Better documentation of the MP lock state for new threads.
[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 *
365fa13f 31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.39 2003/11/03 00:39:07 dillon Exp $
8ad65e08
MD
32 */
33
34#include <sys/param.h>
35#include <sys/systm.h>
36#include <sys/kernel.h>
37#include <sys/proc.h>
38#include <sys/rtprio.h>
39#include <sys/queue.h>
f1d1c3fa 40#include <sys/thread2.h>
7d0bac62 41#include <sys/sysctl.h>
99df837e 42#include <sys/kthread.h>
f1d1c3fa 43#include <machine/cpu.h>
99df837e 44#include <sys/lock.h>
f1d1c3fa 45
7d0bac62
MD
46#include <vm/vm.h>
47#include <vm/vm_param.h>
48#include <vm/vm_kern.h>
49#include <vm/vm_object.h>
50#include <vm/vm_page.h>
51#include <vm/vm_map.h>
52#include <vm/vm_pager.h>
53#include <vm/vm_extern.h>
54#include <vm/vm_zone.h>
55
99df837e 56#include <machine/stdarg.h>
57c254db 57#include <machine/ipl.h>
96728c05
MD
58#ifdef SMP
59#include <machine/smp.h>
60#endif
99df837e 61
7d0bac62 62static int untimely_switch = 0;
4b5f931b 63SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
57c254db
MD
64#ifdef INVARIANTS
65static int token_debug = 0;
66SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
67#endif
4b5f931b
MD
68static quad_t switch_count = 0;
69SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
70static quad_t preempt_hit = 0;
71SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
72static quad_t preempt_miss = 0;
73SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b
MD
74static quad_t preempt_weird = 0;
75SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05
MD
76static quad_t ipiq_count = 0;
77SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
78static quad_t ipiq_fifofull = 0;
79SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 80
4b5f931b
MD
81/*
82 * These helper procedures handle the runq, they can only be called from
83 * within a critical section.
84 */
f1d1c3fa
MD
85static __inline
86void
87_lwkt_dequeue(thread_t td)
88{
89 if (td->td_flags & TDF_RUNQ) {
4b5f931b
MD
90 int nq = td->td_pri & TDPRI_MASK;
91 struct globaldata *gd = mycpu;
92
f1d1c3fa 93 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
94 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
95 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
96 }
97}
98
99static __inline
100void
101_lwkt_enqueue(thread_t td)
102{
103 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b
MD
104 int nq = td->td_pri & TDPRI_MASK;
105 struct globaldata *gd = mycpu;
106
f1d1c3fa 107 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
108 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
109 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
110 }
111}
8ad65e08 112
57c254db
MD
113static __inline
114int
115_lwkt_wantresched(thread_t ntd, thread_t cur)
116{
117 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
118}
119
8ad65e08
MD
120/*
121 * LWKTs operate on a per-cpu basis
122 *
73e4f7b9 123 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
124 */
125void
126lwkt_gdinit(struct globaldata *gd)
127{
4b5f931b
MD
128 int i;
129
130 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
131 TAILQ_INIT(&gd->gd_tdrunq[i]);
132 gd->gd_runqmask = 0;
73e4f7b9 133 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
134}
135
7d0bac62
MD
136/*
137 * Initialize a thread wait structure prior to first use.
138 *
139 * NOTE! called from low level boot code, we cannot do anything fancy!
140 */
141void
142lwkt_init_wait(lwkt_wait_t w)
143{
144 TAILQ_INIT(&w->wa_waitq);
145}
146
147/*
148 * Create a new thread. The thread must be associated with a process context
149 * or LWKT start address before it can be scheduled.
0cfcada1
MD
150 *
151 * If you intend to create a thread without a process context this function
152 * does everything except load the startup and switcher function.
7d0bac62
MD
153 */
154thread_t
ef0fdad1 155lwkt_alloc_thread(struct thread *td)
7d0bac62 156{
99df837e 157 void *stack;
ef0fdad1 158 int flags = 0;
7d0bac62 159
ef0fdad1 160 if (td == NULL) {
26a0694b 161 crit_enter();
ef0fdad1
MD
162 if (mycpu->gd_tdfreecount > 0) {
163 --mycpu->gd_tdfreecount;
164 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 165 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
166 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
167 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
168 crit_exit();
169 stack = td->td_kstack;
170 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
171 } else {
172 crit_exit();
173 td = zalloc(thread_zone);
174 td->td_kstack = NULL;
175 flags |= TDF_ALLOCATED_THREAD;
176 }
177 }
178 if ((stack = td->td_kstack) == NULL) {
99df837e 179 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 180 flags |= TDF_ALLOCATED_STACK;
99df837e 181 }
26a0694b 182 lwkt_init_thread(td, stack, flags, mycpu);
99df837e 183 return(td);
7d0bac62
MD
184}
185
186/*
187 * Initialize a preexisting thread structure. This function is used by
188 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
189 *
190 * NOTE! called from low level boot code, we cannot do anything fancy!
a72187e9 191 * Only the low level boot code will call this function with gd != mycpu.
f8c3996b
MD
192 *
193 * All threads start out in a critical section at a priority of
194 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
195 * appropriate.
7d0bac62
MD
196 */
197void
26a0694b 198lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 199{
99df837e
MD
200 bzero(td, sizeof(struct thread));
201 td->td_kstack = stack;
202 td->td_flags |= flags;
26a0694b 203 td->td_gd = gd;
f8c3996b 204 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
ece04fd0 205 lwkt_init_port(&td->td_msgport, td);
99df837e 206 pmap_init_thread(td);
73e4f7b9 207 crit_enter();
a72187e9 208 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
73e4f7b9
MD
209 crit_exit();
210}
211
212void
213lwkt_set_comm(thread_t td, const char *ctl, ...)
