0694b 270 td->td_gd = gd;
f8c3996b 271 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 272 lwkt_initport(&td->td_msgport, td);
99df837e 273 pmap_init_thread(td);
75cdbe6c
MD
274 if (smp_active == 0 || gd == mycpu) {
275 crit_enter();
276 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
277 crit_exit();
278 } else {
279 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
280 }
73e4f7b9
MD
281}
282
2d93b37a
MD
283#endif /* _KERNEL */
284
73e4f7b9
MD
285void
286lwkt_set_comm(thread_t td, const char *ctl, ...)
287{
e2565a42 288 __va_list va;
73e4f7b9 289
e2565a42 290 __va_start(va, ctl);
73e4f7b9 291 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 292 __va_end(va);
7d0bac62
MD
293}
294
99df837e 295void
73e4f7b9 296lwkt_hold(thread_t td)
99df837e 297{
73e4f7b9
MD
298 ++td->td_refs;
299}
300
301void
302lwkt_rele(thread_t td)
303{
304 KKASSERT(td->td_refs > 0);
305 --td->td_refs;
306}
307
c95cd171
MD
308#ifdef _KERNEL
309
73e4f7b9
MD
310void
311lwkt_wait_free(thread_t td)
312{
313 while (td->td_refs)
377d4740 314 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
315}
316
c95cd171
MD
317#endif
318
73e4f7b9
MD
319void
320lwkt_free_thread(thread_t td)
321{
322 struct globaldata *gd = mycpu;
323
d9eea1a5 324 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
325 ("lwkt_free_thread: did not exit! %p", td));
326
327 crit_enter();
73e4f7b9
MD
328 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
329 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
330 (td->td_flags & TDF_ALLOCATED_THREAD)
331 ) {
73e4f7b9
MD
332 ++gd->gd_tdfreecount;
333 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
334 crit_exit();
335 } else {
336 crit_exit();
337 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
338#ifdef _KERNEL
339 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
340#else
fb04f4fd 341 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 342#endif
73e4f7b9 343 /* gd invalid */
99df837e
MD
344 td->td_kstack = NULL;
345 }
05220613
MD
346 if (td->td_flags & TDF_ALLOCATED_THREAD) {
347#ifdef _KERNEL
99df837e 348 zfree(thread_zone, td);
05220613
MD
349#else
350 free(td);
351#endif
352 }
99df837e
MD
353 }
354}
355
356
8ad65e08
MD
357/*
358 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
359 * switch to the idlethread. Switching must occur within a critical
360 * section to avoid races with the scheduling queue.
361 *
362 * We always have full control over our cpu's run queue. Other cpus
363 * that wish to manipulate our queue must use the cpu_*msg() calls to
364 * talk to our cpu, so a critical section is all that is needed and
365 * the result is very, very fast thread switching.
366 *
96728c05
MD
367 * The LWKT scheduler uses a fixed priority model and round-robins at
368 * each priority level. User process scheduling is a totally
369 * different beast and LWKT priorities should not be confused with
370 * user process priorities.
f1d1c3fa 371 *
96728c05
MD
372 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
373 * cleans it up. Note that the td_switch() function cannot do anything that
374 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
375 * the target thread (not the current thread). It's nice to have a scheduler
376 * that does not need the MP lock to work because it allows us to do some
377 * really cool high-performance MP lock optimizations.
8ad65e08 378 */
96728c05 379
8ad65e08
MD
380void
381lwkt_switch(void)
382{
4b5f931b 383 struct globaldata *gd;
f1d1c3fa 384 thread_t td = curthread;
8ad65e08 385 thread_t ntd;
8a8d5d85
MD
386#ifdef SMP
387 int mpheld;
388#endif
8ad65e08 389
46a3f46d
MD
390 /*
391 * Switching from within a 'fast' (non thread switched) interrupt is
392 * illegal.
393 */
394 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 395 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 396 }
ef0fdad1 397
cb973d15
MD
398 /*
399 * Passive release (used to transition from user to kernel mode
400 * when we block or switch rather then when we enter the kernel).
401 * This function is NOT called if we are switching into a preemption
402 * or returning from a preemption. Typically this causes us to lose
403 * our P_CURPROC designation (if we have one) and become a true LWKT
404 * thread, and may also hand P_CURPROC to another process and schedule
405 * its thread.
406 */
407 if (td->td_release)
408 td->td_release(td);
409
f1d1c3fa 410 crit_enter();
4b5f931b 411 ++switch_count;
8a8d5d85
MD
412
413#ifdef SMP
414 /*
415 * td_mpcount cannot be used to determine if we currently hold the
416 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
417 * to get the lock, and switch out if it can't. Our ownership of
418 * the actual lock will remain stable while we are in a critical section
419 * (but, of course, another cpu may own or release the lock so the
420 * actual value of mp_lock is not stable).
8a8d5d85
MD
421 */
422 mpheld = MP_LOCK_HELD();
423#endif
99df837e
MD
424 if ((ntd = td->td_preempted) != NULL) {
425 /*
426 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
427 * thread. This occurs transparently, whether the preempted thread
428 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
429 * itself).
430 *
431 * We have to setup the MP lock for the original thread after backing
432 * out the adjustment that was made to curthread when the original
433 * was preempted.
