Add chmod 1777 for /tmp and make other minor adjustments.
[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 *
fb04f4fd 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.44 2003/11/24 23:56:07 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
8ad65e08
MD
156/*
157 * LWKTs operate on a per-cpu basis
158 *
73e4f7b9 159 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
160 */
161void
162lwkt_gdinit(struct globaldata *gd)
163{
4b5f931b
MD
164 int i;
165
166 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
167 TAILQ_INIT(&gd->gd_tdrunq[i]);
168 gd->gd_runqmask = 0;
73e4f7b9 169 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
170}
171
7d0bac62
MD
172/*
173 * Initialize a thread wait structure prior to first use.
174 *
175 * NOTE! called from low level boot code, we cannot do anything fancy!
176 */
177void
178lwkt_init_wait(lwkt_wait_t w)
179{
180 TAILQ_INIT(&w->wa_waitq);
181}
182
183/*
184 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
185 * or LWKT start address before it can be scheduled. If the target cpu is
186 * -1 the thread will be created on the current cpu.
0cfcada1
MD
187 *
188 * If you intend to create a thread without a process context this function
189 * does everything except load the startup and switcher function.
7d0bac62
MD
190 */
191thread_t
75cdbe6c 192lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 193{
99df837e 194 void *stack;
ef0fdad1 195 int flags = 0;
7d0bac62 196
ef0fdad1 197 if (td == NULL) {
26a0694b 198 crit_enter();
ef0fdad1
MD
199 if (mycpu->gd_tdfreecount > 0) {
200 --mycpu->gd_tdfreecount;
201 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 202 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
203 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
204 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
205 crit_exit();
206 stack = td->td_kstack;
207 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
208 } else {
209 crit_exit();
05220613 210#ifdef _KERNEL
ef0fdad1 211 td = zalloc(thread_zone);
05220613
MD
212#else
213 td = malloc(sizeof(struct thread));
214#endif
ef0fdad1
MD
215 td->td_kstack = NULL;
216 flags |= TDF_ALLOCATED_THREAD;
217 }
218 }
219 if ((stack = td->td_kstack) == NULL) {
05220613
MD
220#ifdef _KERNEL
221 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
222#else
fb04f4fd 223 stack = libcaps_alloc_stack(THREAD_STACK);
05220613 224#endif
ef0fdad1 225 flags |= TDF_ALLOCATED_STACK;
99df837e 226 }
75cdbe6c
MD
227 if (cpu < 0)
228 lwkt_init_thread(td, stack, flags, mycpu);
229 else
230 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 231 return(td);
7d0bac62
MD
232}
233
234/*
235 * Initialize a preexisting thread structure. This function is used by
236 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
237 *
f8c3996b
MD
238 * All threads start out in a critical section at a priority of
239 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
240 * appropriate. This function may send an IPI message when the
241 * requested cpu is not the current cpu and consequently gd_tdallq may
242 * not be initialized synchronously from the point of view of the originating
243 * cpu.
244 *
245 * NOTE! we have to be careful in regards to creating threads for other cpus
246 * if SMP has not yet been activated.
7d0bac62 247 */
75cdbe6c
MD
248static void
249lwkt_init_thread_remote(void *arg)
250{
251 thread_t td = arg;
252
253 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
254}
255
7d0bac62 256void
26a0694b 257lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 258{
99df837e
MD
259 bzero(td, sizeof(struct thread));
260 td->td_kstack = stack;
261 td->td_flags |= flags;
26a0694b 262 td->td_gd = gd;
f8c3996b 263 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
c95cd171 264 lwkt_initport(&td->td_msgport, td);
99df837e 265 pmap_init_thread(td);
75cdbe6c
MD
266 if (smp_active == 0 || gd == mycpu) {
267 crit_enter();
268 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
269 crit_exit();
270 } else {
271 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
272 }
73e4f7b9
MD
273}
274
275void
276lwkt_set_comm(thread_t td, const char *ctl, ...)
277{
e2565a42 278 __va_list va;
73e4f7b9 279
e2565a42 280 __va_start(va, ctl);
73e4f7b9 281 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 282 __va_end(va);
7d0bac62
MD
283}
284
99df837e 285void
73e4f7b9 286lwkt_hold(thread_t td)
99df837e 287{
73e4f7b9
MD
288 ++td->td_refs;
289}
290
291void
292lwkt_rele(thread_t td)
293{
294 KKASSERT(td->td_refs > 0);
295 --td->td_refs;
296}
297
c95cd171
MD
298#ifdef _KERNEL
299
73e4f7b9
MD
300void
301lwkt_wait_free(thread_t td)
302{
303 while (td->td_refs)
377d4740 304 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
305}
306
c95cd171
MD
307#endif
308
73e4f7b9
MD
309void
310lwkt_free_thread(thread_t td)
311{
312 struct globaldata *gd = mycpu;
313
d9eea1a5 314 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
315 ("lwkt_free_thread: did not exit! %p", td));
316
317 crit_enter();
73e4f7b9
MD
318 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
319 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
320 (td->td_flags & TDF_ALLOCATED_THREAD)
321 ) {
73e4f7b9
MD
322 ++gd->gd_tdfreecount;
323 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
324 crit_exit();
325 } else {
326 crit_exit();
327 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613
MD
328#ifdef _KERNEL
329 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
330#else
fb04f4fd 331 libcaps_free_stack(td->td_kstack, THREAD_STACK);
05220613 332#endif
73e4f7b9 333 /* gd invalid */
99df837e
MD
334 td->td_kstack = NULL;
335 }
05220613
MD
336 if (td->td_flags & TDF_ALLOCATED_THREAD) {
337#ifdef _KERNEL
99df837e 338 zfree(thread_zone, td);
05220613
MD
339#else
340 free(td);
341#endif
342 }
99df837e
MD
343 }
344}
345
346
8ad65e08
MD
347/*
348 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
349 * switch to the idlethread. Switching must occur within a critical
350 * section to avoid races with the scheduling queue.