214{
215 va_list va;
216
217 va_start(va, ctl);
218 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
219 va_end(va);
7d0bac62
MD
220}
221
99df837e 222void
73e4f7b9 223lwkt_hold(thread_t td)
99df837e 224{
73e4f7b9
MD
225 ++td->td_refs;
226}
227
228void
229lwkt_rele(thread_t td)
230{
231 KKASSERT(td->td_refs > 0);
232 --td->td_refs;
233}
234
235void
236lwkt_wait_free(thread_t td)
237{
238 while (td->td_refs)
377d4740 239 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
240}
241
242void
243lwkt_free_thread(thread_t td)
244{
245 struct globaldata *gd = mycpu;
246
d9eea1a5 247 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
248 ("lwkt_free_thread: did not exit! %p", td));
249
250 crit_enter();
73e4f7b9
MD
251 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
252 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
253 (td->td_flags & TDF_ALLOCATED_THREAD)
254 ) {
73e4f7b9
MD
255 ++gd->gd_tdfreecount;
256 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
257 crit_exit();
258 } else {
259 crit_exit();
260 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
261 kmem_free(kernel_map,
262 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 263 /* gd invalid */
99df837e
MD
264 td->td_kstack = NULL;
265 }
266 if (td->td_flags & TDF_ALLOCATED_THREAD)
267 zfree(thread_zone, td);
268 }
269}
270
271
8ad65e08
MD
272/*
273 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
274 * switch to the idlethread. Switching must occur within a critical
275 * section to avoid races with the scheduling queue.
276 *
277 * We always have full control over our cpu's run queue. Other cpus
278 * that wish to manipulate our queue must use the cpu_*msg() calls to
279 * talk to our cpu, so a critical section is all that is needed and
280 * the result is very, very fast thread switching.
281 *
96728c05
MD
282 * The LWKT scheduler uses a fixed priority model and round-robins at
283 * each priority level. User process scheduling is a totally
284 * different beast and LWKT priorities should not be confused with
285 * user process priorities.
f1d1c3fa 286 *
96728c05
MD
287 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
288 * cleans it up. Note that the td_switch() function cannot do anything that
289 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
290 * the target thread (not the current thread). It's nice to have a scheduler
291 * that does not need the MP lock to work because it allows us to do some
292 * really cool high-performance MP lock optimizations.
8ad65e08 293 */
96728c05 294
8ad65e08
MD
295void
296lwkt_switch(void)
297{
4b5f931b 298 struct globaldata *gd;
f1d1c3fa 299 thread_t td = curthread;
8ad65e08 300 thread_t ntd;
8a8d5d85
MD
301#ifdef SMP
302 int mpheld;
303#endif
8ad65e08 304
46a3f46d
MD
305 /*
306 * Switching from within a 'fast' (non thread switched) interrupt is
307 * illegal.
308 */
309 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 310 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 311 }
ef0fdad1 312
cb973d15
MD
313 /*
314 * Passive release (used to transition from user to kernel mode
315 * when we block or switch rather then when we enter the kernel).
316 * This function is NOT called if we are switching into a preemption
317 * or returning from a preemption. Typically this causes us to lose
318 * our P_CURPROC designation (if we have one) and become a true LWKT
319 * thread, and may also hand P_CURPROC to another process and schedule
320 * its thread.
321 */
322 if (td->td_release)
323 td->td_release(td);
324
f1d1c3fa 325 crit_enter();
4b5f931b 326 ++switch_count;
8a8d5d85
MD
327
328#ifdef SMP
329 /*
330 * td_mpcount cannot be used to determine if we currently hold the
331 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
332 * to get the lock, and switch out if it can't. Our ownership of
333 * the actual lock will remain stable while we are in a critical section
334 * (but, of course, another cpu may own or release the lock so the
335 * actual value of mp_lock is not stable).
8a8d5d85
MD
336 */
337 mpheld = MP_LOCK_HELD();
338#endif
99df837e
MD
339 if ((ntd = td->td_preempted) != NULL) {
340 /*
341 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
342 * thread. This occurs transparently, whether the preempted thread
343 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
344 * itself).
345 *
346 * We have to setup the MP lock for the original thread after backing
347 * out the adjustment that was made to curthread when the original
348 * was preempted.
99df837e 349 */
26a0694b 350 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 351#ifdef SMP
96728c05
MD
352 if (ntd->td_mpcount && mpheld == 0) {
353 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
354 td, ntd, td->td_mpcount, ntd->td_mpcount);
355 }
8a8d5d85
MD
356 if (ntd->td_mpcount) {
357 td->td_mpcount -= ntd->td_mpcount;
358 KKASSERT(td->td_mpcount >= 0);
359 }
360#endif
26a0694b 361 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 362 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 363 } else {
4b5f931b
MD
364 /*
365 * Priority queue / round-robin at each priority. Note that user
366 * processes run at a fixed, low priority and the user process
367 * scheduler deals with interactions between user processes
368 * by scheduling and descheduling them from the LWKT queue as
369 * necessary.
8a8d5d85
MD
370 *
371 * We have to adjust the MP lock for the target thread. If we
372 * need the MP lock and cannot obtain it we try to locate a
373 * thread that does not need the MP lock.
