Fix the OID_AUTO collision with static sysctl numbers. This can occur
[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 *
05220613 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.42 2003/11/21 22:46:08 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>
74#include <liblwkt/thread.h>
75#include <sys/thread.h>
76#include <sys/msgport.h>
77#include <sys/errno.h>
78#include <liblwkt/globaldata.h>
79#include <sys/thread2.h>
80#include <sys/msgport2.h>
81#include <stdlib.h>
82
83#endif
84
7d0bac62 85static int untimely_switch = 0;
57c254db
MD
86#ifdef INVARIANTS
87static int token_debug = 0;
05220613
MD
88#endif
89static __int64_t switch_count = 0;
90static __int64_t preempt_hit = 0;
91static __int64_t preempt_miss = 0;
92static __int64_t preempt_weird = 0;
93static __int64_t ipiq_count = 0;
94static __int64_t ipiq_fifofull = 0;
95
96#ifdef _KERNEL
97
98SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
99#ifdef INVARIANTS
57c254db
MD
100SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
101#endif
4b5f931b 102SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 103SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 104SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 105SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05 106SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
96728c05 107SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 108
05220613
MD
109#endif
110
4b5f931b
MD
111/*
112 * These helper procedures handle the runq, they can only be called from
113 * within a critical section.
75cdbe6c
MD
114 *
115 * WARNING! Prior to SMP being brought up it is possible to enqueue and
116 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
117 * instead of 'mycpu' when referencing the globaldata structure. Once
118 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 119 */
f1d1c3fa
MD
120static __inline
121void
122_lwkt_dequeue(thread_t td)
123{
124 if (td->td_flags & TDF_RUNQ) {
4b5f931b 125 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 126 struct globaldata *gd = td->td_gd;
4b5f931b 127
f1d1c3fa 128 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
129 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
130 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
131 }
132}
133
134static __inline
135void
136_lwkt_enqueue(thread_t td)
137{
138 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 139 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 140 struct globaldata *gd = td->td_gd;
4b5f931b 141
f1d1c3fa 142 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
143 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
144 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
145 }
146}
8ad65e08 147
57c254db
MD
148static __inline
149int
150_lwkt_wantresched(thread_t ntd, thread_t cur)
151{
152 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
153}
154
8ad65e08
MD
155/*
156 * LWKTs operate on a per-cpu basis
157 *
73e4f7b9 158 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
159 */
160void
161lwkt_gdinit(struct globaldata *gd)
162{
4b5f931b
MD
163 int i;
164
165 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
166 TAILQ_INIT(&gd->gd_tdrunq[i]);
167 gd->gd_runqmask = 0;
73e4f7b9 168 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
169}
170
7d0bac62
MD
171/*
172 * Initialize a thread wait structure prior to first use.
173 *
174 * NOTE! called from low level boot code, we cannot do anything fancy!
175 */
176void
177lwkt_init_wait(lwkt_wait_t w)
178{
179 TAILQ_INIT(&w->wa_waitq);
180}
181
182/*
183 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
184 * or LWKT start address before it can be scheduled. If the target cpu is
185 * -1 the thread will be created on the current cpu.
0cfcada1
MD
186 *
187 * If you intend to create a thread without a process context this function
188 * does everything except load the startup and switcher function.
7d0bac62
MD
189 */
190thread_t
75cdbe6c 191lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 192{
99df837e 193 void *stack;
ef0fdad1 194 int flags = 0;
7d0bac62 195
ef0fdad1 196 if (td == NULL) {
26a0694b 197 crit_enter();
ef0fdad1
MD
198 if (mycpu->gd_tdfreecount > 0) {
199 --mycpu->gd_tdfreecount;
200 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 201 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
202 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
203 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
204 crit_exit();
205 stack = td->td_kstack;
206 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
207 } else {
208 crit_exit();
05220613 209#ifdef _KERNEL
ef0fdad1 210 td = zalloc(thread_zone);
05220613
MD
211#else
212 td = malloc(sizeof(struct thread));
213#endif
ef0fdad1
MD
214 td->td_kstack = NULL;
215 flags |= TDF_ALLOCATED_THREAD;
216 }
217 }
218 if ((stack = td->td_kstack) == NULL) {
05220613
MD
219#ifdef _KERNEL
220 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
221#else
222 stack = liblwkt_alloc_stack(THREAD_STACK);
223#endif
ef0fdad1 224 flags |= TDF_ALLOCATED_STACK;
99df837e 225 }
75cdbe6c
MD
226 if (cpu < 0)
227 lwkt_init_thread(td, stack, flags, mycpu);
228 else
229 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 230 return(td);
7d0bac62
MD
231}
232
233/*
234 * Initialize a preexisting thread structure. This function is used by
235 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
236 *
f8c3996b
MD
237 * All threads start out in a critical section at a priority of
238 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
239 * appropriate. This function may send an IPI message when the
240 * requested cpu is not the current cpu and consequently gd_tdallq may
241 * not be initialized synchronously from the point of view of the originating
242 * cpu.
243 *
244 * NOTE! we have to be careful in regards to creating threads for other cpus
245 * if SMP has not yet been activated.
7d0bac62 246 */
75cdbe6c
MD
247static void
248lwkt_init_thread_remote(void *arg)
249{
250 thread_t td = arg;
251
252 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
253}
254
7d0bac62 255void
26a0694b 256lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 257{
99df837e
MD
258 bzero(td, sizeof(struct thread));
259 td->td_kstack = stack;
260 td->td_flags |= flags;
26a0694b 261 td->td_gd = gd;
f8c3996b 262 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
ece04fd0 263 lwkt_init_port(&td->td_msgport, td);
99df837e 264 pmap_init_thread(td);
75cdbe6c
MD
265 if (smp_active == 0 || gd == mycpu) {
266 crit_enter();
267 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
268 crit_exit();
269 } else {
270 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
271 }
73e4f7b9
MD
272}
273
274void
275lwkt_set_comm(thread_t td, const char *ctl, ...)
