Forward FAST interrupts to the MP lock holder + minor fixes.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08
MD
1/*
2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 *
f1d1c3fa
MD
26 * Each cpu in a system has its own self-contained light weight kernel
27 * thread scheduler, which means that generally speaking we only need
96728c05
MD
28 * to use a critical section to avoid problems. Foreign thread
29 * scheduling is queued via (async) IPIs.
f1d1c3fa 30 *
3c23a41a 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.22 2003/07/11 22:30:09 dillon Exp $
8ad65e08
MD
32 */
33
34#include <sys/param.h>
35#include <sys/systm.h>
36#include <sys/kernel.h>
37#include <sys/proc.h>
38#include <sys/rtprio.h>
39#include <sys/queue.h>
f1d1c3fa 40#include <sys/thread2.h>
7d0bac62 41#include <sys/sysctl.h>
99df837e 42#include <sys/kthread.h>
f1d1c3fa 43#include <machine/cpu.h>
99df837e 44#include <sys/lock.h>
f1d1c3fa 45
7d0bac62
MD
46#include <vm/vm.h>
47#include <vm/vm_param.h>
48#include <vm/vm_kern.h>
49#include <vm/vm_object.h>
50#include <vm/vm_page.h>
51#include <vm/vm_map.h>
52#include <vm/vm_pager.h>
53#include <vm/vm_extern.h>
54#include <vm/vm_zone.h>
55
99df837e 56#include <machine/stdarg.h>
57c254db 57#include <machine/ipl.h>
96728c05
MD
58#ifdef SMP
59#include <machine/smp.h>
60#endif
99df837e 61
7d0bac62 62static int untimely_switch = 0;
4b5f931b 63SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
57c254db
MD
64#ifdef INVARIANTS
65static int token_debug = 0;
66SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
67#endif
4b5f931b
MD
68static quad_t switch_count = 0;
69SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
70static quad_t preempt_hit = 0;
71SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
72static quad_t preempt_miss = 0;
73SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b
MD
74static quad_t preempt_weird = 0;
75SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05
MD
76static quad_t ipiq_count = 0;
77SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
78static quad_t ipiq_fifofull = 0;
79SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 80
4b5f931b
MD
81/*
82 * These helper procedures handle the runq, they can only be called from
83 * within a critical section.
84 */
f1d1c3fa
MD
85static __inline
86void
87_lwkt_dequeue(thread_t td)
88{
89 if (td->td_flags & TDF_RUNQ) {
4b5f931b
MD
90 int nq = td->td_pri & TDPRI_MASK;
91 struct globaldata *gd = mycpu;
92
f1d1c3fa 93 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
94 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
95 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
96 }
97}
98
99static __inline
100void
101_lwkt_enqueue(thread_t td)
102{
103 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b
MD
104 int nq = td->td_pri & TDPRI_MASK;
105 struct globaldata *gd = mycpu;
106
f1d1c3fa 107 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
108 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
109 gd->gd_runqmask |= 1 << nq;
26a0694b
MD
110#if 0
111 /*
112 * YYY needs cli/sti protection? gd_reqpri set by interrupt
113 * when made pending. need better mechanism.
114 */
115 if (gd->gd_reqpri < (td->td_pri & TDPRI_MASK))
116 gd->gd_reqpri = (td->td_pri & TDPRI_MASK);
117#endif
f1d1c3fa
MD
118 }
119}
8ad65e08 120
57c254db
MD
121static __inline
122int
123_lwkt_wantresched(thread_t ntd, thread_t cur)
124{
125 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
126}
127
8ad65e08
MD
128/*
129 * LWKTs operate on a per-cpu basis
130 *
73e4f7b9 131 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
132 */
133void
134lwkt_gdinit(struct globaldata *gd)
135{
4b5f931b
MD
136 int i;
137
138 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
139 TAILQ_INIT(&gd->gd_tdrunq[i]);
140 gd->gd_runqmask = 0;
73e4f7b9 141 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
142}
143
7d0bac62
MD
144/*
145 * Initialize a thread wait structure prior to first use.
146 *
147 * NOTE! called from low level boot code, we cannot do anything fancy!
148 */
149void
150lwkt_init_wait(lwkt_wait_t w)
151{
152 TAILQ_INIT(&w->wa_waitq);
153}
154
155/*
156 * Create a new thread. The thread must be associated with a process context
157 * or LWKT start address before it can be scheduled.
0cfcada1
MD
158 *
159 * If you intend to create a thread without a process context this function
160 * does everything except load the startup and switcher function.
7d0bac62
MD
161 */
162thread_t
ef0fdad1 163lwkt_alloc_thread(struct thread *td)
7d0bac62 164{
99df837e 165 void *stack;
ef0fdad1 166 int flags = 0;
7d0bac62 167
ef0fdad1 168 if (td == NULL) {
26a0694b 169 crit_enter();
ef0fdad1
MD
170 if (mycpu->gd_tdfreecount > 0) {
171 --mycpu->gd_tdfreecount;
172 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 173 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
174 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
175 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
176 crit_exit();
177 stack = td->td_kstack;
178 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
179 } else {
180 crit_exit();
181 td = zalloc(thread_zone);
182 td->td_kstack = NULL;
183 flags |= TDF_ALLOCATED_THREAD;
184 }
185 }
186 if ((stack = td->td_kstack) == NULL) {
99df837e 187 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 188 flags |= TDF_ALLOCATED_STACK;
99df837e 189 }
26a0694b 190 lwkt_init_thread(td, stack, flags, mycpu);
99df837e 191 return(td);
7d0bac62
MD
192}
193
194/*
195 * Initialize a preexisting thread structure. This function is used by
196 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
197 *
198 * NOTE! called from low level boot code, we cannot do anything fancy!