4b5f931b
MD
374 */
375 gd = mycpu;
4b5f931b
MD
376again:
377 if (gd->gd_runqmask) {
378 int nq = bsrl(gd->gd_runqmask);
379 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
380 gd->gd_runqmask &= ~(1 << nq);
381 goto again;
382 }
8a8d5d85
MD
383#ifdef SMP
384 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
385 /*
96728c05
MD
386 * Target needs MP lock and we couldn't get it, try
387 * to locate a thread which does not need the MP lock
3c23a41a 388 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
389 */
390 u_int32_t rqmask = gd->gd_runqmask;
391 while (rqmask) {
392 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
393 if (ntd->td_mpcount == 0)
394 break;
395 }
396 if (ntd)
397 break;
398 rqmask &= ~(1 << nq);
399 nq = bsrl(rqmask);
400 }
401 if (ntd == NULL) {
a2a5ad0d
MD
402 ntd = &gd->gd_idlethread;
403 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
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 {
409 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
410 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
411 }
412#else
4b5f931b
MD
413 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
414 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 415#endif
4b5f931b 416 } else {
3c23a41a
MD
417 /*
418 * Nothing to run but we may still need the BGL to deal with
419 * pending interrupts, spin in idle if so.
420 */
a2a5ad0d 421 ntd = &gd->gd_idlethread;
235957ed 422 if (gd->gd_reqflags)
3c23a41a 423 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 424 }
f1d1c3fa 425 }
26a0694b
MD
426 KASSERT(ntd->td_pri >= TDPRI_CRIT,
427 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
428
429 /*
430 * Do the actual switch. If the new target does not need the MP lock
431 * and we are holding it, release the MP lock. If the new target requires
432 * the MP lock we have already acquired it for the target.
433 */
434#ifdef SMP
435 if (ntd->td_mpcount == 0 ) {
436 if (MP_LOCK_HELD())
437 cpu_rel_mplock();
438 } else {
439 ASSERT_MP_LOCK_HELD();
440 }
441#endif
8a8d5d85 442 if (td != ntd) {
f1d1c3fa 443 td->td_switch(ntd);
8a8d5d85 444 }
96728c05 445
f1d1c3fa 446 crit_exit();
8ad65e08
MD
447}
448
cb973d15
MD
449/*
450 * Switch if another thread has a higher priority. Do not switch to other
451 * threads at the same priority.
452 */
453void
454lwkt_maybe_switch()
455{
456 struct globaldata *gd = mycpu;
457 struct thread *td = gd->gd_curthread;
458
459 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
460 lwkt_switch();
461 }
462}
463
b68b7282 464/*
96728c05
MD
465 * Request that the target thread preempt the current thread. Preemption
466 * only works under a specific set of conditions:
b68b7282 467 *
96728c05
MD
468 * - We are not preempting ourselves
469 * - The target thread is owned by the current cpu
470 * - We are not currently being preempted
471 * - The target is not currently being preempted
472 * - We are able to satisfy the target's MP lock requirements (if any).
473 *
474 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
475 * this is called via lwkt_schedule() through the td_preemptable callback.
476 * critpri is the managed critical priority that we should ignore in order
477 * to determine whether preemption is possible (aka usually just the crit
478 * priority of lwkt_schedule() itself).
b68b7282 479 *
26a0694b
MD
480 * XXX at the moment we run the target thread in a critical section during
481 * the preemption in order to prevent the target from taking interrupts
482 * that *WE* can't. Preemption is strictly limited to interrupt threads
483 * and interrupt-like threads, outside of a critical section, and the
484 * preempted source thread will be resumed the instant the target blocks
485 * whether or not the source is scheduled (i.e. preemption is supposed to
486 * be as transparent as possible).
4b5f931b 487 *
8a8d5d85
MD
488 * The target thread inherits our MP count (added to its own) for the
489 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
490 * MP lock during the preemption. Therefore, any preempting targets must be
491 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
492 * out of sync with the physical mp_lock, but we do not have to preserve
493 * the original ownership of the lock if it was out of synch (that is, we
494 * can leave it synchronized on return).
b68b7282
MD
495 */
496void
96728c05 497lwkt_preempt(thread_t ntd, int critpri)
b68b7282 498{
46a3f46d
MD
499 struct globaldata *gd = mycpu;
500 thread_t td = gd->gd_curthread;
8a8d5d85
MD
501#ifdef SMP
502 int mpheld;
57c254db 503 int savecnt;
8a8d5d85 504#endif
b68b7282 505
26a0694b 506 /*
96728c05
MD
507 * The caller has put us in a critical section. We can only preempt
508 * if the caller of the caller was not in a critical section (basically
57c254db
MD
509 * a local interrupt), as determined by the 'critpri' parameter. If
510 * we are unable to preempt
96728c05
MD
511 *
512 * YYY The target thread must be in a critical section (else it must
513 * inherit our critical section? I dunno yet).
26a0694b
MD
514 */
515 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 516
cb973d15 517 need_resched();
57c254db
MD
518 if (!_lwkt_wantresched(ntd, td)) {
519 ++preempt_miss;
520 return;
521 }
96728c05
MD
522 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
523 ++preempt_miss;
524 return;
525 }
526#ifdef SMP
46a3f46d 527 if (ntd->td_gd != gd) {
96728c05
MD
528 ++preempt_miss;
529 return;
530 }
531#endif
26a0694b
MD
532 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
533 ++preempt_weird;
534 return;
535 }
536 if (ntd->td_preempted) {
4b5f931b 537 ++preempt_hit;
26a0694b 538 return;
b68b7282 539 }
8a8d5d85 540#ifdef SMP
a2a5ad0d
MD
541 /*
542 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
543 * to or from the MP lock. In this case td_mpcount will be pre-disposed
544 * (non-zero) but not actually synchronized with the actual state of the
545 * lock. We can use it to imply an MP lock requirement for the
546 * preemption but we cannot use it to test whether we hold the MP lock
547 * or not.
a2a5ad0d 548 */
96728c05 549 savecnt = td->td_mpcount;
71ef2f5c 550 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
551 ntd->td_mpcount += td->td_mpcount;
552 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
553 ntd->td_mpcount -= td->td_mpcount;
554 ++preempt_miss;
555 return;
556 }
557#endif
26a0694b
MD
558
559 ++preempt_hit;
560 ntd->td_preempted = td;
561 td->td_flags |= TDF_PREEMPT_LOCK;
562 td->td_switch(ntd);
563 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
564#ifdef SMP
565 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
566 mpheld = MP_LOCK_HELD();
567 if (mpheld && td->td_mpcount == 0)
96728c05 568 cpu_rel_mplock();
71ef2f5c 569 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
570 panic("lwkt_preempt(): MP lock was not held through");
571#endif
26a0694b
MD
572 ntd->td_preempted = NULL;
573 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
574}
575
f1d1c3fa
MD
576/*
577 * Yield our thread while higher priority threads are pending. This is
578 * typically called when we leave a critical section but it can be safely
579 * called while we are in a critical section.