276{
e2565a42 277 __va_list va;
73e4f7b9 278
e2565a42 279 __va_start(va, ctl);
73e4f7b9 280 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 281 __va_end(va);
7d0bac62
MD
282}
283
99df837e 284void
73e4f7b9 285lwkt_hold(thread_t td)
99df837e 286{
73e4f7b9
MD
287 ++td->td_refs;
288}
289
290void
291lwkt_rele(thread_t td)
292{
293 KKASSERT(td->td_refs > 0);
294 --td->td_refs;
295}
296
297void
298lwkt_wait_free(thread_t td)
299{
300 while (td->td_refs)
377d4740 301 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
302}
303
304void
305lwkt_free_thread(thread_t td)
306{
307 struct globaldata *gd = mycpu;
308
d9eea1a5 309 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
310 ("lwkt_free_thread: did not exit! %p", td));
311
312 crit_enter();
73e4f7b9
MD
313 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
314 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
315 (td->td_flags & TDF_ALLOCATED_THREAD)
316 ) {
73e4f7b9
MD
317 ++gd->gd_tdfreecount;
318 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
319 crit_exit();
320 } else {
321 crit_exit();
322 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
323#ifdef _KERNEL
324 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
325#else
326 liblwkt_free_stack(td->td_kstack, THREAD_STACK);
327#endif
73e4f7b9 328 /* gd invalid */
99df837e
MD
329 td->td_kstack = NULL;
330 }
05220613
MD
331 if (td->td_flags & TDF_ALLOCATED_THREAD) {
332#ifdef _KERNEL
99df837e 333 zfree(thread_zone, td);
05220613
MD
334#else
335 free(td);
336#endif
337 }
99df837e
MD
338 }
339}
340
341
8ad65e08
MD
342/*
343 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
344 * switch to the idlethread. Switching must occur within a critical
345 * section to avoid races with the scheduling queue.
346 *
347 * We always have full control over our cpu's run queue. Other cpus
348 * that wish to manipulate our queue must use the cpu_*msg() calls to
349 * talk to our cpu, so a critical section is all that is needed and
350 * the result is very, very fast thread switching.
351 *
96728c05
MD
352 * The LWKT scheduler uses a fixed priority model and round-robins at
353 * each priority level. User process scheduling is a totally
354 * different beast and LWKT priorities should not be confused with
355 * user process priorities.
f1d1c3fa 356 *
96728c05
MD
357 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
358 * cleans it up. Note that the td_switch() function cannot do anything that
359 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
360 * the target thread (not the current thread). It's nice to have a scheduler
361 * that does not need the MP lock to work because it allows us to do some
362 * really cool high-performance MP lock optimizations.
8ad65e08 363 */
96728c05 364
8ad65e08
MD
365void
366lwkt_switch(void)
367{
4b5f931b 368 struct globaldata *gd;
f1d1c3fa 369 thread_t td = curthread;
8ad65e08 370 thread_t ntd;
8a8d5d85
MD
371#ifdef SMP
372 int mpheld;
373#endif
8ad65e08 374
46a3f46d
MD
375 /*
376 * Switching from within a 'fast' (non thread switched) interrupt is
377 * illegal.
378 */
379 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 380 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 381 }
ef0fdad1 382
cb973d15
MD
383 /*
384 * Passive release (used to transition from user to kernel mode
385 * when we block or switch rather then when we enter the kernel).
386 * This function is NOT called if we are switching into a preemption
387 * or returning from a preemption. Typically this causes us to lose
388 * our P_CURPROC designation (if we have one) and become a true LWKT
389 * thread, and may also hand P_CURPROC to another process and schedule
390 * its thread.
391 */
392 if (td->td_release)
393 td->td_release(td);
394
f1d1c3fa 395 crit_enter();
4b5f931b 396 ++switch_count;
8a8d5d85
MD
397
398#ifdef SMP
399 /*
400 * td_mpcount cannot be used to determine if we currently hold the
401 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
402 * to get the lock, and switch out if it can't. Our ownership of
403 * the actual lock will remain stable while we are in a critical section
404 * (but, of course, another cpu may own or release the lock so the
405 * actual value of mp_lock is not stable).
8a8d5d85
MD
406 */
407 mpheld = MP_LOCK_HELD();
408#endif
99df837e
MD
409 if ((ntd = td->td_preempted) != NULL) {
410 /*
411 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
412 * thread. This occurs transparently, whether the preempted thread
413 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
414 * itself).
415 *
416 * We have to setup the MP lock for the original thread after backing
417 * out the adjustment that was made to curthread when the original
418 * was preempted.
99df837e 419 */
26a0694b 420 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 421#ifdef SMP
96728c05
MD
422 if (ntd->td_mpcount && mpheld == 0) {
423 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
424 td, ntd, td->td_mpcount, ntd->td_mpcount);
425 }
8a8d5d85
MD
426 if (ntd->td_mpcount) {
427 td->td_mpcount -= ntd->td_mpcount;
428 KKASSERT(td->td_mpcount >= 0);
429 }
430#endif
26a0694b 431 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 432 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 433 } else {
4b5f931b
MD
434 /*
435 * Priority queue / round-robin at each priority. Note that user
436 * processes run at a fixed, low priority and the user process
437 * scheduler deals with interactions between user processes
438 * by scheduling and descheduling them from the LWKT queue as
439 * necessary.
8a8d5d85
MD
440 *
441 * We have to adjust the MP lock for the target thread. If we
442 * need the MP lock and cannot obtain it we try to locate a
443 * thread that does not need the MP lock.