99df837e 434 */
26a0694b 435 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 436#ifdef SMP
96728c05
MD
437 if (ntd->td_mpcount && mpheld == 0) {
438 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
439 td, ntd, td->td_mpcount, ntd->td_mpcount);
440 }
8a8d5d85
MD
441 if (ntd->td_mpcount) {
442 td->td_mpcount -= ntd->td_mpcount;
443 KKASSERT(td->td_mpcount >= 0);
444 }
445#endif
26a0694b 446 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 447 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 448 } else {
4b5f931b
MD
449 /*
450 * Priority queue / round-robin at each priority. Note that user
451 * processes run at a fixed, low priority and the user process
452 * scheduler deals with interactions between user processes
453 * by scheduling and descheduling them from the LWKT queue as
454 * necessary.
8a8d5d85
MD
455 *
456 * We have to adjust the MP lock for the target thread. If we
457 * need the MP lock and cannot obtain it we try to locate a
458 * thread that does not need the MP lock.
4b5f931b
MD
459 */
460 gd = mycpu;
4b5f931b
MD
461again:
462 if (gd->gd_runqmask) {
463 int nq = bsrl(gd->gd_runqmask);
464 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
465 gd->gd_runqmask &= ~(1 << nq);
466 goto again;
467 }
8a8d5d85
MD
468#ifdef SMP
469 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
470 /*
96728c05
MD
471 * Target needs MP lock and we couldn't get it, try
472 * to locate a thread which does not need the MP lock
3c23a41a 473 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
474 */
475 u_int32_t rqmask = gd->gd_runqmask;
476 while (rqmask) {
477 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
478 if (ntd->td_mpcount == 0)
479 break;
480 }
481 if (ntd)
482 break;
483 rqmask &= ~(1 << nq);
484 nq = bsrl(rqmask);
485 }
486 if (ntd == NULL) {
a2a5ad0d
MD
487 ntd = &gd->gd_idlethread;
488 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
489 } else {
490 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
491 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
492 }
493 } else {
494 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
495 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
496 }
497#else
4b5f931b
MD
498 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
499 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 500#endif
4b5f931b 501 } else {
3c23a41a
MD
502 /*
503 * Nothing to run but we may still need the BGL to deal with
504 * pending interrupts, spin in idle if so.
505 */
a2a5ad0d 506 ntd = &gd->gd_idlethread;
235957ed 507 if (gd->gd_reqflags)
3c23a41a 508 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 509 }
f1d1c3fa 510 }
26a0694b
MD
511 KASSERT(ntd->td_pri >= TDPRI_CRIT,
512 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
513
514 /*
515 * Do the actual switch. If the new target does not need the MP lock
516 * and we are holding it, release the MP lock. If the new target requires
517 * the MP lock we have already acquired it for the target.
518 */
519#ifdef SMP
520 if (ntd->td_mpcount == 0 ) {
521 if (MP_LOCK_HELD())
522 cpu_rel_mplock();
523 } else {
524 ASSERT_MP_LOCK_HELD();
525 }
526#endif
8a8d5d85 527 if (td != ntd) {
f1d1c3fa 528 td->td_switch(ntd);
8a8d5d85 529 }
96728c05 530
f1d1c3fa 531 crit_exit();
8ad65e08
MD
532}
533
cb973d15
MD
534/*
535 * Switch if another thread has a higher priority. Do not switch to other
536 * threads at the same priority.
537 */
538void
539lwkt_maybe_switch()
540{
541 struct globaldata *gd = mycpu;
542 struct thread *td = gd->gd_curthread;
543
544 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
545 lwkt_switch();
546 }
547}
548
b68b7282 549/*
96728c05
MD
550 * Request that the target thread preempt the current thread. Preemption
551 * only works under a specific set of conditions:
b68b7282 552 *
96728c05
MD
553 * - We are not preempting ourselves
554 * - The target thread is owned by the current cpu
555 * - We are not currently being preempted
556 * - The target is not currently being preempted
557 * - We are able to satisfy the target's MP lock requirements (if any).
558 *
559 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
560 * this is called via lwkt_schedule() through the td_preemptable callback.
561 * critpri is the managed critical priority that we should ignore in order
562 * to determine whether preemption is possible (aka usually just the crit
563 * priority of lwkt_schedule() itself).
b68b7282 564 *
26a0694b
MD
565 * XXX at the moment we run the target thread in a critical section during
566 * the preemption in order to prevent the target from taking interrupts
567 * that *WE* can't. Preemption is strictly limited to interrupt threads
568 * and interrupt-like threads, outside of a critical section, and the
569 * preempted source thread will be resumed the instant the target blocks
570 * whether or not the source is scheduled (i.e. preemption is supposed to
571 * be as transparent as possible).