199 */
200void
26a0694b 201lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 202{
99df837e
MD
203 bzero(td, sizeof(struct thread));
204 td->td_kstack = stack;
205 td->td_flags |= flags;
26a0694b
MD
206 td->td_gd = gd;
207 td->td_pri = TDPRI_CRIT;
d9eea1a5 208 td->td_cpu = gd->gd_cpuid; /* YYY don't really need this if have td_gd */
99df837e 209 pmap_init_thread(td);
73e4f7b9
MD
210 crit_enter();
211 TAILQ_INSERT_TAIL(&mycpu->gd_tdallq, td, td_allq);
212 crit_exit();
213}
214
215void
216lwkt_set_comm(thread_t td, const char *ctl, ...)
217{
218 va_list va;
219
220 va_start(va, ctl);
221 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
222 va_end(va);
7d0bac62
MD
223}
224
99df837e 225void
73e4f7b9 226lwkt_hold(thread_t td)
99df837e 227{
73e4f7b9
MD
228 ++td->td_refs;
229}
230
231void
232lwkt_rele(thread_t td)
233{
234 KKASSERT(td->td_refs > 0);
235 --td->td_refs;
236}
237
238void
239lwkt_wait_free(thread_t td)
240{
241 while (td->td_refs)
242 tsleep(td, PWAIT, "tdreap", hz);
243}
244
245void
246lwkt_free_thread(thread_t td)
247{
248 struct globaldata *gd = mycpu;
249
d9eea1a5 250 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
251 ("lwkt_free_thread: did not exit! %p", td));
252
253 crit_enter();
73e4f7b9
MD
254 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
255 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
256 (td->td_flags & TDF_ALLOCATED_THREAD)
257 ) {
73e4f7b9
MD
258 ++gd->gd_tdfreecount;
259 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
260 crit_exit();
261 } else {
262 crit_exit();
263 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
264 kmem_free(kernel_map,
265 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 266 /* gd invalid */
99df837e
MD
267 td->td_kstack = NULL;
268 }
269 if (td->td_flags & TDF_ALLOCATED_THREAD)
270 zfree(thread_zone, td);
271 }
272}
273
274
8ad65e08
MD
275/*
276 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
277 * switch to the idlethread. Switching must occur within a critical
278 * section to avoid races with the scheduling queue.
279 *
280 * We always have full control over our cpu's run queue. Other cpus
281 * that wish to manipulate our queue must use the cpu_*msg() calls to
282 * talk to our cpu, so a critical section is all that is needed and
283 * the result is very, very fast thread switching.
284 *
96728c05
MD
285 * The LWKT scheduler uses a fixed priority model and round-robins at
286 * each priority level. User process scheduling is a totally
287 * different beast and LWKT priorities should not be confused with
288 * user process priorities.
f1d1c3fa 289 *
96728c05
MD
290 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
291 * cleans it up. Note that the td_switch() function cannot do anything that
292 * requires the MP lock since the MP lock will have already been setup for
293 * the target thread (not the current thread).
8ad65e08 294 */
96728c05 295
8ad65e08
MD
296void
297lwkt_switch(void)
298{
4b5f931b 299 struct globaldata *gd;
f1d1c3fa 300 thread_t td = curthread;
8ad65e08 301 thread_t ntd;
8a8d5d85
MD
302#ifdef SMP
303 int mpheld;
304#endif
8ad65e08 305
96728c05
MD
306 if (mycpu->gd_intr_nesting_level &&
307 td->td_preempted == NULL && panicstr == NULL
308 ) {
26a0694b 309 panic("lwkt_switch: cannot switch from within an interrupt, yet\n");
96728c05 310 }
ef0fdad1 311
cb973d15
MD
312 /*
313 * Passive release (used to transition from user to kernel mode
314 * when we block or switch rather then when we enter the kernel).
315 * This function is NOT called if we are switching into a preemption
316 * or returning from a preemption. Typically this causes us to lose
317 * our P_CURPROC designation (if we have one) and become a true LWKT
318 * thread, and may also hand P_CURPROC to another process and schedule
319 * its thread.
320 */
321 if (td->td_release)
322 td->td_release(td);
323
f1d1c3fa 324 crit_enter();
4b5f931b 325 ++switch_count;
8a8d5d85
MD
326
327#ifdef SMP
328 /*
329 * td_mpcount cannot be used to determine if we currently hold the
330 * MP lock because get_mplock() will increment it prior to attempting
331 * to get the lock, and switch out if it can't. Look at the actual lock.
332 */
333 mpheld = MP_LOCK_HELD();
334#endif
99df837e
MD
335 if ((ntd = td->td_preempted) != NULL) {
336 /*
337 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
338 * thread. This occurs transparently, whether the preempted thread
339 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
340 * itself).
341 *
342 * We have to setup the MP lock for the original thread after backing
343 * out the adjustment that was made to curthread when the original
344 * was preempted.
99df837e 345 */
26a0694b 346 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 347#ifdef SMP
96728c05
MD
348 if (ntd->td_mpcount && mpheld == 0) {
349 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
350 td, ntd, td->td_mpcount, ntd->td_mpcount);
351 }
8a8d5d85
MD
352 if (ntd->td_mpcount) {
353 td->td_mpcount -= ntd->td_mpcount;
354 KKASSERT(td->td_mpcount >= 0);
355 }
356#endif
26a0694b 357 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 358 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 359 } else {
4b5f931b
MD
360 /*
361 * Priority queue / round-robin at each priority. Note that user
362 * processes run at a fixed, low priority and the user process
363 * scheduler deals with interactions between user processes
364 * by scheduling and descheduling them from the LWKT queue as
365 * necessary.