351 *
352 * We always have full control over our cpu's run queue. Other cpus
353 * that wish to manipulate our queue must use the cpu_*msg() calls to
354 * talk to our cpu, so a critical section is all that is needed and
355 * the result is very, very fast thread switching.
356 *
96728c05
MD
357 * The LWKT scheduler uses a fixed priority model and round-robins at
358 * each priority level. User process scheduling is a totally
359 * different beast and LWKT priorities should not be confused with
360 * user process priorities.
f1d1c3fa 361 *
96728c05
MD
362 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
363 * cleans it up. Note that the td_switch() function cannot do anything that
364 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
365 * the target thread (not the current thread). It's nice to have a scheduler
366 * that does not need the MP lock to work because it allows us to do some
367 * really cool high-performance MP lock optimizations.
8ad65e08 368 */
96728c05 369
8ad65e08
MD
370void
371lwkt_switch(void)
372{
4b5f931b 373 struct globaldata *gd;
f1d1c3fa 374 thread_t td = curthread;
8ad65e08 375 thread_t ntd;
8a8d5d85
MD
376#ifdef SMP
377 int mpheld;
378#endif
8ad65e08 379
46a3f46d
MD
380 /*
381 * Switching from within a 'fast' (non thread switched) interrupt is
382 * illegal.
383 */
384 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 385 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 386 }
ef0fdad1 387
cb973d15
MD
388 /*
389 * Passive release (used to transition from user to kernel mode
390 * when we block or switch rather then when we enter the kernel).
391 * This function is NOT called if we are switching into a preemption
392 * or returning from a preemption. Typically this causes us to lose
393 * our P_CURPROC designation (if we have one) and become a true LWKT
394 * thread, and may also hand P_CURPROC to another process and schedule
395 * its thread.
396 */
397 if (td->td_release)
398 td->td_release(td);
399
f1d1c3fa 400 crit_enter();
4b5f931b 401 ++switch_count;
8a8d5d85
MD
402
403#ifdef SMP
404 /*
405 * td_mpcount cannot be used to determine if we currently hold the
406 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
407 * to get the lock, and switch out if it can't. Our ownership of
408 * the actual lock will remain stable while we are in a critical section
409 * (but, of course, another cpu may own or release the lock so the
410 * actual value of mp_lock is not stable).
8a8d5d85
MD
411 */
412 mpheld = MP_LOCK_HELD();
413#endif
99df837e
MD
414 if ((ntd = td->td_preempted) != NULL) {
415 /*
416 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
417 * thread. This occurs transparently, whether the preempted thread
418 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
419 * itself).
420 *
421 * We have to setup the MP lock for the original thread after backing
422 * out the adjustment that was made to curthread when the original
423 * was preempted.
99df837e 424 */
26a0694b 425 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 426#ifdef SMP
96728c05
MD
427 if (ntd->td_mpcount && mpheld == 0) {
428 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
429 td, ntd, td->td_mpcount, ntd->td_mpcount);
430 }
8a8d5d85
MD
431 if (ntd->td_mpcount) {
432 td->td_mpcount -= ntd->td_mpcount;
433 KKASSERT(td->td_mpcount >= 0);
434 }
435#endif
26a0694b 436 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 437 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 438 } else {
4b5f931b
MD
439 /*
440 * Priority queue / round-robin at each priority. Note that user
441 * processes run at a fixed, low priority and the user process
442 * scheduler deals with interactions between user processes
443 * by scheduling and descheduling them from the LWKT queue as
444 * necessary.
8a8d5d85
MD
445 *
446 * We have to adjust the MP lock for the target thread. If we
447 * need the MP lock and cannot obtain it we try to locate a
448 * thread that does not need the MP lock.
4b5f931b
MD
449 */
450 gd = mycpu;
4b5f931b
MD
451again:
452 if (gd->gd_runqmask) {
453 int nq = bsrl(gd->gd_runqmask);
454 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
455 gd->gd_runqmask &= ~(1 << nq);
456 goto again;
457 }
8a8d5d85
MD
458#ifdef SMP
459 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
460 /*
96728c05
MD
461 * Target needs MP lock and we couldn't get it, try
462 * to locate a thread which does not need the MP lock
3c23a41a 463 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
464 */
465 u_int32_t rqmask = gd->gd_runqmask;
466 while (rqmask) {
467 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
468 if (ntd->td_mpcount == 0)
469 break;
470 }
471 if (ntd)
472 break;
473 rqmask &= ~(1 << nq);
474 nq = bsrl(rqmask);
475 }
476 if (ntd == NULL) {
a2a5ad0d
MD
477 ntd = &gd->gd_idlethread;
478 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
479 } else {
480 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
481 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
482 }
483 } else {
484 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
485 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
486 }
487#else
4b5f931b
MD
488 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
489 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 490#endif
4b5f931b 491 } else {
3c23a41a
MD
492 /*
493 * Nothing to run but we may still need the BGL to deal with
494 * pending interrupts, spin in idle if so.
495 */
a2a5ad0d 496 ntd = &gd->gd_idlethread;
235957ed 497 if (gd->gd_reqflags)
3c23a41a 498 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 499 }
f1d1c3fa 500 }
26a0694b
MD
501 KASSERT(ntd->td_pri >= TDPRI_CRIT,
502 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
503
504 /*
505 * Do the actual switch. If the new target does not need the MP lock
506 * and we are holding it, release the MP lock. If the new target requires
507 * the MP lock we have already acquired it for the target.