4b5f931b
MD
444 */
445 gd = mycpu;
4b5f931b
MD
446again:
447 if (gd->gd_runqmask) {
448 int nq = bsrl(gd->gd_runqmask);
449 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
450 gd->gd_runqmask &= ~(1 << nq);
451 goto again;
452 }
8a8d5d85
MD
453#ifdef SMP
454 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
455 /*
96728c05
MD
456 * Target needs MP lock and we couldn't get it, try
457 * to locate a thread which does not need the MP lock
3c23a41a 458 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
459 */
460 u_int32_t rqmask = gd->gd_runqmask;
461 while (rqmask) {
462 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
463 if (ntd->td_mpcount == 0)
464 break;
465 }
466 if (ntd)
467 break;
468 rqmask &= ~(1 << nq);
469 nq = bsrl(rqmask);
470 }
471 if (ntd == NULL) {
a2a5ad0d
MD
472 ntd = &gd->gd_idlethread;
473 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
474 } else {
475 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
476 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
477 }
478 } else {
479 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
480 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
481 }
482#else
4b5f931b
MD
483 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
484 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 485#endif
4b5f931b 486 } else {
3c23a41a
MD
487 /*
488 * Nothing to run but we may still need the BGL to deal with
489 * pending interrupts, spin in idle if so.
490 */
a2a5ad0d 491 ntd = &gd->gd_idlethread;
235957ed 492 if (gd->gd_reqflags)
3c23a41a 493 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 494 }
f1d1c3fa 495 }
26a0694b
MD
496 KASSERT(ntd->td_pri >= TDPRI_CRIT,
497 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
498
499 /*
500 * Do the actual switch. If the new target does not need the MP lock
501 * and we are holding it, release the MP lock. If the new target requires
502 * the MP lock we have already acquired it for the target.
503 */
504#ifdef SMP
505 if (ntd->td_mpcount == 0 ) {
506 if (MP_LOCK_HELD())
507 cpu_rel_mplock();
508 } else {
509 ASSERT_MP_LOCK_HELD();
510 }
511#endif
8a8d5d85 512 if (td != ntd) {
f1d1c3fa 513 td->td_switch(ntd);
8a8d5d85 514 }
96728c05 515
f1d1c3fa 516 crit_exit();
8ad65e08
MD
517}
518
cb973d15
MD
519/*
520 * Switch if another thread has a higher priority. Do not switch to other
521 * threads at the same priority.
522 */
523void
524lwkt_maybe_switch()
525{
526 struct globaldata *gd = mycpu;
527 struct thread *td = gd->gd_curthread;
528
529 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
530 lwkt_switch();
531 }
532}
533
b68b7282 534/*
96728c05
MD
535 * Request that the target thread preempt the current thread. Preemption
536 * only works under a specific set of conditions:
b68b7282 537 *
96728c05
MD
538 * - We are not preempting ourselves
539 * - The target thread is owned by the current cpu
540 * - We are not currently being preempted
541 * - The target is not currently being preempted
542 * - We are able to satisfy the target's MP lock requirements (if any).
543 *
544 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
545 * this is called via lwkt_schedule() through the td_preemptable callback.
546 * critpri is the managed critical priority that we should ignore in order
547 * to determine whether preemption is possible (aka usually just the crit
548 * priority of lwkt_schedule() itself).
b68b7282 549 *
26a0694b
MD
550 * XXX at the moment we run the target thread in a critical section during
551 * the preemption in order to prevent the target from taking interrupts
552 * that *WE* can't. Preemption is strictly limited to interrupt threads
553 * and interrupt-like threads, outside of a critical section, and the
554 * preempted source thread will be resumed the instant the target blocks
555 * whether or not the source is scheduled (i.e. preemption is supposed to
556 * be as transparent as possible).
4b5f931b 557 *
8a8d5d85
MD
558 * The target thread inherits our MP count (added to its own) for the
559 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
560 * MP lock during the preemption. Therefore, any preempting targets must be
561 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
562 * out of sync with the physical mp_lock, but we do not have to preserve
563 * the original ownership of the lock if it was out of synch (that is, we
564 * can leave it synchronized on return).
b68b7282
MD
565 */
566void
96728c05 567lwkt_preempt(thread_t ntd, int critpri)
b68b7282 568{
46a3f46d
MD
569 struct globaldata *gd = mycpu;
570 thread_t td = gd->gd_curthread;
8a8d5d85
MD
571#ifdef SMP
572 int mpheld;
57c254db 573 int savecnt;
8a8d5d85 574#endif
b68b7282 575
26a0694b 576 /*
96728c05
MD
577 * The caller has put us in a critical section. We can only preempt
578 * if the caller of the caller was not in a critical section (basically
57c254db
MD
579 * a local interrupt), as determined by the 'critpri' parameter. If
580 * we are unable to preempt
96728c05
MD
581 *
582 * YYY The target thread must be in a critical section (else it must
583 * inherit our critical section? I dunno yet).
26a0694b
MD
584 */
585 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 586
cb973d15 587 need_resched();
57c254db
MD
588 if (!_lwkt_wantresched(ntd, td)) {
589 ++preempt_miss;
590 return;
591 }
96728c05
MD
592 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
593 ++preempt_miss;
594 return;
595 }
596#ifdef SMP
46a3f46d 597 if (ntd->td_gd != gd) {
96728c05
MD
598 ++preempt_miss;
599 return;
600 }
601#endif
26a0694b
MD
602 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
603 ++preempt_weird;
604 return;
605 }
606 if (ntd->td_preempted) {
4b5f931b 607 ++preempt_hit;
26a0694b 608 return;
b68b7282 609 }
8a8d5d85 610#ifdef SMP
a2a5ad0d
MD
611 /*
612 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
613 * to or from the MP lock. In this case td_mpcount will be pre-disposed
614 * (non-zero) but not actually synchronized with the actual state of the
615 * lock. We can use it to imply an MP lock requirement for the
616 * preemption but we cannot use it to test whether we hold the MP lock
617 * or not.