580 *
581 * This function will not generally yield to equal priority threads but it
582 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 583 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
584 * lwkt_yield_quick().
585 *
235957ed 586 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
587 * came along but was blocked and made pending.
588 *
f1d1c3fa
MD
589 * (self contained on a per cpu basis)
590 */
591void
592lwkt_yield_quick(void)
593{
7966cb69
MD
594 globaldata_t gd = mycpu;
595 thread_t td = gd->gd_curthread;
ef0fdad1 596
a2a5ad0d 597 /*
235957ed 598 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
599 * it with a non-zero cpl then we might not wind up calling splz after
600 * a task switch when the critical section is exited even though the
46a3f46d 601 * new task could accept the interrupt.
a2a5ad0d
MD
602 *
603 * XXX from crit_exit() only called after last crit section is released.
604 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
605 *
606 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
607 * except for this special case, we MUST call splz() here to handle any
608 * pending ints, particularly after we switch, or we might accidently
609 * halt the cpu with interrupts pending.
a2a5ad0d 610 */
46a3f46d 611 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 612 splz();
f1d1c3fa
MD
613
614 /*
615 * YYY enabling will cause wakeup() to task-switch, which really
616 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
617 * preemption and MP without actually doing preemption or MP, because a
618 * lot of code assumes that wakeup() does not block.
f1d1c3fa 619 */
46a3f46d
MD
620 if (untimely_switch && td->td_nest_count == 0 &&
621 gd->gd_intr_nesting_level == 0
622 ) {
f1d1c3fa
MD
623 crit_enter();
624 /*
625 * YYY temporary hacks until we disassociate the userland scheduler
626 * from the LWKT scheduler.
627 */
628 if (td->td_flags & TDF_RUNQ) {
629 lwkt_switch(); /* will not reenter yield function */
630 } else {
631 lwkt_schedule_self(); /* make sure we are scheduled */
632 lwkt_switch(); /* will not reenter yield function */
633 lwkt_deschedule_self(); /* make sure we are descheduled */
634 }
7966cb69 635 crit_exit_noyield(td);
f1d1c3fa 636 }
f1d1c3fa
MD
637}
638
8ad65e08 639/*
f1d1c3fa 640 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 641 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
642 * the crit_exit() call in lwkt_switch().
643 *
644 * (self contained on a per cpu basis)
8ad65e08
MD
645 */
646void
f1d1c3fa 647lwkt_yield(void)
8ad65e08 648{
f1d1c3fa
MD
649 lwkt_schedule_self();
650 lwkt_switch();
651}
652
653/*
654 * Schedule a thread to run. As the current thread we can always safely
655 * schedule ourselves, and a shortcut procedure is provided for that
656 * function.
657 *
658 * (non-blocking, self contained on a per cpu basis)
659 */
660void
661lwkt_schedule_self(void)
662{
663 thread_t td = curthread;
664
665 crit_enter();
666 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 667 _lwkt_enqueue(td);
26a0694b
MD
668 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
669 panic("SCHED SELF PANIC");
f1d1c3fa 670 crit_exit();
8ad65e08 671}
8ad65e08
MD
672
673/*
f1d1c3fa
MD
674 * Generic schedule. Possibly schedule threads belonging to other cpus and
675 * deal with threads that might be blocked on a wait queue.
676 *
96728c05 677 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
678 * Load balancing can be accomplished by requesting other sorts of actions
679 * for the thread in question.
8ad65e08
MD
680 */
681void
682lwkt_schedule(thread_t td)
683{
96728c05 684#ifdef INVARIANTS
26a0694b
MD
685 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
686 && td->td_proc->p_stat == SSLEEP
687 ) {
688 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
689 curthread,
690 curthread->td_proc ? curthread->td_proc->p_pid : -1,
691 curthread->td_proc ? curthread->td_proc->p_stat : -1,
692 td,
693 td->td_proc ? curthread->td_proc->p_pid : -1,
694 td->td_proc ? curthread->td_proc->p_stat : -1
695 );
696 panic("SCHED PANIC");
697 }
96728c05 698#endif
f1d1c3fa
MD
699 crit_enter();
700 if (td == curthread) {
701 _lwkt_enqueue(td);
702 } else {
703 lwkt_wait_t w;
704
705 /*
706 * If the thread is on a wait list we have to send our scheduling
707 * request to the owner of the wait structure. Otherwise we send
708 * the scheduling request to the cpu owning the thread. Races
709 * are ok, the target will forward the message as necessary (the
710 * message may chase the thread around before it finally gets
711 * acted upon).
712 *
713 * (remember, wait structures use stable storage)
714 */
715 if ((w = td->td_wait) != NULL) {
96728c05 716 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
717 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
718 --w->wa_count;
719 td->td_wait = NULL;
a72187e9 720 if (td->td_gd == mycpu) {
f1d1c3fa 721 _lwkt_enqueue(td);
57c254db 722 if (td->td_preemptable) {
96728c05 723 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
724 } else if (_lwkt_wantresched(td, curthread)) {
725 need_resched();
726 }
f1d1c3fa 727 } else {
a72187e9 728 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 729 }
96728c05 730 lwkt_reltoken(&w->wa_token);
f1d1c3fa 731 } else {
96728c05 732 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
733 }
734 } else {
735 /*
736 * If the wait structure is NULL and we own the thread, there
737 * is no race (since we are in a critical section). If we
738 * do not own the thread there might be a race but the
739 * target cpu will deal with it.