8a8d5d85
MD
366 *
367 * We have to adjust the MP lock for the target thread. If we
368 * need the MP lock and cannot obtain it we try to locate a
369 * thread that does not need the MP lock.
4b5f931b
MD
370 */
371 gd = mycpu;
4b5f931b
MD
372again:
373 if (gd->gd_runqmask) {
374 int nq = bsrl(gd->gd_runqmask);
375 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
376 gd->gd_runqmask &= ~(1 << nq);
377 goto again;
378 }
8a8d5d85
MD
379#ifdef SMP
380 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
381 /*
96728c05
MD
382 * Target needs MP lock and we couldn't get it, try
383 * to locate a thread which does not need the MP lock
3c23a41a 384 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
385 */
386 u_int32_t rqmask = gd->gd_runqmask;
387 while (rqmask) {
388 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
389 if (ntd->td_mpcount == 0)
390 break;
391 }
392 if (ntd)
393 break;
394 rqmask &= ~(1 << nq);
395 nq = bsrl(rqmask);
396 }
397 if (ntd == NULL) {
a2a5ad0d
MD
398 ntd = &gd->gd_idlethread;
399 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
400 } else {
401 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
402 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
403 }
404 } else {
405 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
406 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
407 }
408#else
4b5f931b
MD
409 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
410 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 411#endif
4b5f931b 412 } else {
3c23a41a
MD
413 /*
414 * Nothing to run but we may still need the BGL to deal with
415 * pending interrupts, spin in idle if so.
416 */
a2a5ad0d 417 ntd = &gd->gd_idlethread;
3c23a41a
MD
418 if (gd->gd_reqpri)
419 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 420 }
f1d1c3fa 421 }
26a0694b
MD
422 KASSERT(ntd->td_pri >= TDPRI_CRIT,
423 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
424
425 /*
426 * Do the actual switch. If the new target does not need the MP lock
427 * and we are holding it, release the MP lock. If the new target requires
428 * the MP lock we have already acquired it for the target.
429 */
430#ifdef SMP
431 if (ntd->td_mpcount == 0 ) {
432 if (MP_LOCK_HELD())
433 cpu_rel_mplock();
434 } else {
435 ASSERT_MP_LOCK_HELD();
436 }
437#endif
8a8d5d85 438 if (td != ntd) {
f1d1c3fa 439 td->td_switch(ntd);
8a8d5d85 440 }
96728c05 441
f1d1c3fa 442 crit_exit();
8ad65e08
MD
443}
444
cb973d15
MD
445/*
446 * Switch if another thread has a higher priority. Do not switch to other
447 * threads at the same priority.
448 */
449void
450lwkt_maybe_switch()
451{
452 struct globaldata *gd = mycpu;
453 struct thread *td = gd->gd_curthread;
454
455 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
456 lwkt_switch();
457 }
458}
459
b68b7282 460/*
96728c05
MD
461 * Request that the target thread preempt the current thread. Preemption
462 * only works under a specific set of conditions:
b68b7282 463 *
96728c05
MD
464 * - We are not preempting ourselves
465 * - The target thread is owned by the current cpu
466 * - We are not currently being preempted
467 * - The target is not currently being preempted
468 * - We are able to satisfy the target's MP lock requirements (if any).
469 *
470 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
471 * this is called via lwkt_schedule() through the td_preemptable callback.
472 * critpri is the managed critical priority that we should ignore in order
473 * to determine whether preemption is possible (aka usually just the crit
474 * priority of lwkt_schedule() itself).
b68b7282 475 *
26a0694b
MD
476 * XXX at the moment we run the target thread in a critical section during
477 * the preemption in order to prevent the target from taking interrupts
478 * that *WE* can't. Preemption is strictly limited to interrupt threads
479 * and interrupt-like threads, outside of a critical section, and the
480 * preempted source thread will be resumed the instant the target blocks
481 * whether or not the source is scheduled (i.e. preemption is supposed to
482 * be as transparent as possible).
4b5f931b 483 *
8a8d5d85
MD
484 * The target thread inherits our MP count (added to its own) for the
485 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
486 * MP lock during the preemption. Therefore, any preempting targets must be
487 * careful in regards to MP assertions. Note that the MP count may be
488 * out of sync with the physical mp_lock. If we preempt we have to preserve
489 * the expected situation.
b68b7282
MD
490 */
491void
96728c05 492lwkt_preempt(thread_t ntd, int critpri)
b68b7282 493{
73e4f7b9 494 thread_t td = curthread;
8a8d5d85
MD
495#ifdef SMP
496 int mpheld;
57c254db 497 int savecnt;
8a8d5d85 498#endif
b68b7282 499
26a0694b 500 /*
96728c05
MD
501 * The caller has put us in a critical section. We can only preempt
502 * if the caller of the caller was not in a critical section (basically
57c254db
MD
503 * a local interrupt), as determined by the 'critpri' parameter. If
504 * we are unable to preempt
96728c05
MD
505 *
506 * YYY The target thread must be in a critical section (else it must
507 * inherit our critical section? I dunno yet).