508 */
509#ifdef SMP
510 if (ntd->td_mpcount == 0 ) {
511 if (MP_LOCK_HELD())
512 cpu_rel_mplock();
513 } else {
514 ASSERT_MP_LOCK_HELD();
515 }
516#endif
8a8d5d85 517 if (td != ntd) {
f1d1c3fa 518 td->td_switch(ntd);
8a8d5d85 519 }
96728c05 520
f1d1c3fa 521 crit_exit();
8ad65e08
MD
522}
523
cb973d15
MD
524/*
525 * Switch if another thread has a higher priority. Do not switch to other
526 * threads at the same priority.
527 */
528void
529lwkt_maybe_switch()
530{
531 struct globaldata *gd = mycpu;
532 struct thread *td = gd->gd_curthread;
533
534 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
535 lwkt_switch();
536 }
537}
538
b68b7282 539/*
96728c05
MD
540 * Request that the target thread preempt the current thread. Preemption
541 * only works under a specific set of conditions:
b68b7282 542 *
96728c05
MD
543 * - We are not preempting ourselves
544 * - The target thread is owned by the current cpu
545 * - We are not currently being preempted
546 * - The target is not currently being preempted
547 * - We are able to satisfy the target's MP lock requirements (if any).
548 *
549 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
550 * this is called via lwkt_schedule() through the td_preemptable callback.
551 * critpri is the managed critical priority that we should ignore in order
552 * to determine whether preemption is possible (aka usually just the crit
553 * priority of lwkt_schedule() itself).
b68b7282 554 *
26a0694b
MD
555 * XXX at the moment we run the target thread in a critical section during
556 * the preemption in order to prevent the target from taking interrupts
557 * that *WE* can't. Preemption is strictly limited to interrupt threads
558 * and interrupt-like threads, outside of a critical section, and the
559 * preempted source thread will be resumed the instant the target blocks
560 * whether or not the source is scheduled (i.e. preemption is supposed to
561 * be as transparent as possible).
4b5f931b 562 *
8a8d5d85
MD
563 * The target thread inherits our MP count (added to its own) for the
564 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
565 * MP lock during the preemption. Therefore, any preempting targets must be
566 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
567 * out of sync with the physical mp_lock, but we do not have to preserve
568 * the original ownership of the lock if it was out of synch (that is, we
569 * can leave it synchronized on return).
b68b7282
MD
570 */
571void
96728c05 572lwkt_preempt(thread_t ntd, int critpri)
b68b7282 573{
46a3f46d
MD
574 struct globaldata *gd = mycpu;
575 thread_t td = gd->gd_curthread;
8a8d5d85
MD
576#ifdef SMP
577 int mpheld;
57c254db 578 int savecnt;
8a8d5d85 579#endif
b68b7282 580
26a0694b 581 /*
96728c05
MD
582 * The caller has put us in a critical section. We can only preempt
583 * if the caller of the caller was not in a critical section (basically
57c254db
MD
584 * a local interrupt), as determined by the 'critpri' parameter. If
585 * we are unable to preempt
96728c05
MD
586 *
587 * YYY The target thread must be in a critical section (else it must
588 * inherit our critical section? I dunno yet).
26a0694b
MD
589 */
590 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 591
cb973d15 592 need_resched();
57c254db
MD
593 if (!_lwkt_wantresched(ntd, td)) {
594 ++preempt_miss;
595 return;
596 }
96728c05
MD
597 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
598 ++preempt_miss;
599 return;
600 }
601#ifdef SMP
46a3f46d 602 if (ntd->td_gd != gd) {
96728c05
MD
603 ++preempt_miss;
604 return;
605 }
606#endif
26a0694b
MD
607 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
608 ++preempt_weird;
609 return;
610 }
611 if (ntd->td_preempted) {
4b5f931b 612 ++preempt_hit;
26a0694b 613 return;
b68b7282 614 }
8a8d5d85 615#ifdef SMP
a2a5ad0d
MD
616 /*
617 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
618 * to or from the MP lock. In this case td_mpcount will be pre-disposed
619 * (non-zero) but not actually synchronized with the actual state of the
620 * lock. We can use it to imply an MP lock requirement for the
621 * preemption but we cannot use it to test whether we hold the MP lock
622 * or not.
a2a5ad0d 623 */
96728c05 624 savecnt = td->td_mpcount;
71ef2f5c 625 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
626 ntd->td_mpcount += td->td_mpcount;
627 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
628 ntd->td_mpcount -= td->td_mpcount;
629 ++preempt_miss;
630 return;
631 }
632#endif
26a0694b
MD
633
634 ++preempt_hit;
635 ntd->td_preempted = td;
636 td->td_flags |= TDF_PREEMPT_LOCK;
637 td->td_switch(ntd);
638 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
639#ifdef SMP
640 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
641 mpheld = MP_LOCK_HELD();
642 if (mpheld && td->td_mpcount == 0)
96728c05 643 cpu_rel_mplock();
71ef2f5c 644 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
645 panic("lwkt_preempt(): MP lock was not held through");
646#endif
26a0694b
MD
647 ntd->td_preempted = NULL;
648 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
649}
650
f1d1c3fa
MD
651/*
652 * Yield our thread while higher priority threads are pending. This is
653 * typically called when we leave a critical section but it can be safely
654 * called while we are in a critical section.
655 *
656 * This function will not generally yield to equal priority threads but it
657 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 658 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
659 * lwkt_yield_quick().