740 */
a72187e9 741 if (td->td_gd == mycpu) {
f1d1c3fa 742 _lwkt_enqueue(td);
57c254db 743 if (td->td_preemptable) {
96728c05 744 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
745 } else if (_lwkt_wantresched(td, curthread)) {
746 need_resched();
747 }
f1d1c3fa 748 } else {
a72187e9 749 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
750 }
751 }
8ad65e08 752 }
f1d1c3fa 753 crit_exit();
8ad65e08
MD
754}
755
d9eea1a5
MD
756/*
757 * Managed acquisition. This code assumes that the MP lock is held for
758 * the tdallq operation and that the thread has been descheduled from its
759 * original cpu. We also have to wait for the thread to be entirely switched
760 * out on its original cpu (this is usually fast enough that we never loop)
761 * since the LWKT system does not have to hold the MP lock while switching
762 * and the target may have released it before switching.
763 */
a2a5ad0d
MD
764void
765lwkt_acquire(thread_t td)
766{
767 struct globaldata *gd;
768
769 gd = td->td_gd;
770 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
771 while (td->td_flags & TDF_RUNNING) /* XXX spin */
772 ;
a2a5ad0d
MD
773 if (gd != mycpu) {
774 crit_enter();
775 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
776 gd = mycpu;
777 td->td_gd = gd;
a2a5ad0d
MD
778 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
779 crit_exit();
780 }
781}
782
8ad65e08 783/*
f1d1c3fa
MD
784 * Deschedule a thread.
785 *
786 * (non-blocking, self contained on a per cpu basis)
787 */
788void
789lwkt_deschedule_self(void)
790{
791 thread_t td = curthread;
792
793 crit_enter();
794 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
795 _lwkt_dequeue(td);
796 crit_exit();
797}
798
799/*
800 * Generic deschedule. Descheduling threads other then your own should be
801 * done only in carefully controlled circumstances. Descheduling is
802 * asynchronous.
803 *
804 * This function may block if the cpu has run out of messages.
8ad65e08
MD
805 */
806void
807lwkt_deschedule(thread_t td)
808{
f1d1c3fa
MD
809 crit_enter();
810 if (td == curthread) {
811 _lwkt_dequeue(td);
812 } else {
a72187e9 813 if (td->td_gd == mycpu) {
f1d1c3fa
MD
814 _lwkt_dequeue(td);
815 } else {
a72187e9 816 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
817 }
818 }
819 crit_exit();
820}
821
4b5f931b
MD
822/*
823 * Set the target thread's priority. This routine does not automatically
824 * switch to a higher priority thread, LWKT threads are not designed for
825 * continuous priority changes. Yield if you want to switch.
826 *
827 * We have to retain the critical section count which uses the high bits
26a0694b
MD
828 * of the td_pri field. The specified priority may also indicate zero or
829 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
830 */
831void
832lwkt_setpri(thread_t td, int pri)
833{
26a0694b 834 KKASSERT(pri >= 0);
a72187e9 835 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
836 crit_enter();
837 if (td->td_flags & TDF_RUNQ) {
838 _lwkt_dequeue(td);
839 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
840 _lwkt_enqueue(td);
841 } else {
842 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
843 }
844 crit_exit();
845}
846
847void
848lwkt_setpri_self(int pri)
849{
850 thread_t td = curthread;
851
4b5f931b
MD
852 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
853 crit_enter();
854 if (td->td_flags & TDF_RUNQ) {
855 _lwkt_dequeue(td);
856 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
857 _lwkt_enqueue(td);
858 } else {
859 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
860 }
861 crit_exit();
862}
863
864struct proc *
865lwkt_preempted_proc(void)
866{
73e4f7b9 867 thread_t td = curthread;
4b5f931b
MD
868 while (td->td_preempted)
869 td = td->td_preempted;
870 return(td->td_proc);
871}
872
ece04fd0
MD
873typedef struct lwkt_gettoken_req {
874 lwkt_token_t tok;
875 int cpu;
876} lwkt_gettoken_req;
877
878#if 0
4b5f931b 879
f1d1c3fa
MD
880/*
881 * This function deschedules the current thread and blocks on the specified
882 * wait queue. We obtain ownership of the wait queue in order to block
883 * on it. A generation number is used to interlock the wait queue in case
884 * it gets signalled while we are blocked waiting on the token.
885 *
886 * Note: alternatively we could dequeue our thread and then message the
887 * target cpu owning the wait queue. YYY implement as sysctl.
888 *
889 * Note: wait queue signals normally ping-pong the cpu as an optimization.
890 */
96728c05 891
f1d1c3fa 892void
ae8050a4 893lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
894{
895 thread_t td = curthread;
f1d1c3fa 896
f1d1c3fa 897 lwkt_gettoken(&w->wa_token);
ae8050a4 898 if (w->wa_gen == *gen) {
f1d1c3fa
MD
899 _lwkt_dequeue(td);
900 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
901 ++w->wa_count;
902 td->td_wait = w;
ae8050a4 903 td->td_wmesg = wmesg;
ece04fd0 904again:
f1d1c3fa 905 lwkt_switch();
ece04fd0
MD
906 lwkt_regettoken(&w->wa_token);
907 if (td->td_wmesg != NULL) {
908 _lwkt_dequeue(td);
909 goto again;
910 }
8ad65e08 911 }
ae8050a4
MD
912 /* token might be lost, doesn't matter for gen update */
913 *gen = w->wa_gen;
f1d1c3fa
MD
914 lwkt_reltoken(&w->wa_token);
915}
916
917/*
918 * Signal a wait queue. We gain ownership of the wait queue in order to
919 * signal it. Once a thread is removed from the wait queue we have to
920 * deal with the cpu owning the thread.