a2a5ad0d 618 */
96728c05 619 savecnt = td->td_mpcount;
71ef2f5c 620 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
621 ntd->td_mpcount += td->td_mpcount;
622 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
623 ntd->td_mpcount -= td->td_mpcount;
624 ++preempt_miss;
625 return;
626 }
627#endif
26a0694b
MD
628
629 ++preempt_hit;
630 ntd->td_preempted = td;
631 td->td_flags |= TDF_PREEMPT_LOCK;
632 td->td_switch(ntd);
633 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
634#ifdef SMP
635 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
636 mpheld = MP_LOCK_HELD();
637 if (mpheld && td->td_mpcount == 0)
96728c05 638 cpu_rel_mplock();
71ef2f5c 639 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
640 panic("lwkt_preempt(): MP lock was not held through");
641#endif
26a0694b
MD
642 ntd->td_preempted = NULL;
643 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
644}
645
f1d1c3fa
MD
646/*
647 * Yield our thread while higher priority threads are pending. This is
648 * typically called when we leave a critical section but it can be safely
649 * called while we are in a critical section.
650 *
651 * This function will not generally yield to equal priority threads but it
652 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 653 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
654 * lwkt_yield_quick().
655 *
235957ed 656 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
657 * came along but was blocked and made pending.
658 *
f1d1c3fa
MD
659 * (self contained on a per cpu basis)
660 */
661void
662lwkt_yield_quick(void)
663{
7966cb69
MD
664 globaldata_t gd = mycpu;
665 thread_t td = gd->gd_curthread;
ef0fdad1 666
a2a5ad0d 667 /*
235957ed 668 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
669 * it with a non-zero cpl then we might not wind up calling splz after
670 * a task switch when the critical section is exited even though the
46a3f46d 671 * new task could accept the interrupt.
a2a5ad0d
MD
672 *
673 * XXX from crit_exit() only called after last crit section is released.
674 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
675 *
676 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
677 * except for this special case, we MUST call splz() here to handle any
678 * pending ints, particularly after we switch, or we might accidently
679 * halt the cpu with interrupts pending.
a2a5ad0d 680 */
46a3f46d 681 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 682 splz();
f1d1c3fa
MD
683
684 /*
685 * YYY enabling will cause wakeup() to task-switch, which really
686 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
687 * preemption and MP without actually doing preemption or MP, because a
688 * lot of code assumes that wakeup() does not block.
f1d1c3fa 689 */
46a3f46d
MD
690 if (untimely_switch && td->td_nest_count == 0 &&
691 gd->gd_intr_nesting_level == 0
692 ) {
f1d1c3fa
MD
693 crit_enter();
694 /*
695 * YYY temporary hacks until we disassociate the userland scheduler
696 * from the LWKT scheduler.
697 */
698 if (td->td_flags & TDF_RUNQ) {
699 lwkt_switch(); /* will not reenter yield function */
700 } else {
701 lwkt_schedule_self(); /* make sure we are scheduled */
702 lwkt_switch(); /* will not reenter yield function */
703 lwkt_deschedule_self(); /* make sure we are descheduled */
704 }
7966cb69 705 crit_exit_noyield(td);
f1d1c3fa 706 }
f1d1c3fa
MD
707}
708
8ad65e08 709/*
f1d1c3fa 710 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 711 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
712 * the crit_exit() call in lwkt_switch().
713 *
714 * (self contained on a per cpu basis)
8ad65e08
MD
715 */
716void
f1d1c3fa 717lwkt_yield(void)
8ad65e08 718{
f1d1c3fa
MD
719 lwkt_schedule_self();
720 lwkt_switch();
721}
722
723/*
724 * Schedule a thread to run. As the current thread we can always safely
725 * schedule ourselves, and a shortcut procedure is provided for that
726 * function.
727 *
728 * (non-blocking, self contained on a per cpu basis)
729 */
730void
731lwkt_schedule_self(void)
732{
733 thread_t td = curthread;
734
735 crit_enter();
736 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 737 _lwkt_enqueue(td);
05220613 738#ifdef _KERNEL
26a0694b
MD
739 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
740 panic("SCHED SELF PANIC");
05220613 741#endif
f1d1c3fa 742 crit_exit();
8ad65e08 743}
8ad65e08
MD
744
745/*
f1d1c3fa
MD
746 * Generic schedule. Possibly schedule threads belonging to other cpus and
747 * deal with threads that might be blocked on a wait queue.
748 *
96728c05 749 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
750 * Load balancing can be accomplished by requesting other sorts of actions
751 * for the thread in question.
8ad65e08
MD
752 */
753void
754lwkt_schedule(thread_t td)
755{
96728c05 756#ifdef INVARIANTS
26a0694b
MD
757 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
758 && td->td_proc->p_stat == SSLEEP
759 ) {
760 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
761 curthread,
762 curthread->td_proc ? curthread->td_proc->p_pid : -1,
763 curthread->td_proc ? curthread->td_proc->p_stat : -1,
764 td,
765 td->td_proc ? curthread->td_proc->p_pid : -1,
766 td->td_proc ? curthread->td_proc->p_stat : -1
767 );
768 panic("SCHED PANIC");
769 }
96728c05 770#endif
f1d1c3fa
MD
771 crit_enter();
772 if (td == curthread) {
773 _lwkt_enqueue(td);
774 } else {
775 lwkt_wait_t w;
776
777 /*
778 * If the thread is on a wait list we have to send our scheduling
779 * request to the owner of the wait structure. Otherwise we send
780 * the scheduling request to the cpu owning the thread. Races
781 * are ok, the target will forward the message as necessary (the
782 * message may chase the thread around before it finally gets
783 * acted upon).