660 *
235957ed 661 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
662 * came along but was blocked and made pending.
663 *
f1d1c3fa
MD
664 * (self contained on a per cpu basis)
665 */
666void
667lwkt_yield_quick(void)
668{
7966cb69
MD
669 globaldata_t gd = mycpu;
670 thread_t td = gd->gd_curthread;
ef0fdad1 671
a2a5ad0d 672 /*
235957ed 673 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
674 * it with a non-zero cpl then we might not wind up calling splz after
675 * a task switch when the critical section is exited even though the
46a3f46d 676 * new task could accept the interrupt.
a2a5ad0d
MD
677 *
678 * XXX from crit_exit() only called after last crit section is released.
679 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
680 *
681 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
682 * except for this special case, we MUST call splz() here to handle any
683 * pending ints, particularly after we switch, or we might accidently
684 * halt the cpu with interrupts pending.
a2a5ad0d 685 */
46a3f46d 686 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 687 splz();
f1d1c3fa
MD
688
689 /*
690 * YYY enabling will cause wakeup() to task-switch, which really
691 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
692 * preemption and MP without actually doing preemption or MP, because a
693 * lot of code assumes that wakeup() does not block.
f1d1c3fa 694 */
46a3f46d
MD
695 if (untimely_switch && td->td_nest_count == 0 &&
696 gd->gd_intr_nesting_level == 0
697 ) {
f1d1c3fa
MD
698 crit_enter();
699 /*
700 * YYY temporary hacks until we disassociate the userland scheduler
701 * from the LWKT scheduler.
702 */
703 if (td->td_flags & TDF_RUNQ) {
704 lwkt_switch(); /* will not reenter yield function */
705 } else {
706 lwkt_schedule_self(); /* make sure we are scheduled */
707 lwkt_switch(); /* will not reenter yield function */
708 lwkt_deschedule_self(); /* make sure we are descheduled */
709 }
7966cb69 710 crit_exit_noyield(td);
f1d1c3fa 711 }
f1d1c3fa
MD
712}
713
8ad65e08 714/*
f1d1c3fa 715 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 716 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
717 * the crit_exit() call in lwkt_switch().
718 *
719 * (self contained on a per cpu basis)
8ad65e08
MD
720 */
721void
f1d1c3fa 722lwkt_yield(void)
8ad65e08 723{
f1d1c3fa
MD
724 lwkt_schedule_self();
725 lwkt_switch();
726}
727
728/*
729 * Schedule a thread to run. As the current thread we can always safely
730 * schedule ourselves, and a shortcut procedure is provided for that
731 * function.
732 *
733 * (non-blocking, self contained on a per cpu basis)
734 */
735void
736lwkt_schedule_self(void)
737{
738 thread_t td = curthread;
739
740 crit_enter();
741 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 742 _lwkt_enqueue(td);
05220613 743#ifdef _KERNEL
26a0694b
MD
744 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
745 panic("SCHED SELF PANIC");
05220613 746#endif
f1d1c3fa 747 crit_exit();
8ad65e08 748}
8ad65e08
MD
749
750/*
f1d1c3fa
MD
751 * Generic schedule. Possibly schedule threads belonging to other cpus and
752 * deal with threads that might be blocked on a wait queue.
753 *
96728c05 754 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
755 * Load balancing can be accomplished by requesting other sorts of actions
756 * for the thread in question.
8ad65e08
MD
757 */
758void
759lwkt_schedule(thread_t td)
760{
96728c05 761#ifdef INVARIANTS
26a0694b
MD
762 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
763 && td->td_proc->p_stat == SSLEEP
764 ) {
765 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
766 curthread,
767 curthread->td_proc ? curthread->td_proc->p_pid : -1,
768 curthread->td_proc ? curthread->td_proc->p_stat : -1,
769 td,
770 td->td_proc ? curthread->td_proc->p_pid : -1,
771 td->td_proc ? curthread->td_proc->p_stat : -1
772 );
773 panic("SCHED PANIC");
774 }
96728c05 775#endif
f1d1c3fa
MD
776 crit_enter();
777 if (td == curthread) {
778 _lwkt_enqueue(td);
779 } else {
780 lwkt_wait_t w;
781
782 /*
783 * If the thread is on a wait list we have to send our scheduling
784 * request to the owner of the wait structure. Otherwise we send
785 * the scheduling request to the cpu owning the thread. Races
786 * are ok, the target will forward the message as necessary (the
787 * message may chase the thread around before it finally gets
788 * acted upon).
789 *
790 * (remember, wait structures use stable storage)
791 */
792 if ((w = td->td_wait) != NULL) {
96728c05 793 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
794 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
795 --w->wa_count;
796 td->td_wait = NULL;
75cdbe6c 797 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 798 _lwkt_enqueue(td);
57c254db 799 if (td->td_preemptable) {
96728c05 800 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
801 } else if (_lwkt_wantresched(td, curthread)) {
802 need_resched();
803 }
f1d1c3fa 804 } else {
a72187e9 805 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 806 }
96728c05 807 lwkt_reltoken(&w->wa_token);
f1d1c3fa 808 } else {
96728c05 809 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
810 }
811 } else {
812 /*
813 * If the wait structure is NULL and we own the thread, there
814 * is no race (since we are in a critical section). If we
815 * do not own the thread there might be a race but the
816 * target cpu will deal with it.
