'Building databases' has 10 seconds worth of sleeps that it doesn't need.
[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 *
e2565a42 26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.41 2003/11/09 02:22:36 dillon Exp $
75cdbe6c
MD
27 */
28
29/*
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
f1d1c3fa 34 *
75cdbe6c
MD
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
8ad65e08
MD
39 */
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/kernel.h>
44#include <sys/proc.h>
45#include <sys/rtprio.h>
46#include <sys/queue.h>
f1d1c3fa 47#include <sys/thread2.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f1d1c3fa 52
7d0bac62
MD
53#include <vm/vm.h>
54#include <vm/vm_param.h>
55#include <vm/vm_kern.h>
56#include <vm/vm_object.h>
57#include <vm/vm_page.h>
58#include <vm/vm_map.h>
59#include <vm/vm_pager.h>
60#include <vm/vm_extern.h>
61#include <vm/vm_zone.h>
62
99df837e 63#include <machine/stdarg.h>
57c254db 64#include <machine/ipl.h>
96728c05 65#include <machine/smp.h>
99df837e 66
7d0bac62 67static int untimely_switch = 0;
4b5f931b 68SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
57c254db
MD
69#ifdef INVARIANTS
70static int token_debug = 0;
71SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
72#endif
4b5f931b
MD
73static quad_t switch_count = 0;
74SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
75static quad_t preempt_hit = 0;
76SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
77static quad_t preempt_miss = 0;
78SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b
MD
79static quad_t preempt_weird = 0;
80SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
96728c05
MD
81static quad_t ipiq_count = 0;
82SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
83static quad_t ipiq_fifofull = 0;
84SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
7d0bac62 85
4b5f931b
MD
86/*
87 * These helper procedures handle the runq, they can only be called from
88 * within a critical section.
75cdbe6c
MD
89 *
90 * WARNING! Prior to SMP being brought up it is possible to enqueue and
91 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
92 * instead of 'mycpu' when referencing the globaldata structure. Once
93 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 94 */
f1d1c3fa
MD
95static __inline
96void
97_lwkt_dequeue(thread_t td)
98{
99 if (td->td_flags & TDF_RUNQ) {
4b5f931b 100 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 101 struct globaldata *gd = td->td_gd;
4b5f931b 102
f1d1c3fa 103 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
104 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
105 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
106 }
107}
108
109static __inline
110void
111_lwkt_enqueue(thread_t td)
112{
113 if ((td->td_flags & TDF_RUNQ) == 0) {
4b5f931b 114 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 115 struct globaldata *gd = td->td_gd;
4b5f931b 116
f1d1c3fa 117 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
118 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
119 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
120 }
121}
8ad65e08 122
57c254db
MD
123static __inline
124int
125_lwkt_wantresched(thread_t ntd, thread_t cur)
126{
127 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
128}
129
8ad65e08
MD
130/*
131 * LWKTs operate on a per-cpu basis
132 *
73e4f7b9 133 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
134 */
135void
136lwkt_gdinit(struct globaldata *gd)
137{
4b5f931b
MD
138 int i;
139
140 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
141 TAILQ_INIT(&gd->gd_tdrunq[i]);
142 gd->gd_runqmask = 0;
73e4f7b9 143 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
144}
145
7d0bac62
MD
146/*
147 * Initialize a thread wait structure prior to first use.
148 *
149 * NOTE! called from low level boot code, we cannot do anything fancy!
150 */
151void
152lwkt_init_wait(lwkt_wait_t w)
153{
154 TAILQ_INIT(&w->wa_waitq);
155}
156
157/*
158 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
159 * or LWKT start address before it can be scheduled. If the target cpu is
160 * -1 the thread will be created on the current cpu.
0cfcada1
MD
161 *
162 * If you intend to create a thread without a process context this function
163 * does everything except load the startup and switcher function.