784 *
785 * (remember, wait structures use stable storage)
786 */
787 if ((w = td->td_wait) != NULL) {
96728c05 788 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
789 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
790 --w->wa_count;
791 td->td_wait = NULL;
75cdbe6c 792 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 793 _lwkt_enqueue(td);
57c254db 794 if (td->td_preemptable) {
96728c05 795 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
796 } else if (_lwkt_wantresched(td, curthread)) {
797 need_resched();
798 }
f1d1c3fa 799 } else {
a72187e9 800 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 801 }
96728c05 802 lwkt_reltoken(&w->wa_token);
f1d1c3fa 803 } else {
96728c05 804 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
805 }
806 } else {
807 /*
808 * If the wait structure is NULL and we own the thread, there
809 * is no race (since we are in a critical section). If we
810 * do not own the thread there might be a race but the
811 * target cpu will deal with it.
812 */
75cdbe6c 813 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 814 _lwkt_enqueue(td);
57c254db 815 if (td->td_preemptable) {
96728c05 816 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
817 } else if (_lwkt_wantresched(td, curthread)) {
818 need_resched();
819 }
f1d1c3fa 820 } else {
a72187e9 821 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
822 }
823 }
8ad65e08 824 }
f1d1c3fa 825 crit_exit();
8ad65e08
MD
826}
827
d9eea1a5
MD
828/*
829 * Managed acquisition. This code assumes that the MP lock is held for
830 * the tdallq operation and that the thread has been descheduled from its
831 * original cpu. We also have to wait for the thread to be entirely switched
832 * out on its original cpu (this is usually fast enough that we never loop)
833 * since the LWKT system does not have to hold the MP lock while switching
834 * and the target may have released it before switching.
835 */
a2a5ad0d
MD
836void
837lwkt_acquire(thread_t td)
838{
839 struct globaldata *gd;
840
841 gd = td->td_gd;
842 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
843 while (td->td_flags & TDF_RUNNING) /* XXX spin */
844 ;
a2a5ad0d
MD
845 if (gd != mycpu) {
846 crit_enter();
847 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
848 gd = mycpu;
849 td->td_gd = gd;
a2a5ad0d
MD
850 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
851 crit_exit();
852 }
853}
854
8ad65e08 855/*
f1d1c3fa
MD
856 * Deschedule a thread.
857 *
858 * (non-blocking, self contained on a per cpu basis)
859 */
860void
861lwkt_deschedule_self(void)
862{
863 thread_t td = curthread;
864
865 crit_enter();
866 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
867 _lwkt_dequeue(td);
868 crit_exit();
869}
870
871/*
872 * Generic deschedule. Descheduling threads other then your own should be
873 * done only in carefully controlled circumstances. Descheduling is
874 * asynchronous.
875 *
876 * This function may block if the cpu has run out of messages.
8ad65e08
MD
877 */
878void
879lwkt_deschedule(thread_t td)
880{
f1d1c3fa
MD
881 crit_enter();
882 if (td == curthread) {
883 _lwkt_dequeue(td);
884 } else {
a72187e9 885 if (td->td_gd == mycpu) {
f1d1c3fa
MD
886 _lwkt_dequeue(td);
887 } else {
a72187e9 888 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
889 }
890 }
891 crit_exit();
892}
893
4b5f931b
MD
894/*
895 * Set the target thread's priority. This routine does not automatically
896 * switch to a higher priority thread, LWKT threads are not designed for
897 * continuous priority changes. Yield if you want to switch.
898 *
899 * We have to retain the critical section count which uses the high bits
26a0694b
MD
900 * of the td_pri field. The specified priority may also indicate zero or
901 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
902 */
903void
904lwkt_setpri(thread_t td, int pri)
905{
26a0694b 906 KKASSERT(pri >= 0);
a72187e9 907 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
908 crit_enter();
909 if (td->td_flags & TDF_RUNQ) {
910 _lwkt_dequeue(td);
911 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
912 _lwkt_enqueue(td);
913 } else {
914 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
915 }
916 crit_exit();
917}
918
919void
920lwkt_setpri_self(int pri)
921{
922 thread_t td = curthread;
923
4b5f931b
MD
924 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
925 crit_enter();
926 if (td->td_flags & TDF_RUNQ) {
927 _lwkt_dequeue(td);
928 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
929 _lwkt_enqueue(td);
930 } else {
931 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
932 }
933 crit_exit();
934}
935
936struct proc *
937lwkt_preempted_proc(void)
938{
73e4f7b9 939 thread_t td = curthread;
4b5f931b
MD
940 while (td->td_preempted)
941 td = td->td_preempted;
942 return(td->td_proc);
943}
944
ece04fd0
MD
945typedef struct lwkt_gettoken_req {
946 lwkt_token_t tok;
947 int cpu;
948} lwkt_gettoken_req;
949
950#if 0
4b5f931b 951
f1d1c3fa
MD
952/*
953 * This function deschedules the current thread and blocks on the specified
954 * wait queue. We obtain ownership of the wait queue in order to block
955 * on it. A generation number is used to interlock the wait queue in case
956 * it gets signalled while we are blocked waiting on the token.
957 *
958 * Note: alternatively we could dequeue our thread and then message the
959 * target cpu owning the wait queue. YYY implement as sysctl.
960 *
961 * Note: wait queue signals normally ping-pong the cpu as an optimization.
962 */
96728c05 963
f1d1c3fa 964void
ae8050a4 965lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
966{
967 thread_t td = curthread;
f1d1c3fa 968
f1d1c3fa 969 lwkt_gettoken(&w->wa_token);
ae8050a4 970 if (w->wa_gen == *gen) {
f1d1c3fa
MD
971 _lwkt_dequeue(td);
972 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
973 ++w->wa_count;
974 td->td_wait = w;
ae8050a4 975 td->td_wmesg = wmesg;
ece04fd0 976again:
f1d1c3fa 977 lwkt_switch();
ece04fd0
MD
978 lwkt_regettoken(&w->wa_token);
979 if (td->td_wmesg != NULL) {
980 _lwkt_dequeue(td);
981 goto again;
982 }
8ad65e08 983 }
ae8050a4
MD
984 /* token might be lost, doesn't matter for gen update */
985 *gen = w->wa_gen;
f1d1c3fa
MD
986 lwkt_reltoken(&w->wa_token);
987}
988
989/*
990 * Signal a wait queue. We gain ownership of the wait queue in order to
991 * signal it. Once a thread is removed from the wait queue we have to
992 * deal with the cpu owning the thread.