817 */
75cdbe6c 818 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 819 _lwkt_enqueue(td);
57c254db 820 if (td->td_preemptable) {
96728c05 821 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
822 } else if (_lwkt_wantresched(td, curthread)) {
823 need_resched();
824 }
f1d1c3fa 825 } else {
a72187e9 826 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
827 }
828 }
8ad65e08 829 }
f1d1c3fa 830 crit_exit();
8ad65e08
MD
831}
832
d9eea1a5
MD
833/*
834 * Managed acquisition. This code assumes that the MP lock is held for
835 * the tdallq operation and that the thread has been descheduled from its
836 * original cpu. We also have to wait for the thread to be entirely switched
837 * out on its original cpu (this is usually fast enough that we never loop)
838 * since the LWKT system does not have to hold the MP lock while switching
839 * and the target may have released it before switching.
840 */
a2a5ad0d
MD
841void
842lwkt_acquire(thread_t td)
843{
844 struct globaldata *gd;
845
846 gd = td->td_gd;
847 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
848 while (td->td_flags & TDF_RUNNING) /* XXX spin */
849 ;
a2a5ad0d
MD
850 if (gd != mycpu) {
851 crit_enter();
852 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
853 gd = mycpu;
854 td->td_gd = gd;
a2a5ad0d
MD
855 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
856 crit_exit();
857 }
858}
859
8ad65e08 860/*
f1d1c3fa
MD
861 * Deschedule a thread.
862 *
863 * (non-blocking, self contained on a per cpu basis)
864 */
865void
866lwkt_deschedule_self(void)
867{
868 thread_t td = curthread;
869
870 crit_enter();
871 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
872 _lwkt_dequeue(td);
873 crit_exit();
874}
875
876/*
877 * Generic deschedule. Descheduling threads other then your own should be
878 * done only in carefully controlled circumstances. Descheduling is
879 * asynchronous.
880 *
881 * This function may block if the cpu has run out of messages.
8ad65e08
MD
882 */
883void
884lwkt_deschedule(thread_t td)
885{
f1d1c3fa
MD
886 crit_enter();
887 if (td == curthread) {
888 _lwkt_dequeue(td);
889 } else {
a72187e9 890 if (td->td_gd == mycpu) {
f1d1c3fa
MD
891 _lwkt_dequeue(td);
892 } else {
a72187e9 893 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
894 }
895 }
896 crit_exit();
897}
898
4b5f931b
MD
899/*
900 * Set the target thread's priority. This routine does not automatically
901 * switch to a higher priority thread, LWKT threads are not designed for
902 * continuous priority changes. Yield if you want to switch.
903 *
904 * We have to retain the critical section count which uses the high bits
26a0694b
MD
905 * of the td_pri field. The specified priority may also indicate zero or
906 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
907 */
908void
909lwkt_setpri(thread_t td, int pri)
910{
26a0694b 911 KKASSERT(pri >= 0);
a72187e9 912 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
913 crit_enter();
914 if (td->td_flags & TDF_RUNQ) {
915 _lwkt_dequeue(td);
916 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
917 _lwkt_enqueue(td);
918 } else {
919 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
920 }
921 crit_exit();
922}
923
924void
925lwkt_setpri_self(int pri)
926{
927 thread_t td = curthread;
928
4b5f931b
MD
929 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
930 crit_enter();
931 if (td->td_flags & TDF_RUNQ) {
932 _lwkt_dequeue(td);
933 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
934 _lwkt_enqueue(td);
935 } else {
936 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
937 }
938 crit_exit();
939}
940
941struct proc *
942lwkt_preempted_proc(void)
943{
73e4f7b9 944 thread_t td = curthread;
4b5f931b
MD
945 while (td->td_preempted)
946 td = td->td_preempted;
947 return(td->td_proc);
948}
949
ece04fd0
MD
950typedef struct lwkt_gettoken_req {
951 lwkt_token_t tok;
952 int cpu;
953} lwkt_gettoken_req;
954
955#if 0
4b5f931b 956
f1d1c3fa
MD
957/*
958 * This function deschedules the current thread and blocks on the specified
959 * wait queue. We obtain ownership of the wait queue in order to block
960 * on it. A generation number is used to interlock the wait queue in case
961 * it gets signalled while we are blocked waiting on the token.
962 *
963 * Note: alternatively we could dequeue our thread and then message the
964 * target cpu owning the wait queue. YYY implement as sysctl.
965 *
966 * Note: wait queue signals normally ping-pong the cpu as an optimization.
967 */
96728c05 968
f1d1c3fa 969void
ae8050a4 970lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
971{
972 thread_t td = curthread;
f1d1c3fa 973
f1d1c3fa 974 lwkt_gettoken(&w->wa_token);
ae8050a4 975 if (w->wa_gen == *gen) {
f1d1c3fa
MD
976 _lwkt_dequeue(td);
977 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
978 ++w->wa_count;
979 td->td_wait = w;
ae8050a4 980 td->td_wmesg = wmesg;
ece04fd0 981again:
f1d1c3fa 982 lwkt_switch();
ece04fd0
MD
983 lwkt_regettoken(&w->wa_token);
984 if (td->td_wmesg != NULL) {
985 _lwkt_dequeue(td);
986 goto again;
987 }
8ad65e08 988 }
ae8050a4
MD
989 /* token might be lost, doesn't matter for gen update */
990 *gen = w->wa_gen;
f1d1c3fa
MD
991 lwkt_reltoken(&w->wa_token);
992}
993
994/*
995 * Signal a wait queue. We gain ownership of the wait queue in order to
996 * signal it. Once a thread is removed from the wait queue we have to
997 * deal with the cpu owning the thread.
998 *
999 * Note: alternatively we could message the target cpu owning the wait
1000 * queue. YYY implement as sysctl.