7d0bac62
MD
164 */
165thread_t
75cdbe6c 166lwkt_alloc_thread(struct thread *td, int cpu)
7d0bac62 167{
99df837e 168 void *stack;
ef0fdad1 169 int flags = 0;
7d0bac62 170
ef0fdad1 171 if (td == NULL) {
26a0694b 172 crit_enter();
ef0fdad1
MD
173 if (mycpu->gd_tdfreecount > 0) {
174 --mycpu->gd_tdfreecount;
175 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
d9eea1a5 176 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1
MD
177 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
178 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
179 crit_exit();
180 stack = td->td_kstack;
181 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
182 } else {
183 crit_exit();
184 td = zalloc(thread_zone);
185 td->td_kstack = NULL;
186 flags |= TDF_ALLOCATED_THREAD;
187 }
188 }
189 if ((stack = td->td_kstack) == NULL) {
99df837e 190 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
ef0fdad1 191 flags |= TDF_ALLOCATED_STACK;
99df837e 192 }
75cdbe6c
MD
193 if (cpu < 0)
194 lwkt_init_thread(td, stack, flags, mycpu);
195 else
196 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
99df837e 197 return(td);
7d0bac62
MD
198}
199
200/*
201 * Initialize a preexisting thread structure. This function is used by
202 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
203 *
f8c3996b
MD
204 * All threads start out in a critical section at a priority of
205 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
206 * appropriate. This function may send an IPI message when the
207 * requested cpu is not the current cpu and consequently gd_tdallq may
208 * not be initialized synchronously from the point of view of the originating
209 * cpu.
210 *
211 * NOTE! we have to be careful in regards to creating threads for other cpus
212 * if SMP has not yet been activated.
7d0bac62 213 */
75cdbe6c
MD
214static void
215lwkt_init_thread_remote(void *arg)
216{
217 thread_t td = arg;
218
219 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
220}
221
7d0bac62 222void
26a0694b 223lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
7d0bac62 224{
99df837e
MD
225 bzero(td, sizeof(struct thread));
226 td->td_kstack = stack;
227 td->td_flags |= flags;
26a0694b 228 td->td_gd = gd;
f8c3996b 229 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
ece04fd0 230 lwkt_init_port(&td->td_msgport, td);
99df837e 231 pmap_init_thread(td);
75cdbe6c
MD
232 if (smp_active == 0 || gd == mycpu) {
233 crit_enter();
234 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
235 crit_exit();
236 } else {
237 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
238 }
73e4f7b9
MD
239}
240
241void
242lwkt_set_comm(thread_t td, const char *ctl, ...)
243{
e2565a42 244 __va_list va;
73e4f7b9 245
e2565a42 246 __va_start(va, ctl);
73e4f7b9 247 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 248 __va_end(va);
7d0bac62
MD
249}
250
99df837e 251void
73e4f7b9 252lwkt_hold(thread_t td)
99df837e 253{
73e4f7b9
MD
254 ++td->td_refs;
255}
256
257void
258lwkt_rele(thread_t td)
259{
260 KKASSERT(td->td_refs > 0);
261 --td->td_refs;
262}
263
264void
265lwkt_wait_free(thread_t td)
266{
267 while (td->td_refs)
377d4740 268 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
269}
270
271void
272lwkt_free_thread(thread_t td)
273{
274 struct globaldata *gd = mycpu;
275
d9eea1a5 276 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
277 ("lwkt_free_thread: did not exit! %p", td));
278
279 crit_enter();
73e4f7b9
MD
280 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
281 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
282 (td->td_flags & TDF_ALLOCATED_THREAD)
283 ) {
73e4f7b9
MD
284 ++gd->gd_tdfreecount;
285 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
99df837e
MD
286 crit_exit();
287 } else {
288 crit_exit();
289 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
290 kmem_free(kernel_map,
291 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
73e4f7b9 292 /* gd invalid */
99df837e
MD
293 td->td_kstack = NULL;
294 }
295 if (td->td_flags & TDF_ALLOCATED_THREAD)
296 zfree(thread_zone, td);
297 }
298}
299
300
8ad65e08
MD
301/*
302 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
303 * switch to the idlethread. Switching must occur within a critical
304 * section to avoid races with the scheduling queue.
305 *
306 * We always have full control over our cpu's run queue. Other cpus
307 * that wish to manipulate our queue must use the cpu_*msg() calls to
308 * talk to our cpu, so a critical section is all that is needed and
309 * the result is very, very fast thread switching.