921 *
922 * Note: alternatively we could message the target cpu owning the wait
923 * queue. YYY implement as sysctl.
924 */
925void
ece04fd0 926lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
927{
928 thread_t td;
929 int count;
930
931 lwkt_gettoken(&w->wa_token);
932 ++w->wa_gen;
ece04fd0
MD
933 if (count < 0)
934 count = w->wa_count;
f1d1c3fa
MD
935 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
936 --count;
937 --w->wa_count;
938 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
939 td->td_wait = NULL;
ae8050a4 940 td->td_wmesg = NULL;
a72187e9 941 if (td->td_gd == mycpu) {
f1d1c3fa
MD
942 _lwkt_enqueue(td);
943 } else {
a72187e9 944 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
945 }
946 lwkt_regettoken(&w->wa_token);
947 }
948 lwkt_reltoken(&w->wa_token);
949}
950
ece04fd0
MD
951#endif
952
f1d1c3fa 953/*
96728c05 954 * Acquire ownership of a token
f1d1c3fa 955 *
96728c05 956 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
957 * section side effects, depending on its purpose. These side effects
958 * guarentee that you will maintain ownership of the token as long as you
959 * do not block. If you block you may lose access to the token (but you
960 * must still release it even if you lose your access to it).
961 *
96728c05 962 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 963 * a token, but do we really only need to disable IPIs ?
96728c05
MD
964 *
965 * YYY certain tokens could be made to act like mutexes when performance
966 * would be better (e.g. t_cpu == -1). This is not yet implemented.
967 *
a2a5ad0d
MD
968 * YYY the tokens replace 4.x's simplelocks for the most part, but this
969 * means that 4.x does not expect a switch so for now we cannot switch
970 * when waiting for an IPI to be returned.
971 *
972 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
973 * it and then block. The IPI causes the token to be given away to the
974 * requesting cpu, unless it has already changed hands. Since only the
975 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 976 * This needs serious optimization.
f1d1c3fa 977 */
57c254db
MD
978
979#ifdef SMP
980
96728c05
MD
981static
982void
983lwkt_gettoken_remote(void *arg)
984{
985 lwkt_gettoken_req *req = arg;
986 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 987#ifdef INVARIANTS
a2a5ad0d
MD
988 if (token_debug)
989 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 990#endif
96728c05 991 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
992 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
993 /* else set reqcpu to point to current cpu for release */
96728c05
MD
994 }
995}
996
57c254db
MD
997#endif
998
8a8d5d85 999int
f1d1c3fa
MD
1000lwkt_gettoken(lwkt_token_t tok)
1001{
1002 /*
1003 * Prevent preemption so the token can't be taken away from us once
1004 * we gain ownership of it. Use a synchronous request which might
1005 * block. The request will be forwarded as necessary playing catchup
1006 * to the token.
1007 */
96728c05 1008
f1d1c3fa 1009 crit_enter();
57c254db 1010#ifdef INVARIANTS
1d4c9574
MD
1011 if (curthread->td_pri > 1800) {
1012 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1013 tok, ((int **)&tok)[-1]);
1014 }
a2a5ad0d
MD
1015 if (curthread->td_pri > 2000) {
1016 curthread->td_pri = 1000;
1017 panic("too HIGH!");
57c254db
MD
1018 }
1019#endif
96728c05 1020#ifdef SMP
d0e06f83 1021 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1022 struct lwkt_gettoken_req req;
1023 int seq;
96728c05
MD
1024 int dcpu;
1025
1026 req.cpu = mycpu->gd_cpuid;
1027 req.tok = tok;
1028 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1029 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1030#ifdef INVARIANTS
a2a5ad0d
MD
1031 if (token_debug)
1032 printf("REQT%d ", dcpu);
634081ff 1033#endif
96728c05
MD
1034 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1035 lwkt_wait_ipiq(dcpu, seq);
634081ff 1036#ifdef INVARIANTS
a2a5ad0d
MD
1037 if (token_debug)
1038 printf("REQR%d ", tok->t_cpu);
634081ff 1039#endif
f1d1c3fa
MD
1040 }
1041#endif
1042 /*
1043 * leave us in a critical section on return. This will be undone
8a8d5d85 1044 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1045 */
8a8d5d85 1046 return(++tok->t_gen);
f1d1c3fa
MD
1047}
1048
96728c05
MD
1049/*
1050 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1051 * failure.
1052 */
1053int
1054lwkt_trytoken(lwkt_token_t tok)
1055{
1056 crit_enter();
1057#ifdef SMP
1058 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1059 crit_exit();
96728c05
MD
1060 return(0);
1061 }
1062#endif
1063 /* leave us in the critical section */
1064 ++tok->t_gen;
1065 return(1);
1066}
1067
f1d1c3fa
MD
1068/*
1069 * Release your ownership of a token. Releases must occur in reverse
1070 * order to aquisitions, eventually so priorities can be unwound properly
1071 * like SPLs. At the moment the actual implemention doesn't care.
1072 *
1073 * We can safely hand a token that we own to another cpu without notifying
1074 * it, but once we do we can't get it back without requesting it (unless
1075 * the other cpu hands it back to us before we check).
1076 *
1077 * We might have lost the token, so check that.