1001 */
1002void
ece04fd0 1003lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
1004{
1005 thread_t td;
1006 int count;
1007
1008 lwkt_gettoken(&w->wa_token);
1009 ++w->wa_gen;
ece04fd0
MD
1010 if (count < 0)
1011 count = w->wa_count;
f1d1c3fa
MD
1012 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1013 --count;
1014 --w->wa_count;
1015 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1016 td->td_wait = NULL;
ae8050a4 1017 td->td_wmesg = NULL;
a72187e9 1018 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1019 _lwkt_enqueue(td);
1020 } else {
a72187e9 1021 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
1022 }
1023 lwkt_regettoken(&w->wa_token);
1024 }
1025 lwkt_reltoken(&w->wa_token);
1026}
1027
ece04fd0
MD
1028#endif
1029
f1d1c3fa 1030/*
96728c05 1031 * Acquire ownership of a token
f1d1c3fa 1032 *
96728c05 1033 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
1034 * section side effects, depending on its purpose. These side effects
1035 * guarentee that you will maintain ownership of the token as long as you
1036 * do not block. If you block you may lose access to the token (but you
1037 * must still release it even if you lose your access to it).
1038 *
96728c05 1039 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 1040 * a token, but do we really only need to disable IPIs ?
96728c05
MD
1041 *
1042 * YYY certain tokens could be made to act like mutexes when performance
1043 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1044 *
a2a5ad0d
MD
1045 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1046 * means that 4.x does not expect a switch so for now we cannot switch
1047 * when waiting for an IPI to be returned.
1048 *
1049 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
1050 * it and then block. The IPI causes the token to be given away to the
1051 * requesting cpu, unless it has already changed hands. Since only the
1052 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 1053 * This needs serious optimization.
f1d1c3fa 1054 */
57c254db
MD
1055
1056#ifdef SMP
1057
96728c05
MD
1058static
1059void
1060lwkt_gettoken_remote(void *arg)
1061{
1062 lwkt_gettoken_req *req = arg;
1063 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 1064#ifdef INVARIANTS
a2a5ad0d
MD
1065 if (token_debug)
1066 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 1067#endif
96728c05 1068 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
1069 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1070 /* else set reqcpu to point to current cpu for release */
96728c05
MD
1071 }
1072}
1073
57c254db
MD
1074#endif
1075
8a8d5d85 1076int
f1d1c3fa
MD
1077lwkt_gettoken(lwkt_token_t tok)
1078{
1079 /*
1080 * Prevent preemption so the token can't be taken away from us once
1081 * we gain ownership of it. Use a synchronous request which might
1082 * block. The request will be forwarded as necessary playing catchup
1083 * to the token.
1084 */
96728c05 1085
f1d1c3fa 1086 crit_enter();
57c254db 1087#ifdef INVARIANTS
1d4c9574
MD
1088 if (curthread->td_pri > 1800) {
1089 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1090 tok, ((int **)&tok)[-1]);
1091 }
a2a5ad0d
MD
1092 if (curthread->td_pri > 2000) {
1093 curthread->td_pri = 1000;
1094 panic("too HIGH!");
57c254db
MD
1095 }
1096#endif
96728c05 1097#ifdef SMP
d0e06f83 1098 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1099 struct lwkt_gettoken_req req;
1100 int seq;
96728c05
MD
1101 int dcpu;
1102
1103 req.cpu = mycpu->gd_cpuid;
1104 req.tok = tok;
1105 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1106 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1107#ifdef INVARIANTS
a2a5ad0d
MD
1108 if (token_debug)
1109 printf("REQT%d ", dcpu);
634081ff 1110#endif
96728c05
MD
1111 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1112 lwkt_wait_ipiq(dcpu, seq);
634081ff 1113#ifdef INVARIANTS
a2a5ad0d
MD
1114 if (token_debug)
1115 printf("REQR%d ", tok->t_cpu);
634081ff 1116#endif
f1d1c3fa
MD
1117 }
1118#endif
1119 /*
1120 * leave us in a critical section on return. This will be undone
8a8d5d85 1121 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1122 */
8a8d5d85 1123 return(++tok->t_gen);
f1d1c3fa
MD
1124}
1125
96728c05
MD
1126/*
1127 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1128 * failure.
1129 */
1130int
1131lwkt_trytoken(lwkt_token_t tok)
1132{
1133 crit_enter();
1134#ifdef SMP
1135 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1136 crit_exit();
96728c05
MD
1137 return(0);
1138 }
1139#endif
1140 /* leave us in the critical section */
1141 ++tok->t_gen;
1142 return(1);
1143}
1144
f1d1c3fa
MD
1145/*
1146 * Release your ownership of a token. Releases must occur in reverse
1147 * order to aquisitions, eventually so priorities can be unwound properly
1148 * like SPLs. At the moment the actual implemention doesn't care.
1149 *
1150 * We can safely hand a token that we own to another cpu without notifying
1151 * it, but once we do we can't get it back without requesting it (unless
1152 * the other cpu hands it back to us before we check).
1153 *
1154 * We might have lost the token, so check that.
7ba9c17c
MD
1155 *
1156 * Return the token's generation number. The number is useful to callers
1157 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1158 */
7ba9c17c 1159int
f1d1c3fa
MD
1160lwkt_reltoken(lwkt_token_t tok)
1161{
7ba9c17c
MD
1162 int gen;
1163
d0e06f83 1164 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1165 tok->t_cpu = tok->t_reqcpu;
1166 }
7ba9c17c 1167 gen = tok->t_gen;
f1d1c3fa 1168 crit_exit();
7ba9c17c 1169 return(gen);
f1d1c3fa
MD
1170}
1171
1172/*
7ba9c17c
MD
1173 * Reacquire a token that might have been lost. 0 is returned if the
1174 * generation has not changed (nobody stole the token from us), -1 is
1175 * returned otherwise. The token is reacquired regardless but the
1176 * generation number is not bumped further if we already own the token.