993 *
994 * Note: alternatively we could message the target cpu owning the wait
995 * queue. YYY implement as sysctl.
996 */
997void
ece04fd0 998lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
999{
1000 thread_t td;
1001 int count;
1002
1003 lwkt_gettoken(&w->wa_token);
1004 ++w->wa_gen;
ece04fd0
MD
1005 if (count < 0)
1006 count = w->wa_count;
f1d1c3fa
MD
1007 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1008 --count;
1009 --w->wa_count;
1010 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1011 td->td_wait = NULL;
ae8050a4 1012 td->td_wmesg = NULL;
a72187e9 1013 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1014 _lwkt_enqueue(td);
1015 } else {
a72187e9 1016 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1017 }
1018 lwkt_regettoken(&w->wa_token);
1019 }
1020 lwkt_reltoken(&w->wa_token);
1021}
1022
ece04fd0
MD
1023#endif
1024
f1d1c3fa 1025/*
96728c05 1026 * Acquire ownership of a token
f1d1c3fa 1027 *
96728c05 1028 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
1029 * section side effects, depending on its purpose. These side effects
1030 * guarentee that you will maintain ownership of the token as long as you
1031 * do not block. If you block you may lose access to the token (but you
1032 * must still release it even if you lose your access to it).
1033 *
96728c05 1034 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 1035 * a token, but do we really only need to disable IPIs ?
96728c05
MD
1036 *
1037 * YYY certain tokens could be made to act like mutexes when performance
1038 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1039 *
a2a5ad0d
MD
1040 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1041 * means that 4.x does not expect a switch so for now we cannot switch
1042 * when waiting for an IPI to be returned.
1043 *
1044 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
1045 * it and then block. The IPI causes the token to be given away to the
1046 * requesting cpu, unless it has already changed hands. Since only the
1047 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 1048 * This needs serious optimization.
f1d1c3fa 1049 */
57c254db
MD
1050
1051#ifdef SMP
1052
96728c05
MD
1053static
1054void
1055lwkt_gettoken_remote(void *arg)
1056{
1057 lwkt_gettoken_req *req = arg;
1058 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 1059#ifdef INVARIANTS
a2a5ad0d
MD
1060 if (token_debug)
1061 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 1062#endif
96728c05 1063 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
1064 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1065 /* else set reqcpu to point to current cpu for release */
96728c05
MD
1066 }
1067}
1068
57c254db
MD
1069#endif
1070
8a8d5d85 1071int
f1d1c3fa
MD
1072lwkt_gettoken(lwkt_token_t tok)
1073{
1074 /*
1075 * Prevent preemption so the token can't be taken away from us once
1076 * we gain ownership of it. Use a synchronous request which might
1077 * block. The request will be forwarded as necessary playing catchup
1078 * to the token.
1079 */
96728c05 1080
f1d1c3fa 1081 crit_enter();
57c254db 1082#ifdef INVARIANTS
1d4c9574
MD
1083 if (curthread->td_pri > 1800) {
1084 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1085 tok, ((int **)&tok)[-1]);
1086 }
a2a5ad0d
MD
1087 if (curthread->td_pri > 2000) {
1088 curthread->td_pri = 1000;
1089 panic("too HIGH!");
57c254db
MD
1090 }
1091#endif
96728c05 1092#ifdef SMP
d0e06f83 1093 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1094 struct lwkt_gettoken_req req;
1095 int seq;
96728c05
MD
1096 int dcpu;
1097
1098 req.cpu = mycpu->gd_cpuid;
1099 req.tok = tok;
1100 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1101 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1102#ifdef INVARIANTS
a2a5ad0d
MD
1103 if (token_debug)
1104 printf("REQT%d ", dcpu);
634081ff 1105#endif
96728c05
MD
1106 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1107 lwkt_wait_ipiq(dcpu, seq);
634081ff 1108#ifdef INVARIANTS
a2a5ad0d
MD
1109 if (token_debug)
1110 printf("REQR%d ", tok->t_cpu);
634081ff 1111#endif
f1d1c3fa
MD
1112 }
1113#endif
1114 /*
1115 * leave us in a critical section on return. This will be undone
8a8d5d85 1116 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1117 */
8a8d5d85 1118 return(++tok->t_gen);
f1d1c3fa
MD
1119}
1120
96728c05
MD
1121/*
1122 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1123 * failure.
1124 */
1125int
1126lwkt_trytoken(lwkt_token_t tok)
1127{
1128 crit_enter();
1129#ifdef SMP
1130 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1131 crit_exit();
96728c05
MD
1132 return(0);
1133 }
1134#endif
1135 /* leave us in the critical section */
1136 ++tok->t_gen;
1137 return(1);
1138}
1139
f1d1c3fa
MD
1140/*
1141 * Release your ownership of a token. Releases must occur in reverse
1142 * order to aquisitions, eventually so priorities can be unwound properly
1143 * like SPLs. At the moment the actual implemention doesn't care.
1144 *
1145 * We can safely hand a token that we own to another cpu without notifying
1146 * it, but once we do we can't get it back without requesting it (unless
1147 * the other cpu hands it back to us before we check).
1148 *
1149 * We might have lost the token, so check that.