26a0694b
MD
508 */
509 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 510
cb973d15 511 need_resched();
57c254db
MD
512 if (!_lwkt_wantresched(ntd, td)) {
513 ++preempt_miss;
514 return;
515 }
96728c05
MD
516 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
517 ++preempt_miss;
518 return;
519 }
520#ifdef SMP
521 if (ntd->td_cpu != mycpu->gd_cpuid) {
522 ++preempt_miss;
523 return;
524 }
525#endif
26a0694b
MD
526 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
527 ++preempt_weird;
528 return;
529 }
530 if (ntd->td_preempted) {
4b5f931b 531 ++preempt_hit;
26a0694b 532 return;
b68b7282 533 }
8a8d5d85 534#ifdef SMP
a2a5ad0d
MD
535 /*
536 * note: an interrupt might have occured just as we were transitioning
a5934754
MD
537 * to the MP lock. In this case td_mpcount will be pre-disposed but
538 * not actually synchronized with the actual state of the lock. We
539 * can use it to imply an MP lock requirement for the preemption but
540 * we cannot use it to test whether we hold the MP lock or not.
a2a5ad0d 541 */
a5934754 542 mpheld = MP_LOCK_HELD();
96728c05
MD
543 if (mpheld && td->td_mpcount == 0)
544 panic("lwkt_preempt(): held and no count");
545 savecnt = td->td_mpcount;
8a8d5d85
MD
546 ntd->td_mpcount += td->td_mpcount;
547 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
548 ntd->td_mpcount -= td->td_mpcount;
549 ++preempt_miss;
550 return;
551 }
552#endif
26a0694b
MD
553
554 ++preempt_hit;
555 ntd->td_preempted = td;
556 td->td_flags |= TDF_PREEMPT_LOCK;
557 td->td_switch(ntd);
558 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
559#ifdef SMP
560 KKASSERT(savecnt == td->td_mpcount);
561 if (mpheld == 0 && MP_LOCK_HELD())
562 cpu_rel_mplock();
563 else if (mpheld && !MP_LOCK_HELD())
564 panic("lwkt_preempt(): MP lock was not held through");
565#endif
26a0694b
MD
566 ntd->td_preempted = NULL;
567 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
568}
569
f1d1c3fa
MD
570/*
571 * Yield our thread while higher priority threads are pending. This is
572 * typically called when we leave a critical section but it can be safely
573 * called while we are in a critical section.
574 *
575 * This function will not generally yield to equal priority threads but it
576 * can occur as a side effect. Note that lwkt_switch() is called from
577 * inside the critical section to pervent its own crit_exit() from reentering
578 * lwkt_yield_quick().
579 *
ef0fdad1
MD
580 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
581 * came along but was blocked and made pending.
582 *
f1d1c3fa
MD
583 * (self contained on a per cpu basis)
584 */
585void
586lwkt_yield_quick(void)
587{
588 thread_t td = curthread;
ef0fdad1 589
a2a5ad0d
MD
590 /*
591 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
592 * it with a non-zero cpl then we might not wind up calling splz after
593 * a task switch when the critical section is exited even though the
594 * new task could accept the interrupt. YYY alternative is to have
595 * lwkt_switch() just call splz unconditionally.
596 *
597 * XXX from crit_exit() only called after last crit section is released.
598 * If called directly will run splz() even if in a critical section.
599 */
ef0fdad1 600 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
f1d1c3fa
MD
601 splz();
602 }
603
604 /*
605 * YYY enabling will cause wakeup() to task-switch, which really
606 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
607 * preemption and MP without actually doing preemption or MP, because a
608 * lot of code assumes that wakeup() does not block.
f1d1c3fa 609 */
ef0fdad1 610 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
f1d1c3fa
MD
611 crit_enter();
612 /*
613 * YYY temporary hacks until we disassociate the userland scheduler
614 * from the LWKT scheduler.
615 */
616 if (td->td_flags & TDF_RUNQ) {
617 lwkt_switch(); /* will not reenter yield function */
618 } else {
619 lwkt_schedule_self(); /* make sure we are scheduled */
620 lwkt_switch(); /* will not reenter yield function */
621 lwkt_deschedule_self(); /* make sure we are descheduled */
622 }
623 crit_exit_noyield();
624 }
f1d1c3fa
MD
625}
626
8ad65e08 627/*
f1d1c3fa
MD
628 * This implements a normal yield which, unlike _quick, will yield to equal
629 * priority threads as well. Note that gd_reqpri tests will be handled by
630 * the crit_exit() call in lwkt_switch().
631 *
632 * (self contained on a per cpu basis)
8ad65e08
MD
633 */
634void
f1d1c3fa 635lwkt_yield(void)
8ad65e08 636{
f1d1c3fa
MD
637 lwkt_schedule_self();
638 lwkt_switch();
639}
640
641/*
642 * Schedule a thread to run. As the current thread we can always safely
643 * schedule ourselves, and a shortcut procedure is provided for that
644 * function.
645 *
646 * (non-blocking, self contained on a per cpu basis)
647 */
648void
649lwkt_schedule_self(void)
650{
651 thread_t td = curthread;
652
653 crit_enter();
654 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 655 _lwkt_enqueue(td);
26a0694b
MD
656 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
657 panic("SCHED SELF PANIC");
f1d1c3fa 658 crit_exit();
8ad65e08 659}
8ad65e08
MD
660
661/*
f1d1c3fa
MD
662 * Generic schedule. Possibly schedule threads belonging to other cpus and
663 * deal with threads that might be blocked on a wait queue.
664 *
96728c05 665 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
666 * Load balancing can be accomplished by requesting other sorts of actions
667 * for the thread in question.