4b5f931b 572 *
8a8d5d85
MD
573 * The target thread inherits our MP count (added to its own) for the
574 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
575 * MP lock during the preemption. Therefore, any preempting targets must be
576 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
577 * out of sync with the physical mp_lock, but we do not have to preserve
578 * the original ownership of the lock if it was out of synch (that is, we
579 * can leave it synchronized on return).
b68b7282
MD
580 */
581void
96728c05 582lwkt_preempt(thread_t ntd, int critpri)
b68b7282 583{
46a3f46d
MD
584 struct globaldata *gd = mycpu;
585 thread_t td = gd->gd_curthread;
8a8d5d85
MD
586#ifdef SMP
587 int mpheld;
57c254db 588 int savecnt;
8a8d5d85 589#endif
b68b7282 590
26a0694b 591 /*
96728c05
MD
592 * The caller has put us in a critical section. We can only preempt
593 * if the caller of the caller was not in a critical section (basically
57c254db
MD
594 * a local interrupt), as determined by the 'critpri' parameter. If
595 * we are unable to preempt
96728c05
MD
596 *
597 * YYY The target thread must be in a critical section (else it must
598 * inherit our critical section? I dunno yet).
26a0694b
MD
599 */
600 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 601
cb973d15 602 need_resched();
57c254db
MD
603 if (!_lwkt_wantresched(ntd, td)) {
604 ++preempt_miss;
605 return;
606 }
96728c05
MD
607 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
608 ++preempt_miss;
609 return;
610 }
611#ifdef SMP
46a3f46d 612 if (ntd->td_gd != gd) {
96728c05
MD
613 ++preempt_miss;
614 return;
615 }
616#endif
26a0694b
MD
617 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
618 ++preempt_weird;
619 return;
620 }
621 if (ntd->td_preempted) {
4b5f931b 622 ++preempt_hit;
26a0694b 623 return;
b68b7282 624 }
8a8d5d85 625#ifdef SMP
a2a5ad0d
MD
626 /*
627 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
628 * to or from the MP lock. In this case td_mpcount will be pre-disposed
629 * (non-zero) but not actually synchronized with the actual state of the
630 * lock. We can use it to imply an MP lock requirement for the
631 * preemption but we cannot use it to test whether we hold the MP lock
632 * or not.
a2a5ad0d 633 */
96728c05 634 savecnt = td->td_mpcount;
71ef2f5c 635 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
636 ntd->td_mpcount += td->td_mpcount;
637 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
638 ntd->td_mpcount -= td->td_mpcount;
639 ++preempt_miss;
640 return;
641 }
642#endif
26a0694b
MD
643
644 ++preempt_hit;
645 ntd->td_preempted = td;
646 td->td_flags |= TDF_PREEMPT_LOCK;
647 td->td_switch(ntd);
648 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
649#ifdef SMP
650 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
651 mpheld = MP_LOCK_HELD();
652 if (mpheld && td->td_mpcount == 0)
96728c05 653 cpu_rel_mplock();
71ef2f5c 654 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
655 panic("lwkt_preempt(): MP lock was not held through");
656#endif
26a0694b
MD
657 ntd->td_preempted = NULL;
658 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
659}
660
f1d1c3fa
MD
661/*
662 * Yield our thread while higher priority threads are pending. This is
663 * typically called when we leave a critical section but it can be safely
664 * called while we are in a critical section.
665 *
666 * This function will not generally yield to equal priority threads but it
667 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 668 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
669 * lwkt_yield_quick().
670 *
235957ed 671 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
672 * came along but was blocked and made pending.
673 *
f1d1c3fa
MD
674 * (self contained on a per cpu basis)
675 */
676void
677lwkt_yield_quick(void)
678{
7966cb69
MD
679 globaldata_t gd = mycpu;
680 thread_t td = gd->gd_curthread;
ef0fdad1 681
a2a5ad0d 682 /*
235957ed 683 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
684 * it with a non-zero cpl then we might not wind up calling splz after
685 * a task switch when the critical section is exited even though the
46a3f46d 686 * new task could accept the interrupt.
a2a5ad0d
MD
687 *
688 * XXX from crit_exit() only called after last crit section is released.
689 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
690 *
691 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
692 * except for this special case, we MUST call splz() here to handle any
693 * pending ints, particularly after we switch, or we might accidently
694 * halt the cpu with interrupts pending.
a2a5ad0d 695 */
46a3f46d 696 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 697 splz();
f1d1c3fa
MD
698
699 /*
700 * YYY enabling will cause wakeup() to task-switch, which really
701 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
702 * preemption and MP without actually doing preemption or MP, because a
703 * lot of code assumes that wakeup() does not block.
f1d1c3fa 704 */
46a3f46d
MD
705 if (untimely_switch && td->td_nest_count == 0 &&
706 gd->gd_intr_nesting_level == 0
707 ) {
f1d1c3fa
MD
708 crit_enter();
709 /*
710 * YYY temporary hacks until we disassociate the userland scheduler
711 * from the LWKT scheduler.
712 */
713 if (td->td_flags & TDF_RUNQ) {
714 lwkt_switch(); /* will not reenter yield function */
715 } else {
716 lwkt_schedule_self(); /* make sure we are scheduled */
717 lwkt_switch(); /* will not reenter yield function */
718 lwkt_deschedule_self(); /* make sure we are descheduled */
719 }
7966cb69 720 crit_exit_noyield(td);
f1d1c3fa 721 }
f1d1c3fa
MD
722}
723
8ad65e08 724/*
f1d1c3fa 725 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 726 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
727 * the crit_exit() call in lwkt_switch().