7ba9c17c
MD
1078 *
1079 * Return the token's generation number. The number is useful to callers
1080 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1081 */
7ba9c17c 1082int
f1d1c3fa
MD
1083lwkt_reltoken(lwkt_token_t tok)
1084{
7ba9c17c
MD
1085 int gen;
1086
d0e06f83 1087 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1088 tok->t_cpu = tok->t_reqcpu;
1089 }
7ba9c17c 1090 gen = tok->t_gen;
f1d1c3fa 1091 crit_exit();
7ba9c17c 1092 return(gen);
f1d1c3fa
MD
1093}
1094
1095/*
7ba9c17c
MD
1096 * Reacquire a token that might have been lost. 0 is returned if the
1097 * generation has not changed (nobody stole the token from us), -1 is
1098 * returned otherwise. The token is reacquired regardless but the
1099 * generation number is not bumped further if we already own the token.
8a8d5d85 1100 *
7ba9c17c
MD
1101 * For efficiency we inline the best-case situation for lwkt_regettoken()
1102 * (i.e .we still own the token).
8a8d5d85
MD
1103 */
1104int
1105lwkt_gentoken(lwkt_token_t tok, int *gen)
1106{
7ba9c17c 1107 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1108 return(0);
7ba9c17c
MD
1109 *gen = lwkt_regettoken(tok);
1110 return(-1);
8a8d5d85
MD
1111}
1112
8a8d5d85 1113/*
435ff993
MD
1114 * Re-acquire a token that might have been lost. The generation number
1115 * is bumped and returned regardless of whether the token had been lost
1116 * or not (because we only have cpu granularity we have to bump the token
1117 * either way).
f1d1c3fa
MD
1118 */
1119int
1120lwkt_regettoken(lwkt_token_t tok)
1121{
96728c05 1122 /* assert we are in a critical section */
d0e06f83 1123 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1124#ifdef SMP
d0e06f83 1125 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1126 struct lwkt_gettoken_req req;
1127 int seq;
96728c05 1128 int dcpu;
57c254db 1129
96728c05
MD
1130 req.cpu = mycpu->gd_cpuid;
1131 req.tok = tok;
1132 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1133 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1134#ifdef INVARIANTS
cb973d15
MD
1135 if (token_debug)
1136 printf("REQT%d ", dcpu);
634081ff 1137#endif
96728c05
MD
1138 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1139 lwkt_wait_ipiq(dcpu, seq);
634081ff 1140#ifdef INVARIATNS
cb973d15
MD
1141 if (token_debug)
1142 printf("REQR%d ", tok->t_cpu);
634081ff 1143#endif
f1d1c3fa 1144 }
f1d1c3fa 1145#endif
96728c05 1146 }
435ff993 1147 ++tok->t_gen;
8a8d5d85 1148 return(tok->t_gen);
8ad65e08
MD
1149}
1150
72740893
MD
1151void
1152lwkt_inittoken(lwkt_token_t tok)
1153{
1154 /*
1155 * Zero structure and set cpu owner and reqcpu to cpu 0.
1156 */
1157 bzero(tok, sizeof(*tok));
1158}
1159
99df837e
MD
1160/*
1161 * Create a kernel process/thread/whatever. It shares it's address space
1162 * with proc0 - ie: kernel only.
1163 *
365fa13f
MD
1164 * NOTE! By default new threads are created with the MP lock held. A
1165 * thread which does not require the MP lock should release it by calling
1166 * rel_mplock() at the start of the new thread.
99df837e
MD
1167 */
1168int
1169lwkt_create(void (*func)(void *), void *arg,
73e4f7b9 1170 struct thread **tdp, thread_t template, int tdflags,
ef0fdad1 1171 const char *fmt, ...)
99df837e 1172{
73e4f7b9 1173 thread_t td;
99df837e
MD
1174 va_list ap;
1175
a2a5ad0d
MD
1176 td = lwkt_alloc_thread(template);
1177 if (tdp)
1178 *tdp = td;
99df837e 1179 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1180 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1181#ifdef SMP
1182 td->td_mpcount = 1;
1183#endif
99df837e
MD
1184
1185 /*
1186 * Set up arg0 for 'ps' etc
1187 */
1188 va_start(ap, fmt);
1189 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1190 va_end(ap);
1191
1192 /*
1193 * Schedule the thread to run
1194 */
ef0fdad1
MD
1195 if ((td->td_flags & TDF_STOPREQ) == 0)
1196 lwkt_schedule(td);
1197 else
1198 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1199 return 0;
1200}
1201
1202/*
1203 * Destroy an LWKT thread. Warning! This function is not called when
1204 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1205 * uses a different reaping mechanism.
1206 */
1207void
1208lwkt_exit(void)
1209{
1210 thread_t td = curthread;
1211
1212 if (td->td_flags & TDF_VERBOSE)
1213 printf("kthread %p %s has exited\n", td, td->td_comm);
1214 crit_enter();
1215 lwkt_deschedule_self();
1216 ++mycpu->gd_tdfreecount;
1217 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1218 cpu_thread_exit();
1219}
1220
1221/*
1222 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1223 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1224 *
1225 * NOTE! By default kthreads are created with the MP lock held. A
1226 * thread which does not require the MP lock should release it by calling
1227 * rel_mplock() at the start of the new thread.
99df837e
MD
1228 */
1229int
1230kthread_create(void (*func)(void *), void *arg,
1231 struct thread **tdp, const char *fmt, ...)
1232{
73e4f7b9 1233 thread_t td;
99df837e
MD
1234 va_list ap;
1235
a2a5ad0d
MD
1236 td = lwkt_alloc_thread(NULL);
1237 if (tdp)
1238 *tdp = td;
99df837e
MD
1239 cpu_set_thread_handler(td, kthread_exit, func, arg);
1240 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1241#ifdef SMP
1242 td->td_mpcount = 1;
1243#endif
99df837e
MD
1244
1245 /*
1246 * Set up arg0 for 'ps' etc
1247 */
1248 va_start(ap, fmt);
1249 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1250 va_end(ap);
1251
1252 /*
1253 * Schedule the thread to run
1254 */
1255 lwkt_schedule(td);
1256 return 0;
1257}
1258
26a0694b
MD
1259void
1260crit_panic(void)
1261{
73e4f7b9 1262 thread_t td = curthread;
26a0694b
MD
1263 int lpri = td->td_pri;
1264
1265 td->td_pri = 0;
1266 panic("td_pri is/would-go negative! %p %d", td, lpri);
1267}
1268
99df837e
MD
1269/*
1270 * Destroy an LWKT thread. Warning! This function is not called when
1271 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1272 * uses a different reaping mechanism.