8ad65e08
MD
668 */
669void
670lwkt_schedule(thread_t td)
671{
96728c05 672#ifdef INVARIANTS
26a0694b
MD
673 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
674 && td->td_proc->p_stat == SSLEEP
675 ) {
676 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
677 curthread,
678 curthread->td_proc ? curthread->td_proc->p_pid : -1,
679 curthread->td_proc ? curthread->td_proc->p_stat : -1,
680 td,
681 td->td_proc ? curthread->td_proc->p_pid : -1,
682 td->td_proc ? curthread->td_proc->p_stat : -1
683 );
684 panic("SCHED PANIC");
685 }
96728c05 686#endif
f1d1c3fa
MD
687 crit_enter();
688 if (td == curthread) {
689 _lwkt_enqueue(td);
690 } else {
691 lwkt_wait_t w;
692
693 /*
694 * If the thread is on a wait list we have to send our scheduling
695 * request to the owner of the wait structure. Otherwise we send
696 * the scheduling request to the cpu owning the thread. Races
697 * are ok, the target will forward the message as necessary (the
698 * message may chase the thread around before it finally gets
699 * acted upon).
700 *
701 * (remember, wait structures use stable storage)
702 */
703 if ((w = td->td_wait) != NULL) {
96728c05 704 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
705 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
706 --w->wa_count;
707 td->td_wait = NULL;
d0e06f83 708 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 709 _lwkt_enqueue(td);
57c254db 710 if (td->td_preemptable) {
96728c05 711 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
712 } else if (_lwkt_wantresched(td, curthread)) {
713 need_resched();
714 }
f1d1c3fa 715 } else {
96728c05 716 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 717 }
96728c05 718 lwkt_reltoken(&w->wa_token);
f1d1c3fa 719 } else {
96728c05 720 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
721 }
722 } else {
723 /*
724 * If the wait structure is NULL and we own the thread, there
725 * is no race (since we are in a critical section). If we
726 * do not own the thread there might be a race but the
727 * target cpu will deal with it.
728 */
d0e06f83 729 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa 730 _lwkt_enqueue(td);
57c254db 731 if (td->td_preemptable) {
96728c05 732 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
733 } else if (_lwkt_wantresched(td, curthread)) {
734 need_resched();
735 }
f1d1c3fa 736 } else {
96728c05 737 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
738 }
739 }
8ad65e08 740 }
f1d1c3fa 741 crit_exit();
8ad65e08
MD
742}
743
d9eea1a5
MD
744/*
745 * Managed acquisition. This code assumes that the MP lock is held for
746 * the tdallq operation and that the thread has been descheduled from its
747 * original cpu. We also have to wait for the thread to be entirely switched
748 * out on its original cpu (this is usually fast enough that we never loop)
749 * since the LWKT system does not have to hold the MP lock while switching
750 * and the target may have released it before switching.
751 */
a2a5ad0d
MD
752void
753lwkt_acquire(thread_t td)
754{
755 struct globaldata *gd;
cb973d15 756 int ocpu;
a2a5ad0d
MD
757
758 gd = td->td_gd;
759 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
760 while (td->td_flags & TDF_RUNNING) /* XXX spin */
761 ;
a2a5ad0d 762 if (gd != mycpu) {
cb973d15 763 ocpu = td->td_cpu;
a2a5ad0d
MD
764 crit_enter();
765 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
766 gd = mycpu;
767 td->td_gd = gd;
768 td->td_cpu = gd->gd_cpuid;
769 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
770 crit_exit();
771 }
772}
773
8ad65e08 774/*
f1d1c3fa
MD
775 * Deschedule a thread.
776 *
777 * (non-blocking, self contained on a per cpu basis)
778 */
779void
780lwkt_deschedule_self(void)
781{
782 thread_t td = curthread;
783
784 crit_enter();
785 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
786 _lwkt_dequeue(td);
787 crit_exit();
788}
789
790/*
791 * Generic deschedule. Descheduling threads other then your own should be
792 * done only in carefully controlled circumstances. Descheduling is
793 * asynchronous.
794 *
795 * This function may block if the cpu has run out of messages.
8ad65e08
MD
796 */
797void
798lwkt_deschedule(thread_t td)
799{
f1d1c3fa
MD
800 crit_enter();
801 if (td == curthread) {
802 _lwkt_dequeue(td);
803 } else {
d0e06f83 804 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
805 _lwkt_dequeue(td);
806 } else {
96728c05 807 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
808 }
809 }
810 crit_exit();
811}
812
4b5f931b
MD
813/*
814 * Set the target thread's priority. This routine does not automatically
815 * switch to a higher priority thread, LWKT threads are not designed for
816 * continuous priority changes. Yield if you want to switch.
817 *
818 * We have to retain the critical section count which uses the high bits
26a0694b
MD
819 * of the td_pri field. The specified priority may also indicate zero or
820 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
821 */
822void
823lwkt_setpri(thread_t td, int pri)
824{
26a0694b 825 KKASSERT(pri >= 0);
57c254db 826 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
26a0694b
MD
827 crit_enter();
828 if (td->td_flags & TDF_RUNQ) {
829 _lwkt_dequeue(td);
830 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
831 _lwkt_enqueue(td);
832 } else {
833 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
834 }
835 crit_exit();
836}
837
838void
839lwkt_setpri_self(int pri)
840{
841 thread_t td = curthread;
842
4b5f931b
MD
843 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
844 crit_enter();
845 if (td->td_flags & TDF_RUNQ) {
846 _lwkt_dequeue(td);
847 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
848 _lwkt_enqueue(td);
849 } else {
850 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
851 }
852 crit_exit();
853}
854
855struct proc *
856lwkt_preempted_proc(void)
857{
73e4f7b9 858 thread_t td = curthread;
4b5f931b
MD
859 while (td->td_preempted)
860 td = td->td_preempted;
861 return(td->td_proc);
862}
863
864
f1d1c3fa
MD
865/*
866 * This function deschedules the current thread and blocks on the specified
867 * wait queue. We obtain ownership of the wait queue in order to block
868 * on it. A generation number is used to interlock the wait queue in case
869 * it gets signalled while we are blocked waiting on the token.