728 *
729 * (self contained on a per cpu basis)
8ad65e08
MD
730 */
731void
f1d1c3fa 732lwkt_yield(void)
8ad65e08 733{
f1d1c3fa
MD
734 lwkt_schedule_self();
735 lwkt_switch();
736}
737
738/*
739 * Schedule a thread to run. As the current thread we can always safely
740 * schedule ourselves, and a shortcut procedure is provided for that
741 * function.
742 *
743 * (non-blocking, self contained on a per cpu basis)
744 */
745void
746lwkt_schedule_self(void)
747{
748 thread_t td = curthread;
749
750 crit_enter();
751 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 752 _lwkt_enqueue(td);
05220613 753#ifdef _KERNEL
26a0694b
MD
754 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
755 panic("SCHED SELF PANIC");
05220613 756#endif
f1d1c3fa 757 crit_exit();
8ad65e08 758}
8ad65e08
MD
759
760/*
f1d1c3fa
MD
761 * Generic schedule. Possibly schedule threads belonging to other cpus and
762 * deal with threads that might be blocked on a wait queue.
763 *
96728c05 764 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
765 * Load balancing can be accomplished by requesting other sorts of actions
766 * for the thread in question.
8ad65e08
MD
767 */
768void
769lwkt_schedule(thread_t td)
770{
96728c05 771#ifdef INVARIANTS
26a0694b
MD
772 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
773 && td->td_proc->p_stat == SSLEEP
774 ) {
775 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
776 curthread,
777 curthread->td_proc ? curthread->td_proc->p_pid : -1,
778 curthread->td_proc ? curthread->td_proc->p_stat : -1,
779 td,
780 td->td_proc ? curthread->td_proc->p_pid : -1,
781 td->td_proc ? curthread->td_proc->p_stat : -1
782 );
783 panic("SCHED PANIC");
784 }
96728c05 785#endif
f1d1c3fa
MD
786 crit_enter();
787 if (td == curthread) {
788 _lwkt_enqueue(td);
789 } else {
790 lwkt_wait_t w;
791
792 /*
793 * If the thread is on a wait list we have to send our scheduling
794 * request to the owner of the wait structure. Otherwise we send
795 * the scheduling request to the cpu owning the thread. Races
796 * are ok, the target will forward the message as necessary (the
797 * message may chase the thread around before it finally gets
798 * acted upon).
799 *
800 * (remember, wait structures use stable storage)
801 */
802 if ((w = td->td_wait) != NULL) {
96728c05 803 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
804 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
805 --w->wa_count;
806 td->td_wait = NULL;
75cdbe6c 807 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 808 _lwkt_enqueue(td);
57c254db 809 if (td->td_preemptable) {
96728c05 810 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
811 } else if (_lwkt_wantresched(td, curthread)) {
812 need_resched();
813 }
f1d1c3fa 814 } else {
a72187e9 815 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 816 }
96728c05 817 lwkt_reltoken(&w->wa_token);
f1d1c3fa 818 } else {
96728c05 819 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
820 }
821 } else {
822 /*
823 * If the wait structure is NULL and we own the thread, there
824 * is no race (since we are in a critical section). If we
825 * do not own the thread there might be a race but the
826 * target cpu will deal with it.
827 */
75cdbe6c 828 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 829 _lwkt_enqueue(td);
57c254db 830 if (td->td_preemptable) {
96728c05 831 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
832 } else if (_lwkt_wantresched(td, curthread)) {
833 need_resched();
834 }
f1d1c3fa 835 } else {
a72187e9 836 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
837 }
838 }
8ad65e08 839 }
f1d1c3fa 840 crit_exit();
8ad65e08
MD
841}
842
d9eea1a5
MD
843/*
844 * Managed acquisition. This code assumes that the MP lock is held for
845 * the tdallq operation and that the thread has been descheduled from its
846 * original cpu. We also have to wait for the thread to be entirely switched
847 * out on its original cpu (this is usually fast enough that we never loop)
848 * since the LWKT system does not have to hold the MP lock while switching
849 * and the target may have released it before switching.
850 */
a2a5ad0d
MD
851void
852lwkt_acquire(thread_t td)
853{
854 struct globaldata *gd;
855
856 gd = td->td_gd;
857 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
858 while (td->td_flags & TDF_RUNNING) /* XXX spin */
859 ;
a2a5ad0d
MD
860 if (gd != mycpu) {
861 crit_enter();
862 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
863 gd = mycpu;
864 td->td_gd = gd;
a2a5ad0d
MD
865 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
866 crit_exit();
867 }
868}
869
8ad65e08 870/*
f1d1c3fa
MD
871 * Deschedule a thread.
872 *
873 * (non-blocking, self contained on a per cpu basis)
874 */
875void
876lwkt_deschedule_self(void)
877{
878 thread_t td = curthread;
879
880 crit_enter();
881 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
882 _lwkt_dequeue(td);
883 crit_exit();
884}
885
886/*
887 * Generic deschedule. Descheduling threads other then your own should be
888 * done only in carefully controlled circumstances. Descheduling is
889 * asynchronous.
890 *
891 * This function may block if the cpu has run out of messages.