7ba9c17c
MD
1150 *
1151 * Return the token's generation number. The number is useful to callers
1152 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1153 */
7ba9c17c 1154int
f1d1c3fa
MD
1155lwkt_reltoken(lwkt_token_t tok)
1156{
7ba9c17c
MD
1157 int gen;
1158
d0e06f83 1159 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1160 tok->t_cpu = tok->t_reqcpu;
1161 }
7ba9c17c 1162 gen = tok->t_gen;
f1d1c3fa 1163 crit_exit();
7ba9c17c 1164 return(gen);
f1d1c3fa
MD
1165}
1166
1167/*
7ba9c17c
MD
1168 * Reacquire a token that might have been lost. 0 is returned if the
1169 * generation has not changed (nobody stole the token from us), -1 is
1170 * returned otherwise. The token is reacquired regardless but the
1171 * generation number is not bumped further if we already own the token.
8a8d5d85 1172 *
7ba9c17c
MD
1173 * For efficiency we inline the best-case situation for lwkt_regettoken()
1174 * (i.e .we still own the token).
8a8d5d85
MD
1175 */
1176int
1177lwkt_gentoken(lwkt_token_t tok, int *gen)
1178{
7ba9c17c 1179 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1180 return(0);
7ba9c17c
MD
1181 *gen = lwkt_regettoken(tok);
1182 return(-1);
8a8d5d85
MD
1183}
1184
8a8d5d85 1185/*
435ff993
MD
1186 * Re-acquire a token that might have been lost. The generation number
1187 * is bumped and returned regardless of whether the token had been lost
1188 * or not (because we only have cpu granularity we have to bump the token
1189 * either way).
f1d1c3fa
MD
1190 */
1191int
1192lwkt_regettoken(lwkt_token_t tok)
1193{
96728c05 1194 /* assert we are in a critical section */
d0e06f83 1195 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1196#ifdef SMP
d0e06f83 1197 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1198 struct lwkt_gettoken_req req;
1199 int seq;
96728c05 1200 int dcpu;
57c254db 1201
96728c05
MD
1202 req.cpu = mycpu->gd_cpuid;
1203 req.tok = tok;
1204 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1205 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1206#ifdef INVARIANTS
cb973d15
MD
1207 if (token_debug)
1208 printf("REQT%d ", dcpu);
634081ff 1209#endif
96728c05
MD
1210 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1211 lwkt_wait_ipiq(dcpu, seq);
634081ff 1212#ifdef INVARIATNS
cb973d15
MD
1213 if (token_debug)
1214 printf("REQR%d ", tok->t_cpu);
634081ff 1215#endif
f1d1c3fa 1216 }
f1d1c3fa 1217#endif
96728c05 1218 }
435ff993 1219 ++tok->t_gen;
8a8d5d85 1220 return(tok->t_gen);
8ad65e08
MD
1221}
1222
72740893
MD
1223void
1224lwkt_inittoken(lwkt_token_t tok)
1225{
1226 /*
1227 * Zero structure and set cpu owner and reqcpu to cpu 0.
1228 */
1229 bzero(tok, sizeof(*tok));
1230}
1231
99df837e
MD
1232/*
1233 * Create a kernel process/thread/whatever. It shares it's address space
1234 * with proc0 - ie: kernel only.
1235 *
365fa13f
MD
1236 * NOTE! By default new threads are created with the MP lock held. A
1237 * thread which does not require the MP lock should release it by calling
1238 * rel_mplock() at the start of the new thread.
99df837e
MD
1239 */
1240int
1241lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1242 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1243 const char *fmt, ...)
99df837e 1244{
73e4f7b9 1245 thread_t td;
e2565a42 1246 __va_list ap;
99df837e 1247
75cdbe6c 1248 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1249 if (tdp)
1250 *tdp = td;
99df837e 1251 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1252 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1253#ifdef SMP
1254 td->td_mpcount = 1;
1255#endif
99df837e
MD
1256
1257 /*
1258 * Set up arg0 for 'ps' etc
1259 */
e2565a42 1260 __va_start(ap, fmt);
99df837e 1261 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1262 __va_end(ap);
99df837e
MD
1263
1264 /*
1265 * Schedule the thread to run
1266 */
ef0fdad1
MD
1267 if ((td->td_flags & TDF_STOPREQ) == 0)
1268 lwkt_schedule(td);
1269 else
1270 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1271 return 0;
1272}
1273
1274/*
1275 * Destroy an LWKT thread. Warning! This function is not called when
1276 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1277 * uses a different reaping mechanism.
1278 */
1279void
1280lwkt_exit(void)
1281{
1282 thread_t td = curthread;
1283
1284 if (td->td_flags & TDF_VERBOSE)
1285 printf("kthread %p %s has exited\n", td, td->td_comm);
1286 crit_enter();
1287 lwkt_deschedule_self();
1288 ++mycpu->gd_tdfreecount;
1289 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1290 cpu_thread_exit();
1291}
1292
05220613 1293#ifdef _KERNEL
99df837e
MD
1294/*
1295 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1296 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1297 *
1298 * NOTE! By default kthreads are created with the MP lock held. A
1299 * thread which does not require the MP lock should release it by calling
1300 * rel_mplock() at the start of the new thread.
99df837e
MD
1301 */
1302int
1303kthread_create(void (*func)(void *), void *arg,
1304 struct thread **tdp, const char *fmt, ...)