870 *
871 * Note: alternatively we could dequeue our thread and then message the
872 * target cpu owning the wait queue. YYY implement as sysctl.
873 *
874 * Note: wait queue signals normally ping-pong the cpu as an optimization.
875 */
96728c05
MD
876typedef struct lwkt_gettoken_req {
877 lwkt_token_t tok;
878 int cpu;
879} lwkt_gettoken_req;
880
f1d1c3fa 881void
ae8050a4 882lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
883{
884 thread_t td = curthread;
f1d1c3fa 885
f1d1c3fa 886 lwkt_gettoken(&w->wa_token);
ae8050a4 887 if (w->wa_gen == *gen) {
f1d1c3fa
MD
888 _lwkt_dequeue(td);
889 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
890 ++w->wa_count;
891 td->td_wait = w;
ae8050a4 892 td->td_wmesg = wmesg;
f1d1c3fa 893 lwkt_switch();
8ad65e08 894 }
ae8050a4
MD
895 /* token might be lost, doesn't matter for gen update */
896 *gen = w->wa_gen;
f1d1c3fa
MD
897 lwkt_reltoken(&w->wa_token);
898}
899
900/*
901 * Signal a wait queue. We gain ownership of the wait queue in order to
902 * signal it. Once a thread is removed from the wait queue we have to
903 * deal with the cpu owning the thread.
904 *
905 * Note: alternatively we could message the target cpu owning the wait
906 * queue. YYY implement as sysctl.
907 */
908void
909lwkt_signal(lwkt_wait_t w)
910{
911 thread_t td;
912 int count;
913
914 lwkt_gettoken(&w->wa_token);
915 ++w->wa_gen;
916 count = w->wa_count;
917 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
918 --count;
919 --w->wa_count;
920 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
921 td->td_wait = NULL;
ae8050a4 922 td->td_wmesg = NULL;
d0e06f83 923 if (td->td_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
924 _lwkt_enqueue(td);
925 } else {
96728c05 926 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
927 }
928 lwkt_regettoken(&w->wa_token);
929 }
930 lwkt_reltoken(&w->wa_token);
931}
932
933/*
96728c05 934 * Acquire ownership of a token
f1d1c3fa 935 *
96728c05 936 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
937 * section side effects, depending on its purpose. These side effects
938 * guarentee that you will maintain ownership of the token as long as you
939 * do not block. If you block you may lose access to the token (but you
940 * must still release it even if you lose your access to it).
941 *
96728c05 942 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 943 * a token, but do we really only need to disable IPIs ?
96728c05
MD
944 *
945 * YYY certain tokens could be made to act like mutexes when performance
946 * would be better (e.g. t_cpu == -1). This is not yet implemented.
947 *
a2a5ad0d
MD
948 * YYY the tokens replace 4.x's simplelocks for the most part, but this
949 * means that 4.x does not expect a switch so for now we cannot switch
950 * when waiting for an IPI to be returned.
951 *
952 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
953 * it and then block. The IPI causes the token to be given away to the
954 * requesting cpu, unless it has already changed hands. Since only the
955 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 956 * This needs serious optimization.
f1d1c3fa 957 */
57c254db
MD
958
959#ifdef SMP
960
96728c05
MD
961static
962void
963lwkt_gettoken_remote(void *arg)
964{
965 lwkt_gettoken_req *req = arg;
966 if (req->tok->t_cpu == mycpu->gd_cpuid) {
a2a5ad0d
MD
967 if (token_debug)
968 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
96728c05 969 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
970 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
971 /* else set reqcpu to point to current cpu for release */
96728c05
MD
972 }
973}
974
57c254db
MD
975#endif
976
8a8d5d85 977int
f1d1c3fa
MD
978lwkt_gettoken(lwkt_token_t tok)
979{
980 /*
981 * Prevent preemption so the token can't be taken away from us once
982 * we gain ownership of it. Use a synchronous request which might
983 * block. The request will be forwarded as necessary playing catchup
984 * to the token.
985 */
96728c05 986
f1d1c3fa 987 crit_enter();
57c254db 988#ifdef INVARIANTS
a2a5ad0d
MD
989 if (curthread->td_pri > 2000) {
990 curthread->td_pri = 1000;
991 panic("too HIGH!");
57c254db
MD
992 }
993#endif
96728c05 994#ifdef SMP
d0e06f83 995 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
996 struct lwkt_gettoken_req req;
997 int seq;
96728c05
MD
998 int dcpu;
999
1000 req.cpu = mycpu->gd_cpuid;
1001 req.tok = tok;
1002 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d
MD
1003 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1004 if (token_debug)
1005 printf("REQT%d ", dcpu);
96728c05
MD
1006 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1007 lwkt_wait_ipiq(dcpu, seq);
a2a5ad0d
MD
1008 if (token_debug)
1009 printf("REQR%d ", tok->t_cpu);
f1d1c3fa
MD
1010 }
1011#endif
1012 /*
1013 * leave us in a critical section on return. This will be undone
8a8d5d85 1014 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1015 */
8a8d5d85 1016 return(++tok->t_gen);
f1d1c3fa
MD
1017}
1018
96728c05
MD
1019/*
1020 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1021 * failure.