8ad65e08
MD
892 */
893void
894lwkt_deschedule(thread_t td)
895{
f1d1c3fa
MD
896 crit_enter();
897 if (td == curthread) {
898 _lwkt_dequeue(td);
899 } else {
a72187e9 900 if (td->td_gd == mycpu) {
f1d1c3fa
MD
901 _lwkt_dequeue(td);
902 } else {
a72187e9 903 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
904 }
905 }
906 crit_exit();
907}
908
4b5f931b
MD
909/*
910 * Set the target thread's priority. This routine does not automatically
911 * switch to a higher priority thread, LWKT threads are not designed for
912 * continuous priority changes. Yield if you want to switch.
913 *
914 * We have to retain the critical section count which uses the high bits
26a0694b
MD
915 * of the td_pri field. The specified priority may also indicate zero or
916 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
917 */
918void
919lwkt_setpri(thread_t td, int pri)
920{
26a0694b 921 KKASSERT(pri >= 0);
a72187e9 922 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
923 crit_enter();
924 if (td->td_flags & TDF_RUNQ) {
925 _lwkt_dequeue(td);
926 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
927 _lwkt_enqueue(td);
928 } else {
929 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
930 }
931 crit_exit();
932}
933
934void
935lwkt_setpri_self(int pri)
936{
937 thread_t td = curthread;
938
4b5f931b
MD
939 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
940 crit_enter();
941 if (td->td_flags & TDF_RUNQ) {
942 _lwkt_dequeue(td);
943 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
944 _lwkt_enqueue(td);
945 } else {
946 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
947 }
948 crit_exit();
949}
950
951struct proc *
952lwkt_preempted_proc(void)
953{
73e4f7b9 954 thread_t td = curthread;
4b5f931b
MD
955 while (td->td_preempted)
956 td = td->td_preempted;
957 return(td->td_proc);
958}
959
ece04fd0
MD
960typedef struct lwkt_gettoken_req {
961 lwkt_token_t tok;
962 int cpu;
963} lwkt_gettoken_req;
964
965#if 0
4b5f931b 966
f1d1c3fa
MD
967/*
968 * This function deschedules the current thread and blocks on the specified
969 * wait queue. We obtain ownership of the wait queue in order to block
970 * on it. A generation number is used to interlock the wait queue in case
971 * it gets signalled while we are blocked waiting on the token.
972 *
973 * Note: alternatively we could dequeue our thread and then message the
974 * target cpu owning the wait queue. YYY implement as sysctl.
975 *
976 * Note: wait queue signals normally ping-pong the cpu as an optimization.
977 */
96728c05 978
f1d1c3fa 979void
ae8050a4 980lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
981{
982 thread_t td = curthread;
f1d1c3fa 983
f1d1c3fa 984 lwkt_gettoken(&w->wa_token);
ae8050a4 985 if (w->wa_gen == *gen) {
f1d1c3fa
MD
986 _lwkt_dequeue(td);
987 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
988 ++w->wa_count;
989 td->td_wait = w;
ae8050a4 990 td->td_wmesg = wmesg;
ece04fd0 991again:
f1d1c3fa 992 lwkt_switch();
ece04fd0
MD
993 lwkt_regettoken(&w->wa_token);
994 if (td->td_wmesg != NULL) {
995 _lwkt_dequeue(td);
996 goto again;
997 }
8ad65e08 998 }
ae8050a4
MD
999 /* token might be lost, doesn't matter for gen update */
1000 *gen = w->wa_gen;
f1d1c3fa
MD
1001 lwkt_reltoken(&w->wa_token);
1002}
1003
1004/*
1005 * Signal a wait queue. We gain ownership of the wait queue in order to
1006 * signal it. Once a thread is removed from the wait queue we have to
1007 * deal with the cpu owning the thread.
1008 *
1009 * Note: alternatively we could message the target cpu owning the wait
1010 * queue. YYY implement as sysctl.
1011 */
1012void
ece04fd0 1013lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1014{
1015 thread_t td;
1016 int count;
1017
1018 lwkt_gettoken(&w->wa_token);
1019 ++w->wa_gen;
ece04fd0
MD
1020 if (count < 0)
1021 count = w->wa_count;
f1d1c3fa
MD
1022 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1023 --count;
1024 --w->wa_count;
1025 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1026 td->td_wait = NULL;
ae8050a4 1027 td->td_wmesg = NULL;
a72187e9 1028 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1029 _lwkt_enqueue(td);
1030 } else {
a72187e9 1031 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1032 }
1033 lwkt_regettoken(&w->wa_token);
1034 }
1035 lwkt_reltoken(&w->wa_token);
1036}
1037
ece04fd0
MD
1038#endif
1039
f1d1c3fa 1040/*
96728c05 1041 * Acquire ownership of a token
f1d1c3fa 1042 *
96728c05 1043 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
1044 * section side effects, depending on its purpose. These side effects
1045 * guarentee that you will maintain ownership of the token as long as you
1046 * do not block. If you block you may lose access to the token (but you
1047 * must still release it even if you lose your access to it).
1048 *
96728c05 1049 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 1050 * a token, but do we really only need to disable IPIs ?
96728c05
MD
1051 *
1052 * YYY certain tokens could be made to act like mutexes when performance
1053 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1054 *
a2a5ad0d
MD
1055 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1056 * means that 4.x does not expect a switch so for now we cannot switch
1057 * when waiting for an IPI to be returned.