8a8d5d85 1177 *
7ba9c17c
MD
1178 * For efficiency we inline the best-case situation for lwkt_regettoken()
1179 * (i.e .we still own the token).
8a8d5d85
MD
1180 */
1181int
1182lwkt_gentoken(lwkt_token_t tok, int *gen)
1183{
7ba9c17c 1184 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1185 return(0);
7ba9c17c
MD
1186 *gen = lwkt_regettoken(tok);
1187 return(-1);
8a8d5d85
MD
1188}
1189
8a8d5d85 1190/*
435ff993
MD
1191 * Re-acquire a token that might have been lost. The generation number
1192 * is bumped and returned regardless of whether the token had been lost
1193 * or not (because we only have cpu granularity we have to bump the token
1194 * either way).
f1d1c3fa
MD
1195 */
1196int
1197lwkt_regettoken(lwkt_token_t tok)
1198{
96728c05 1199 /* assert we are in a critical section */
d0e06f83 1200 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1201#ifdef SMP
d0e06f83 1202 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1203 struct lwkt_gettoken_req req;
1204 int seq;
96728c05 1205 int dcpu;
57c254db 1206
96728c05
MD
1207 req.cpu = mycpu->gd_cpuid;
1208 req.tok = tok;
1209 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1210 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1211#ifdef INVARIANTS
cb973d15
MD
1212 if (token_debug)
1213 printf("REQT%d ", dcpu);
634081ff 1214#endif
96728c05
MD
1215 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1216 lwkt_wait_ipiq(dcpu, seq);
634081ff 1217#ifdef INVARIATNS
cb973d15
MD
1218 if (token_debug)
1219 printf("REQR%d ", tok->t_cpu);
634081ff 1220#endif
f1d1c3fa 1221 }
f1d1c3fa 1222#endif
96728c05 1223 }
435ff993 1224 ++tok->t_gen;
8a8d5d85 1225 return(tok->t_gen);
8ad65e08
MD
1226}
1227
72740893
MD
1228void
1229lwkt_inittoken(lwkt_token_t tok)
1230{
1231 /*
1232 * Zero structure and set cpu owner and reqcpu to cpu 0.
1233 */
1234 bzero(tok, sizeof(*tok));
1235}
1236
99df837e
MD
1237/*
1238 * Create a kernel process/thread/whatever. It shares it's address space
1239 * with proc0 - ie: kernel only.
1240 *
365fa13f
MD
1241 * NOTE! By default new threads are created with the MP lock held. A
1242 * thread which does not require the MP lock should release it by calling
1243 * rel_mplock() at the start of the new thread.
99df837e
MD
1244 */
1245int
1246lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1247 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1248 const char *fmt, ...)
99df837e 1249{
73e4f7b9 1250 thread_t td;
e2565a42 1251 __va_list ap;
99df837e 1252
75cdbe6c 1253 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1254 if (tdp)
1255 *tdp = td;
99df837e 1256 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1257 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1258#ifdef SMP
1259 td->td_mpcount = 1;
1260#endif
99df837e
MD
1261
1262 /*
1263 * Set up arg0 for 'ps' etc
1264 */
e2565a42 1265 __va_start(ap, fmt);
99df837e 1266 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1267 __va_end(ap);
99df837e
MD
1268
1269 /*
1270 * Schedule the thread to run
1271 */
ef0fdad1
MD
1272 if ((td->td_flags & TDF_STOPREQ) == 0)
1273 lwkt_schedule(td);
1274 else
1275 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1276 return 0;
1277}
1278
1279/*
1280 * Destroy an LWKT thread. Warning! This function is not called when
1281 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1282 * uses a different reaping mechanism.
1283 */
1284void
1285lwkt_exit(void)
1286{
1287 thread_t td = curthread;
1288
1289 if (td->td_flags & TDF_VERBOSE)
1290 printf("kthread %p %s has exited\n", td, td->td_comm);
1291 crit_enter();
1292 lwkt_deschedule_self();
1293 ++mycpu->gd_tdfreecount;
1294 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1295 cpu_thread_exit();
1296}
1297
05220613 1298#ifdef _KERNEL
99df837e
MD
1299/*
1300 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1301 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1302 *
1303 * NOTE! By default kthreads are created with the MP lock held. A
1304 * thread which does not require the MP lock should release it by calling
1305 * rel_mplock() at the start of the new thread.
99df837e
MD
1306 */
1307int
1308kthread_create(void (*func)(void *), void *arg,
1309 struct thread **tdp, const char *fmt, ...)
1310{
73e4f7b9 1311 thread_t td;
e2565a42 1312 __va_list ap;
99df837e 1313
75cdbe6c 1314 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1315 if (tdp)
1316 *tdp = td;
99df837e
MD
1317 cpu_set_thread_handler(td, kthread_exit, func, arg);
1318 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1319#ifdef SMP
1320 td->td_mpcount = 1;
1321#endif
99df837e
MD
1322
1323 /*
1324 * Set up arg0 for 'ps' etc
1325 */
e2565a42 1326 __va_start(ap, fmt);
99df837e 1327 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1328 __va_end(ap);
99df837e
MD
1329
1330 /*
1331 * Schedule the thread to run
1332 */
1333 lwkt_schedule(td);
1334 return 0;
1335}
1336
05220613
MD
1337#endif
1338
26a0694b
MD
1339void
1340crit_panic(void)
1341{
73e4f7b9 1342 thread_t td = curthread;
26a0694b
MD
1343 int lpri = td->td_pri;
1344
1345 td->td_pri = 0;
1346 panic("td_pri is/would-go negative! %p %d", td, lpri);
1347}
1348
99df837e
MD
1349/*
1350 * Destroy an LWKT thread. Warning! This function is not called when
1351 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1352 * uses a different reaping mechanism.