1022 */
1023int
1024lwkt_trytoken(lwkt_token_t tok)
1025{
1026 crit_enter();
1027#ifdef SMP
1028 if (tok->t_cpu != mycpu->gd_cpuid) {
1029 return(0);
1030 }
1031#endif
1032 /* leave us in the critical section */
1033 ++tok->t_gen;
1034 return(1);
1035}
1036
f1d1c3fa
MD
1037/*
1038 * Release your ownership of a token. Releases must occur in reverse
1039 * order to aquisitions, eventually so priorities can be unwound properly
1040 * like SPLs. At the moment the actual implemention doesn't care.
1041 *
1042 * We can safely hand a token that we own to another cpu without notifying
1043 * it, but once we do we can't get it back without requesting it (unless
1044 * the other cpu hands it back to us before we check).
1045 *
1046 * We might have lost the token, so check that.
1047 */
1048void
1049lwkt_reltoken(lwkt_token_t tok)
1050{
d0e06f83 1051 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1052 tok->t_cpu = tok->t_reqcpu;
1053 }
1054 crit_exit();
1055}
1056
1057/*
8a8d5d85
MD
1058 * Reacquire a token that might have been lost and compare and update the
1059 * generation number. 0 is returned if the generation has not changed
1060 * (nobody else obtained the token while we were blocked, on this cpu or
1061 * any other cpu).
1062 *
1063 * This function returns with the token re-held whether the generation
1064 * number changed or not.
1065 */
1066int
1067lwkt_gentoken(lwkt_token_t tok, int *gen)
1068{
1069 if (lwkt_regettoken(tok) == *gen) {
1070 return(0);
1071 } else {
1072 *gen = tok->t_gen;
1073 return(-1);
1074 }
1075}
1076
1077
1078/*
96728c05 1079 * Re-acquire a token that might have been lost. Returns the generation
8a8d5d85 1080 * number of the token.
f1d1c3fa
MD
1081 */
1082int
1083lwkt_regettoken(lwkt_token_t tok)
1084{
96728c05 1085 /* assert we are in a critical section */
d0e06f83 1086 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1087#ifdef SMP
d0e06f83 1088 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1089 struct lwkt_gettoken_req req;
1090 int seq;
96728c05 1091 int dcpu;
57c254db 1092
96728c05
MD
1093 req.cpu = mycpu->gd_cpuid;
1094 req.tok = tok;
1095 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1096 KKASSERT(dcpu >= 0 && dcpu < ncpus);
cb973d15
MD
1097 if (token_debug)
1098 printf("REQT%d ", dcpu);
96728c05
MD
1099 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1100 lwkt_wait_ipiq(dcpu, seq);
cb973d15
MD
1101 if (token_debug)
1102 printf("REQR%d ", tok->t_cpu);
f1d1c3fa 1103 }
f1d1c3fa 1104#endif
96728c05
MD
1105 ++tok->t_gen;
1106 }
8a8d5d85 1107 return(tok->t_gen);
8ad65e08
MD
1108}
1109
72740893
MD
1110void
1111lwkt_inittoken(lwkt_token_t tok)
1112{
1113 /*
1114 * Zero structure and set cpu owner and reqcpu to cpu 0.
1115 */
1116 bzero(tok, sizeof(*tok));
1117}
1118
99df837e
MD
1119/*
1120 * Create a kernel process/thread/whatever. It shares it's address space
1121 * with proc0 - ie: kernel only.
1122 *
1123 * XXX should be renamed to lwkt_create()
8a8d5d85
MD
1124 *
1125 * The thread will be entered with the MP lock held.
99df837e
MD
1126 */
1127int
1128lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1129 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1130 const char *fmt, ...)
99df837e 1131{
73e4f7b9 1132 thread_t td;
99df837e
MD
1133 va_list ap;
1134
a2a5ad0d
MD
1135 td = lwkt_alloc_thread(template);
1136 if (tdp)
1137 *tdp = td;
99df837e 1138 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1139 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1140#ifdef SMP
1141 td->td_mpcount = 1;
1142#endif
99df837e
MD
1143
1144 /*
1145 * Set up arg0 for 'ps' etc
1146 */
1147 va_start(ap, fmt);
1148 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1149 va_end(ap);
1150
1151 /*
1152 * Schedule the thread to run
1153 */
ef0fdad1
MD
1154 if ((td->td_flags & TDF_STOPREQ) == 0)
1155 lwkt_schedule(td);
1156 else
1157 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1158 return 0;
1159}
1160
1161/*
1162 * Destroy an LWKT thread. Warning! This function is not called when
1163 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1164 * uses a different reaping mechanism.
1165 */
1166void
1167lwkt_exit(void)
1168{
1169 thread_t td = curthread;
1170
1171 if (td->td_flags & TDF_VERBOSE)
1172 printf("kthread %p %s has exited\n", td, td->td_comm);
1173 crit_enter();
1174 lwkt_deschedule_self();
1175 ++mycpu->gd_tdfreecount;
1176 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1177 cpu_thread_exit();
1178}
1179
1180/*
1181 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1182 * with proc0 - ie: kernel only. 5.x compatible.
99df837e
MD
1183 */
1184int
1185kthread_create(void (*func)(void *), void *arg,
1186 struct thread **tdp, const char *fmt, ...)