1305{
73e4f7b9 1306 thread_t td;
e2565a42 1307 __va_list ap;
99df837e 1308
75cdbe6c 1309 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1310 if (tdp)
1311 *tdp = td;
99df837e
MD
1312 cpu_set_thread_handler(td, kthread_exit, func, arg);
1313 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1314#ifdef SMP
1315 td->td_mpcount = 1;
1316#endif
99df837e
MD
1317
1318 /*
1319 * Set up arg0 for 'ps' etc
1320 */
e2565a42 1321 __va_start(ap, fmt);
99df837e 1322 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1323 __va_end(ap);
99df837e
MD
1324
1325 /*
1326 * Schedule the thread to run
1327 */
1328 lwkt_schedule(td);
1329 return 0;
1330}
1331
05220613
MD
1332#endif
1333
26a0694b
MD
1334void
1335crit_panic(void)
1336{
73e4f7b9 1337 thread_t td = curthread;
26a0694b
MD
1338 int lpri = td->td_pri;
1339
1340 td->td_pri = 0;
1341 panic("td_pri is/would-go negative! %p %d", td, lpri);
1342}
1343
99df837e
MD
1344/*
1345 * Destroy an LWKT thread. Warning! This function is not called when
1346 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1347 * uses a different reaping mechanism.
1348 *
1349 * XXX duplicates lwkt_exit()
1350 */
1351void
1352kthread_exit(void)
1353{
1354 lwkt_exit();
1355}
1356
96728c05
MD
1357#ifdef SMP
1358
1359/*
1360 * Send a function execution request to another cpu. The request is queued
1361 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1362 * possible target cpu. The FIFO can be written.
1363 *
1364 * YYY If the FIFO fills up we have to enable interrupts and process the
1365 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1366 * Create a CPU_*() function to do this!
1367 *
46a3f46d
MD
1368 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1369 * end will take care of any pending interrupts.
1370 *
96728c05
MD
1371 * Must be called from a critical section.
1372 */
1373int
1374lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1375{
1376 lwkt_ipiq_t ip;
1377 int windex;
a2a5ad0d 1378 struct globaldata *gd = mycpu;
96728c05 1379
a2a5ad0d 1380 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1381 func(arg);
1382 return(0);
1383 }
cb973d15 1384 crit_enter();
a2a5ad0d
MD
1385 ++gd->gd_intr_nesting_level;
1386#ifdef INVARIANTS
1387 if (gd->gd_intr_nesting_level > 20)
1388 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1389#endif
96728c05
MD
1390 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1391 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1392 ++ipiq_count;
a2a5ad0d 1393 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1394
1395 /*
1396 * We always drain before the FIFO becomes full so it should never
1397 * become full. We need to leave enough entries to deal with
1398 * reentrancy.
1399 */
1400 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1401 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1402 ip->ip_func[windex] = func;
1403 ip->ip_arg[windex] = arg;
1404 /* YYY memory barrier */
1405 ++ip->ip_windex;
96728c05
MD
1406 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1407 unsigned int eflags = read_eflags();
1408 cpu_enable_intr();
1409 ++ipiq_fifofull;
cb973d15 1410 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1411 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1412 lwkt_process_ipiq();
1413 }
1414 write_eflags(eflags);
1415 }
a2a5ad0d 1416 --gd->gd_intr_nesting_level;
96728c05 1417 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1418 crit_exit();
96728c05
MD
1419 return(ip->ip_windex);
1420}
1421
cb973d15
MD
1422/*
1423 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1424 * The message will not be sent to stopped cpus.
cb973d15
MD
1425 */
1426void
1427lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1428{
1429 int cpuid;
1430
435ff993 1431 mask &= ~stopped_cpus;
cb973d15
MD
1432 while (mask) {
1433 cpuid = bsfl(mask);
1434 lwkt_send_ipiq(cpuid, func, arg);
1435 mask &= ~(1 << cpuid);
1436 }
1437}
1438
96728c05
MD
1439/*
1440 * Wait for the remote cpu to finish processing a function.
1441 *
1442 * YYY we have to enable interrupts and process the IPIQ while waiting
1443 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1444 * function to do this! YYY we really should 'block' here.
1445 *
1446 * Must be called from a critical section. Thsi routine may be called
1447 * from an interrupt (for example, if an interrupt wakes a foreign thread
1448 * up).
1449 */
1450void
1451lwkt_wait_ipiq(int dcpu, int seq)
1452{
1453 lwkt_ipiq_t ip;
a2a5ad0d 1454 int maxc = 100000000;
96728c05
MD
1455
1456 if (dcpu != mycpu->gd_cpuid) {
1457 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1458 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1459 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1460 unsigned int eflags = read_eflags();
1461 cpu_enable_intr();
cb973d15 1462 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1463 lwkt_process_ipiq();
a2a5ad0d 1464 if (--maxc == 0)
cb973d15 1465 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1466 if (maxc < -1000000)
1467 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1468 }
1469 write_eflags(eflags);
1470 }
1471 }
1472}
1473
1474/*
1475 * Called from IPI interrupt (like a fast interrupt), which has placed
1476 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1477 * May also be called from doreti or splz, or be reentrantly called
1478 * indirectly through the ip_func[] we run.
96728c05
MD
1479 */
1480void
1481lwkt_process_ipiq(void)
1482{
1483 int n;
1484 int cpuid = mycpu->gd_cpuid;
1485
1486 for (n = 0; n < ncpus; ++n) {
1487 lwkt_ipiq_t ip;
1488 int ri;
1489
1490 if (n == cpuid)
1491 continue;
1492 ip = globaldata_find(n)->gd_ipiq;
1493 if (ip == NULL)
1494 continue;
1495 ip = &ip[cpuid];
cb973d15
MD
1496
1497 /*
1498 * Note: xindex is only updated after we are sure the function has
1499 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1500 * function may send an IPI which may block/drain.
1501 */
96728c05
MD
1502 while (ip->ip_rindex != ip->ip_windex) {
1503 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1504 ++ip->ip_rindex;
cb973d15
MD
1505 ip->ip_func[ri](ip->ip_arg[ri]);
1506 /* YYY memory barrier */
1507 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1508 }
1509 }
1510}
1511
1512#else
1513
1514int
1515lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1516{
1517 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1518 return(0); /* NOT REACHED */
1519}
1520
1521void
1522lwkt_wait_ipiq(int dcpu, int seq)
1523{
1524 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1525}
1526
1527#endif