1058 *
1059 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
1060 * it and then block. The IPI causes the token to be given away to the
1061 * requesting cpu, unless it has already changed hands. Since only the
1062 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 1063 * This needs serious optimization.
f1d1c3fa 1064 */
57c254db
MD
1065
1066#ifdef SMP
1067
96728c05
MD
1068static
1069void
1070lwkt_gettoken_remote(void *arg)
1071{
1072 lwkt_gettoken_req *req = arg;
1073 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 1074#ifdef INVARIANTS
a2a5ad0d
MD
1075 if (token_debug)
1076 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 1077#endif
96728c05 1078 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
1079 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1080 /* else set reqcpu to point to current cpu for release */
96728c05
MD
1081 }
1082}
1083
57c254db
MD
1084#endif
1085
8a8d5d85 1086int
f1d1c3fa
MD
1087lwkt_gettoken(lwkt_token_t tok)
1088{
1089 /*
1090 * Prevent preemption so the token can't be taken away from us once
1091 * we gain ownership of it. Use a synchronous request which might
1092 * block. The request will be forwarded as necessary playing catchup
1093 * to the token.
1094 */
96728c05 1095
f1d1c3fa 1096 crit_enter();
57c254db 1097#ifdef INVARIANTS
1d4c9574
MD
1098 if (curthread->td_pri > 1800) {
1099 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1100 tok, ((int **)&tok)[-1]);
1101 }
a2a5ad0d
MD
1102 if (curthread->td_pri > 2000) {
1103 curthread->td_pri = 1000;
1104 panic("too HIGH!");
57c254db
MD
1105 }
1106#endif
96728c05 1107#ifdef SMP
d0e06f83 1108 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1109 struct lwkt_gettoken_req req;
1110 int seq;
96728c05
MD
1111 int dcpu;
1112
1113 req.cpu = mycpu->gd_cpuid;
1114 req.tok = tok;
1115 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1116 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1117#ifdef INVARIANTS
a2a5ad0d
MD
1118 if (token_debug)
1119 printf("REQT%d ", dcpu);
634081ff 1120#endif
96728c05
MD
1121 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1122 lwkt_wait_ipiq(dcpu, seq);
634081ff 1123#ifdef INVARIANTS
a2a5ad0d
MD
1124 if (token_debug)
1125 printf("REQR%d ", tok->t_cpu);
634081ff 1126#endif
f1d1c3fa
MD
1127 }
1128#endif
1129 /*
1130 * leave us in a critical section on return. This will be undone
8a8d5d85 1131 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1132 */
8a8d5d85 1133 return(++tok->t_gen);
f1d1c3fa
MD
1134}
1135
96728c05
MD
1136/*
1137 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1138 * failure.
1139 */
1140int
1141lwkt_trytoken(lwkt_token_t tok)
1142{
1143 crit_enter();
1144#ifdef SMP
1145 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1146 crit_exit();
96728c05
MD
1147 return(0);
1148 }
1149#endif
1150 /* leave us in the critical section */
1151 ++tok->t_gen;
1152 return(1);
1153}
1154
f1d1c3fa
MD
1155/*
1156 * Release your ownership of a token. Releases must occur in reverse
1157 * order to aquisitions, eventually so priorities can be unwound properly
1158 * like SPLs. At the moment the actual implemention doesn't care.
1159 *
1160 * We can safely hand a token that we own to another cpu without notifying
1161 * it, but once we do we can't get it back without requesting it (unless
1162 * the other cpu hands it back to us before we check).
1163 *
1164 * We might have lost the token, so check that.
7ba9c17c
MD
1165 *
1166 * Return the token's generation number. The number is useful to callers
1167 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1168 */
7ba9c17c 1169int
f1d1c3fa
MD
1170lwkt_reltoken(lwkt_token_t tok)
1171{
7ba9c17c
MD
1172 int gen;
1173
d0e06f83 1174 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1175 tok->t_cpu = tok->t_reqcpu;
1176 }
7ba9c17c 1177 gen = tok->t_gen;
f1d1c3fa 1178 crit_exit();
7ba9c17c 1179 return(gen);
f1d1c3fa
MD
1180}
1181
1182/*
7ba9c17c
MD
1183 * Reacquire a token that might have been lost. 0 is returned if the
1184 * generation has not changed (nobody stole the token from us), -1 is
1185 * returned otherwise. The token is reacquired regardless but the
1186 * generation number is not bumped further if we already own the token.
8a8d5d85 1187 *
7ba9c17c
MD
1188 * For efficiency we inline the best-case situation for lwkt_regettoken()
1189 * (i.e .we still own the token).