1187{
73e4f7b9 1188 thread_t td;
99df837e
MD
1189 va_list ap;
1190
a2a5ad0d
MD
1191 td = lwkt_alloc_thread(NULL);
1192 if (tdp)
1193 *tdp = td;
99df837e
MD
1194 cpu_set_thread_handler(td, kthread_exit, func, arg);
1195 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1196#ifdef SMP
1197 td->td_mpcount = 1;
1198#endif
99df837e
MD
1199
1200 /*
1201 * Set up arg0 for 'ps' etc
1202 */
1203 va_start(ap, fmt);
1204 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1205 va_end(ap);
1206
1207 /*
1208 * Schedule the thread to run
1209 */
1210 lwkt_schedule(td);
1211 return 0;
1212}
1213
26a0694b
MD
1214void
1215crit_panic(void)
1216{
73e4f7b9 1217 thread_t td = curthread;
26a0694b
MD
1218 int lpri = td->td_pri;
1219
1220 td->td_pri = 0;
1221 panic("td_pri is/would-go negative! %p %d", td, lpri);
1222}
1223
99df837e
MD
1224/*
1225 * Destroy an LWKT thread. Warning! This function is not called when
1226 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1227 * uses a different reaping mechanism.
1228 *
1229 * XXX duplicates lwkt_exit()
1230 */
1231void
1232kthread_exit(void)
1233{
1234 lwkt_exit();
1235}
1236
96728c05
MD
1237#ifdef SMP
1238
1239/*
1240 * Send a function execution request to another cpu. The request is queued
1241 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1242 * possible target cpu. The FIFO can be written.
1243 *
1244 * YYY If the FIFO fills up we have to enable interrupts and process the
1245 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1246 * Create a CPU_*() function to do this!
1247 *
1248 * Must be called from a critical section.
1249 */
1250int
1251lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1252{
1253 lwkt_ipiq_t ip;
1254 int windex;
a2a5ad0d 1255 struct globaldata *gd = mycpu;
96728c05 1256
a2a5ad0d 1257 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1258 func(arg);
1259 return(0);
1260 }
cb973d15 1261 crit_enter();
a2a5ad0d
MD
1262 ++gd->gd_intr_nesting_level;
1263#ifdef INVARIANTS
1264 if (gd->gd_intr_nesting_level > 20)
1265 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1266#endif
96728c05
MD
1267 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1268 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1269 ++ipiq_count;
a2a5ad0d 1270 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1271
1272 /*
1273 * We always drain before the FIFO becomes full so it should never
1274 * become full. We need to leave enough entries to deal with
1275 * reentrancy.
1276 */
1277 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1278 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1279 ip->ip_func[windex] = func;
1280 ip->ip_arg[windex] = arg;
1281 /* YYY memory barrier */
1282 ++ip->ip_windex;
96728c05
MD
1283 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1284 unsigned int eflags = read_eflags();
1285 cpu_enable_intr();
1286 ++ipiq_fifofull;
cb973d15 1287 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1288 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1289 lwkt_process_ipiq();
1290 }
1291 write_eflags(eflags);
1292 }
a2a5ad0d 1293 --gd->gd_intr_nesting_level;
96728c05 1294 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1295 crit_exit();
96728c05
MD
1296 return(ip->ip_windex);
1297}
1298
cb973d15
MD
1299/*
1300 * Send a message to several target cpus. Typically used for scheduling.
1301 */
1302void
1303lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1304{
1305 int cpuid;
1306
1307 while (mask) {
1308 cpuid = bsfl(mask);
1309 lwkt_send_ipiq(cpuid, func, arg);
1310 mask &= ~(1 << cpuid);
1311 }
1312}
1313
96728c05
MD
1314/*
1315 * Wait for the remote cpu to finish processing a function.
1316 *
1317 * YYY we have to enable interrupts and process the IPIQ while waiting
1318 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1319 * function to do this! YYY we really should 'block' here.
1320 *
1321 * Must be called from a critical section. Thsi routine may be called
1322 * from an interrupt (for example, if an interrupt wakes a foreign thread
1323 * up).
1324 */
1325void
1326lwkt_wait_ipiq(int dcpu, int seq)
1327{
1328 lwkt_ipiq_t ip;
a2a5ad0d 1329 int maxc = 100000000;
96728c05
MD
1330
1331 if (dcpu != mycpu->gd_cpuid) {
1332 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1333 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1334 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1335 unsigned int eflags = read_eflags();
1336 cpu_enable_intr();
cb973d15 1337 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1338 lwkt_process_ipiq();
a2a5ad0d 1339 if (--maxc == 0)
cb973d15 1340 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1341 if (maxc < -1000000)
1342 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1343 }
1344 write_eflags(eflags);
1345 }
1346 }
1347}
1348
1349/*
1350 * Called from IPI interrupt (like a fast interrupt), which has placed
1351 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1352 * May also be called from doreti or splz, or be reentrantly called
1353 * indirectly through the ip_func[] we run.
96728c05
MD
1354 */
1355void
1356lwkt_process_ipiq(void)
1357{
1358 int n;
1359 int cpuid = mycpu->gd_cpuid;
1360
1361 for (n = 0; n < ncpus; ++n) {
1362 lwkt_ipiq_t ip;
1363 int ri;
1364
1365 if (n == cpuid)
1366 continue;
1367 ip = globaldata_find(n)->gd_ipiq;
1368 if (ip == NULL)
1369 continue;
1370 ip = &ip[cpuid];
cb973d15
MD
1371
1372 /*
1373 * Note: xindex is only updated after we are sure the function has
1374 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1375 * function may send an IPI which may block/drain.
1376 */
96728c05
MD
1377 while (ip->ip_rindex != ip->ip_windex) {
1378 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1379 ++ip->ip_rindex;
cb973d15
MD
1380 ip->ip_func[ri](ip->ip_arg[ri]);
1381 /* YYY memory barrier */
1382 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1383 }
1384 }
1385}
1386
1387#else
1388
1389int
1390lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1391{
1392 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1393 return(0); /* NOT REACHED */
1394}
1395
1396void
1397lwkt_wait_ipiq(int dcpu, int seq)
1398{
1399 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1400}
1401
1402#endif