1273 *
1274 * XXX duplicates lwkt_exit()
1275 */
1276void
1277kthread_exit(void)
1278{
1279 lwkt_exit();
1280}
1281
96728c05
MD
1282#ifdef SMP
1283
1284/*
1285 * Send a function execution request to another cpu. The request is queued
1286 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1287 * possible target cpu. The FIFO can be written.
1288 *
1289 * YYY If the FIFO fills up we have to enable interrupts and process the
1290 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1291 * Create a CPU_*() function to do this!
1292 *
46a3f46d
MD
1293 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1294 * end will take care of any pending interrupts.
1295 *
96728c05
MD
1296 * Must be called from a critical section.
1297 */
1298int
1299lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1300{
1301 lwkt_ipiq_t ip;
1302 int windex;
a2a5ad0d 1303 struct globaldata *gd = mycpu;
96728c05 1304
a2a5ad0d 1305 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1306 func(arg);
1307 return(0);
1308 }
cb973d15 1309 crit_enter();
a2a5ad0d
MD
1310 ++gd->gd_intr_nesting_level;
1311#ifdef INVARIANTS
1312 if (gd->gd_intr_nesting_level > 20)
1313 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1314#endif
96728c05
MD
1315 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1316 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1317 ++ipiq_count;
a2a5ad0d 1318 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1319
1320 /*
1321 * We always drain before the FIFO becomes full so it should never
1322 * become full. We need to leave enough entries to deal with
1323 * reentrancy.
1324 */
1325 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1326 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1327 ip->ip_func[windex] = func;
1328 ip->ip_arg[windex] = arg;
1329 /* YYY memory barrier */
1330 ++ip->ip_windex;
96728c05
MD
1331 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1332 unsigned int eflags = read_eflags();
1333 cpu_enable_intr();
1334 ++ipiq_fifofull;
cb973d15 1335 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1336 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1337 lwkt_process_ipiq();
1338 }
1339 write_eflags(eflags);
1340 }
a2a5ad0d 1341 --gd->gd_intr_nesting_level;
96728c05 1342 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1343 crit_exit();
96728c05
MD
1344 return(ip->ip_windex);
1345}
1346
cb973d15
MD
1347/*
1348 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1349 * The message will not be sent to stopped cpus.
cb973d15
MD
1350 */
1351void
1352lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1353{
1354 int cpuid;
1355
435ff993 1356 mask &= ~stopped_cpus;
cb973d15
MD
1357 while (mask) {
1358 cpuid = bsfl(mask);
1359 lwkt_send_ipiq(cpuid, func, arg);
1360 mask &= ~(1 << cpuid);
1361 }
1362}
1363
96728c05
MD
1364/*
1365 * Wait for the remote cpu to finish processing a function.
1366 *
1367 * YYY we have to enable interrupts and process the IPIQ while waiting
1368 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1369 * function to do this! YYY we really should 'block' here.
1370 *
1371 * Must be called from a critical section. Thsi routine may be called
1372 * from an interrupt (for example, if an interrupt wakes a foreign thread
1373 * up).
1374 */
1375void
1376lwkt_wait_ipiq(int dcpu, int seq)
1377{
1378 lwkt_ipiq_t ip;
a2a5ad0d 1379 int maxc = 100000000;
96728c05
MD
1380
1381 if (dcpu != mycpu->gd_cpuid) {
1382 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1383 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1384 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1385 unsigned int eflags = read_eflags();
1386 cpu_enable_intr();
cb973d15 1387 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1388 lwkt_process_ipiq();
a2a5ad0d 1389 if (--maxc == 0)
cb973d15 1390 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1391 if (maxc < -1000000)
1392 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1393 }
1394 write_eflags(eflags);
1395 }
1396 }
1397}
1398
1399/*
1400 * Called from IPI interrupt (like a fast interrupt), which has placed
1401 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1402 * May also be called from doreti or splz, or be reentrantly called
1403 * indirectly through the ip_func[] we run.
96728c05
MD
1404 */
1405void
1406lwkt_process_ipiq(void)
1407{
1408 int n;
1409 int cpuid = mycpu->gd_cpuid;
1410
1411 for (n = 0; n < ncpus; ++n) {
1412 lwkt_ipiq_t ip;
1413 int ri;
1414
1415 if (n == cpuid)
1416 continue;
1417 ip = globaldata_find(n)->gd_ipiq;
1418 if (ip == NULL)
1419 continue;
1420 ip = &ip[cpuid];
cb973d15
MD
1421
1422 /*
1423 * Note: xindex is only updated after we are sure the function has
1424 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1425 * function may send an IPI which may block/drain.
1426 */
96728c05
MD
1427 while (ip->ip_rindex != ip->ip_windex) {
1428 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1429 ++ip->ip_rindex;
cb973d15
MD
1430 ip->ip_func[ri](ip->ip_arg[ri]);
1431 /* YYY memory barrier */
1432 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1433 }
1434 }
1435}
1436
1437#else
1438
1439int
1440lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1441{
1442 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1443 return(0); /* NOT REACHED */
1444}
1445
1446void
1447lwkt_wait_ipiq(int dcpu, int seq)
1448{
1449 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1450}
1451
1452#endif