8a8d5d85
MD
1190 */
1191int
1192lwkt_gentoken(lwkt_token_t tok, int *gen)
1193{
7ba9c17c 1194 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1195 return(0);
7ba9c17c
MD
1196 *gen = lwkt_regettoken(tok);
1197 return(-1);
8a8d5d85
MD
1198}
1199
8a8d5d85 1200/*
435ff993
MD
1201 * Re-acquire a token that might have been lost. The generation number
1202 * is bumped and returned regardless of whether the token had been lost
1203 * or not (because we only have cpu granularity we have to bump the token
1204 * either way).
f1d1c3fa
MD
1205 */
1206int
1207lwkt_regettoken(lwkt_token_t tok)
1208{
96728c05 1209 /* assert we are in a critical section */
d0e06f83 1210 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1211#ifdef SMP
d0e06f83 1212 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1213 struct lwkt_gettoken_req req;
1214 int seq;
96728c05 1215 int dcpu;
57c254db 1216
96728c05
MD
1217 req.cpu = mycpu->gd_cpuid;
1218 req.tok = tok;
1219 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1220 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1221#ifdef INVARIANTS
cb973d15
MD
1222 if (token_debug)
1223 printf("REQT%d ", dcpu);
634081ff 1224#endif
96728c05
MD
1225 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1226 lwkt_wait_ipiq(dcpu, seq);
634081ff 1227#ifdef INVARIATNS
cb973d15
MD
1228 if (token_debug)
1229 printf("REQR%d ", tok->t_cpu);
634081ff 1230#endif
f1d1c3fa 1231 }
f1d1c3fa 1232#endif
96728c05 1233 }
435ff993 1234 ++tok->t_gen;
8a8d5d85 1235 return(tok->t_gen);
8ad65e08
MD
1236}
1237
72740893
MD
1238void
1239lwkt_inittoken(lwkt_token_t tok)
1240{
1241 /*
1242 * Zero structure and set cpu owner and reqcpu to cpu 0.
1243 */
1244 bzero(tok, sizeof(*tok));
1245}
1246
99df837e
MD
1247/*
1248 * Create a kernel process/thread/whatever. It shares it's address space
1249 * with proc0 - ie: kernel only.
1250 *
365fa13f
MD
1251 * NOTE! By default new threads are created with the MP lock held. A
1252 * thread which does not require the MP lock should release it by calling
1253 * rel_mplock() at the start of the new thread.
99df837e
MD
1254 */
1255int
1256lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1257 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1258 const char *fmt, ...)
99df837e 1259{
73e4f7b9 1260 thread_t td;
e2565a42 1261 __va_list ap;
99df837e 1262
75cdbe6c 1263 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1264 if (tdp)
1265 *tdp = td;
99df837e 1266 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1267 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1268#ifdef SMP
1269 td->td_mpcount = 1;
1270#endif
99df837e
MD
1271
1272 /*
1273 * Set up arg0 for 'ps' etc
1274 */
e2565a42 1275 __va_start(ap, fmt);
99df837e 1276 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1277 __va_end(ap);
99df837e
MD
1278
1279 /*
1280 * Schedule the thread to run
1281 */
ef0fdad1
MD
1282 if ((td->td_flags & TDF_STOPREQ) == 0)
1283 lwkt_schedule(td);
1284 else
1285 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1286 return 0;
1287}
1288
2d93b37a
MD
1289/*
1290 * lwkt_exit() has a userland override.
1291 * kthread_* is specific to the kernel and is not needed by userland.
1292 */
1293#ifdef _KERNEL
1294
99df837e
MD
1295/*
1296 * Destroy an LWKT thread. Warning! This function is not called when
1297 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1298 * uses a different reaping mechanism.
1299 */
1300void
1301lwkt_exit(void)
1302{
1303 thread_t td = curthread;
1304
1305 if (td->td_flags & TDF_VERBOSE)
1306 printf("kthread %p %s has exited\n", td, td->td_comm);
1307 crit_enter();
1308 lwkt_deschedule_self();
1309 ++mycpu->gd_tdfreecount;
1310 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1311 cpu_thread_exit();
1312}
1313
1314/*
1315 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1316 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1317 *
1318 * NOTE! By default kthreads are created with the MP lock held. A
1319 * thread which does not require the MP lock should release it by calling
1320 * rel_mplock() at the start of the new thread.
99df837e
MD
1321 */
1322int
1323kthread_create(void (*func)(void *), void *arg,
1324 struct thread **tdp, const char *fmt, ...)
1325{
73e4f7b9 1326 thread_t td;
e2565a42 1327 __va_list ap;
99df837e 1328
75cdbe6c 1329 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1330 if (tdp)
1331 *tdp = td;
99df837e
MD
1332 cpu_set_thread_handler(td, kthread_exit, func, arg);
1333 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1334#ifdef SMP
1335 td->td_mpcount = 1;
1336#endif
99df837e
MD
1337
1338 /*
1339 * Set up arg0 for 'ps' etc
1340 */
e2565a42 1341 __va_start(ap, fmt);
99df837e 1342 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1343 __va_end(ap);
99df837e
MD
1344
1345 /*
1346 * Schedule the thread to run
1347 */
1348 lwkt_schedule(td);
1349 return 0;
1350}
1351
1352/*
1353 * Destroy an LWKT thread. Warning! This function is not called when
1354 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1355 * uses a different reaping mechanism.