1353 *
1354 * XXX duplicates lwkt_exit()
1355 */
1356void
1357kthread_exit(void)
1358{
1359 lwkt_exit();
1360}
1361
96728c05
MD
1362#ifdef SMP
1363
1364/*
1365 * Send a function execution request to another cpu. The request is queued
1366 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1367 * possible target cpu. The FIFO can be written.
1368 *
1369 * YYY If the FIFO fills up we have to enable interrupts and process the
1370 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1371 * Create a CPU_*() function to do this!
1372 *
46a3f46d
MD
1373 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1374 * end will take care of any pending interrupts.
1375 *
96728c05
MD
1376 * Must be called from a critical section.
1377 */
1378int
1379lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1380{
1381 lwkt_ipiq_t ip;
1382 int windex;
a2a5ad0d 1383 struct globaldata *gd = mycpu;
96728c05 1384
a2a5ad0d 1385 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1386 func(arg);
1387 return(0);
1388 }
cb973d15 1389 crit_enter();
a2a5ad0d
MD
1390 ++gd->gd_intr_nesting_level;
1391#ifdef INVARIANTS
1392 if (gd->gd_intr_nesting_level > 20)
1393 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1394#endif
96728c05
MD
1395 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1396 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1397 ++ipiq_count;
a2a5ad0d 1398 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1399
1400 /*
1401 * We always drain before the FIFO becomes full so it should never
1402 * become full. We need to leave enough entries to deal with
1403 * reentrancy.
1404 */
1405 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1406 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1407 ip->ip_func[windex] = func;
1408 ip->ip_arg[windex] = arg;
1409 /* YYY memory barrier */
1410 ++ip->ip_windex;
96728c05
MD
1411 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1412 unsigned int eflags = read_eflags();
1413 cpu_enable_intr();
1414 ++ipiq_fifofull;
cb973d15 1415 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1416 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1417 lwkt_process_ipiq();
1418 }
1419 write_eflags(eflags);
1420 }
a2a5ad0d 1421 --gd->gd_intr_nesting_level;
96728c05 1422 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1423 crit_exit();
96728c05
MD
1424 return(ip->ip_windex);
1425}
1426
cb973d15
MD
1427/*
1428 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1429 * The message will not be sent to stopped cpus.
cb973d15
MD
1430 */
1431void
1432lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1433{
1434 int cpuid;
1435
435ff993 1436 mask &= ~stopped_cpus;
cb973d15
MD
1437 while (mask) {
1438 cpuid = bsfl(mask);
1439 lwkt_send_ipiq(cpuid, func, arg);
1440 mask &= ~(1 << cpuid);
1441 }
1442}
1443
96728c05
MD
1444/*
1445 * Wait for the remote cpu to finish processing a function.
1446 *
1447 * YYY we have to enable interrupts and process the IPIQ while waiting
1448 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1449 * function to do this! YYY we really should 'block' here.
1450 *
1451 * Must be called from a critical section. Thsi routine may be called
1452 * from an interrupt (for example, if an interrupt wakes a foreign thread
1453 * up).
1454 */
1455void
1456lwkt_wait_ipiq(int dcpu, int seq)
1457{
1458 lwkt_ipiq_t ip;
a2a5ad0d 1459 int maxc = 100000000;
96728c05
MD
1460
1461 if (dcpu != mycpu->gd_cpuid) {
1462 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1463 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1464 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1465 unsigned int eflags = read_eflags();
1466 cpu_enable_intr();
cb973d15 1467 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1468 lwkt_process_ipiq();
a2a5ad0d 1469 if (--maxc == 0)
cb973d15 1470 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1471 if (maxc < -1000000)
1472 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1473 }
1474 write_eflags(eflags);
1475 }
1476 }
1477}
1478
1479/*
1480 * Called from IPI interrupt (like a fast interrupt), which has placed
1481 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1482 * May also be called from doreti or splz, or be reentrantly called
1483 * indirectly through the ip_func[] we run.
96728c05
MD
1484 */
1485void
1486lwkt_process_ipiq(void)
1487{
1488 int n;
1489 int cpuid = mycpu->gd_cpuid;
1490
1491 for (n = 0; n < ncpus; ++n) {
1492 lwkt_ipiq_t ip;
1493 int ri;
1494
1495 if (n == cpuid)
1496 continue;
1497 ip = globaldata_find(n)->gd_ipiq;
1498 if (ip == NULL)
1499 continue;
1500 ip = &ip[cpuid];
cb973d15
MD
1501
1502 /*
1503 * Note: xindex is only updated after we are sure the function has
1504 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1505 * function may send an IPI which may block/drain.
1506 */
96728c05
MD
1507 while (ip->ip_rindex != ip->ip_windex) {
1508 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1509 ++ip->ip_rindex;
cb973d15
MD
1510 ip->ip_func[ri](ip->ip_arg[ri]);
1511 /* YYY memory barrier */
1512 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1513 }
1514 }
1515}
1516
1517#else
1518
1519int
1520lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1521{
1522 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1523 return(0); /* NOT REACHED */
1524}
1525
1526void
1527lwkt_wait_ipiq(int dcpu, int seq)
1528{
1529 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1530}
1531
1532#endif