1356 *
1357 * XXX duplicates lwkt_exit()
1358 */
1359void
1360kthread_exit(void)
1361{
1362 lwkt_exit();
1363}
1364
2d93b37a
MD
1365#endif /* _KERNEL */
1366
1367void
1368crit_panic(void)
1369{
1370 thread_t td = curthread;
1371 int lpri = td->td_pri;
1372
1373 td->td_pri = 0;
1374 panic("td_pri is/would-go negative! %p %d", td, lpri);
1375}
1376
96728c05
MD
1377#ifdef SMP
1378
1379/*
1380 * Send a function execution request to another cpu. The request is queued
1381 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1382 * possible target cpu. The FIFO can be written.
1383 *
1384 * YYY If the FIFO fills up we have to enable interrupts and process the
1385 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1386 * Create a CPU_*() function to do this!
1387 *
46a3f46d
MD
1388 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1389 * end will take care of any pending interrupts.
1390 *
96728c05
MD
1391 * Must be called from a critical section.
1392 */
1393int
1394lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1395{
1396 lwkt_ipiq_t ip;
1397 int windex;
a2a5ad0d 1398 struct globaldata *gd = mycpu;
96728c05 1399
a2a5ad0d 1400 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1401 func(arg);
1402 return(0);
1403 }
cb973d15 1404 crit_enter();
a2a5ad0d
MD
1405 ++gd->gd_intr_nesting_level;
1406#ifdef INVARIANTS
1407 if (gd->gd_intr_nesting_level > 20)
1408 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1409#endif
96728c05
MD
1410 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1411 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1412 ++ipiq_count;
a2a5ad0d 1413 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1414
1415 /*
1416 * We always drain before the FIFO becomes full so it should never
1417 * become full. We need to leave enough entries to deal with
1418 * reentrancy.
1419 */
1420 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1421 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1422 ip->ip_func[windex] = func;
1423 ip->ip_arg[windex] = arg;
1424 /* YYY memory barrier */
1425 ++ip->ip_windex;
96728c05
MD
1426 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1427 unsigned int eflags = read_eflags();
1428 cpu_enable_intr();
1429 ++ipiq_fifofull;
cb973d15 1430 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1431 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1432 lwkt_process_ipiq();
1433 }
1434 write_eflags(eflags);
1435 }
a2a5ad0d 1436 --gd->gd_intr_nesting_level;
96728c05 1437 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1438 crit_exit();
96728c05
MD
1439 return(ip->ip_windex);
1440}
1441
cb973d15
MD
1442/*
1443 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1444 * The message will not be sent to stopped cpus.
cb973d15
MD
1445 */
1446void
1447lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1448{
1449 int cpuid;
1450
435ff993 1451 mask &= ~stopped_cpus;
cb973d15
MD
1452 while (mask) {
1453 cpuid = bsfl(mask);
1454 lwkt_send_ipiq(cpuid, func, arg);
1455 mask &= ~(1 << cpuid);
1456 }
1457}
1458
96728c05
MD
1459/*
1460 * Wait for the remote cpu to finish processing a function.
1461 *
1462 * YYY we have to enable interrupts and process the IPIQ while waiting
1463 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1464 * function to do this! YYY we really should 'block' here.
1465 *
1466 * Must be called from a critical section. Thsi routine may be called
1467 * from an interrupt (for example, if an interrupt wakes a foreign thread
1468 * up).
1469 */
1470void
1471lwkt_wait_ipiq(int dcpu, int seq)
1472{
1473 lwkt_ipiq_t ip;
a2a5ad0d 1474 int maxc = 100000000;
96728c05
MD
1475
1476 if (dcpu != mycpu->gd_cpuid) {
1477 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1478 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1479 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1480 unsigned int eflags = read_eflags();
1481 cpu_enable_intr();
cb973d15 1482 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1483 lwkt_process_ipiq();
a2a5ad0d 1484 if (--maxc == 0)
cb973d15 1485 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1486 if (maxc < -1000000)
1487 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1488 }
1489 write_eflags(eflags);
1490 }
1491 }
1492}
1493
1494/*
1495 * Called from IPI interrupt (like a fast interrupt), which has placed
1496 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1497 * May also be called from doreti or splz, or be reentrantly called
1498 * indirectly through the ip_func[] we run.
96728c05
MD
1499 */
1500void
1501lwkt_process_ipiq(void)
1502{
1503 int n;
1504 int cpuid = mycpu->gd_cpuid;
1505
1506 for (n = 0; n < ncpus; ++n) {
1507 lwkt_ipiq_t ip;
1508 int ri;
1509
1510 if (n == cpuid)
1511 continue;
1512 ip = globaldata_find(n)->gd_ipiq;
1513 if (ip == NULL)
1514 continue;
1515 ip = &ip[cpuid];
cb973d15
MD
1516
1517 /*
1518 * Note: xindex is only updated after we are sure the function has
1519 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1520 * function may send an IPI which may block/drain.
1521 */
96728c05
MD
1522 while (ip->ip_rindex != ip->ip_windex) {
1523 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1524 ++ip->ip_rindex;
cb973d15
MD
1525 ip->ip_func[ri](ip->ip_arg[ri]);
1526 /* YYY memory barrier */
1527 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1528 }
1529 }
1530}
1531
1532#else
1533
1534int
1535lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1536{
1537 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1538 return(0); /* NOT REACHED */
1539}
1540
1541void
1542lwkt_wait_ipiq(int dcpu, int seq)
1543{
1544 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1545}
1546
1547#endif