310 *
96728c05
MD
311 * The LWKT scheduler uses a fixed priority model and round-robins at
312 * each priority level. User process scheduling is a totally
313 * different beast and LWKT priorities should not be confused with
314 * user process priorities.
f1d1c3fa 315 *
96728c05
MD
316 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
317 * cleans it up. Note that the td_switch() function cannot do anything that
318 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
319 * the target thread (not the current thread). It's nice to have a scheduler
320 * that does not need the MP lock to work because it allows us to do some
321 * really cool high-performance MP lock optimizations.
8ad65e08 322 */
96728c05 323
8ad65e08
MD
324void
325lwkt_switch(void)
326{
4b5f931b 327 struct globaldata *gd;
f1d1c3fa 328 thread_t td = curthread;
8ad65e08 329 thread_t ntd;
8a8d5d85
MD
330#ifdef SMP
331 int mpheld;
332#endif
8ad65e08 333
46a3f46d
MD
334 /*
335 * Switching from within a 'fast' (non thread switched) interrupt is
336 * illegal.
337 */
338 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
03aa8d99 339 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
96728c05 340 }
ef0fdad1 341
cb973d15
MD
342 /*
343 * Passive release (used to transition from user to kernel mode
344 * when we block or switch rather then when we enter the kernel).
345 * This function is NOT called if we are switching into a preemption
346 * or returning from a preemption. Typically this causes us to lose
347 * our P_CURPROC designation (if we have one) and become a true LWKT
348 * thread, and may also hand P_CURPROC to another process and schedule
349 * its thread.
350 */
351 if (td->td_release)
352 td->td_release(td);
353
f1d1c3fa 354 crit_enter();
4b5f931b 355 ++switch_count;
8a8d5d85
MD
356
357#ifdef SMP
358 /*
359 * td_mpcount cannot be used to determine if we currently hold the
360 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
361 * to get the lock, and switch out if it can't. Our ownership of
362 * the actual lock will remain stable while we are in a critical section
363 * (but, of course, another cpu may own or release the lock so the
364 * actual value of mp_lock is not stable).
8a8d5d85
MD
365 */
366 mpheld = MP_LOCK_HELD();
367#endif
99df837e
MD
368 if ((ntd = td->td_preempted) != NULL) {
369 /*
370 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
371 * thread. This occurs transparently, whether the preempted thread
372 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
373 * itself).
374 *
375 * We have to setup the MP lock for the original thread after backing
376 * out the adjustment that was made to curthread when the original
377 * was preempted.
99df837e 378 */
26a0694b 379 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 380#ifdef SMP
96728c05
MD
381 if (ntd->td_mpcount && mpheld == 0) {
382 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
383 td, ntd, td->td_mpcount, ntd->td_mpcount);
384 }
8a8d5d85
MD
385 if (ntd->td_mpcount) {
386 td->td_mpcount -= ntd->td_mpcount;
387 KKASSERT(td->td_mpcount >= 0);
388 }
389#endif
26a0694b 390 ntd->td_flags |= TDF_PREEMPT_DONE;
8a8d5d85 391 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 392 } else {
4b5f931b
MD
393 /*
394 * Priority queue / round-robin at each priority. Note that user
395 * processes run at a fixed, low priority and the user process
396 * scheduler deals with interactions between user processes
397 * by scheduling and descheduling them from the LWKT queue as
398 * necessary.
8a8d5d85
MD
399 *
400 * We have to adjust the MP lock for the target thread. If we
401 * need the MP lock and cannot obtain it we try to locate a
402 * thread that does not need the MP lock.
4b5f931b
MD
403 */
404 gd = mycpu;
4b5f931b
MD
405again:
406 if (gd->gd_runqmask) {
407 int nq = bsrl(gd->gd_runqmask);
408 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
409 gd->gd_runqmask &= ~(1 << nq);
410 goto again;
411 }
8a8d5d85
MD
412#ifdef SMP
413 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
414 /*
96728c05
MD
415 * Target needs MP lock and we couldn't get it, try
416 * to locate a thread which does not need the MP lock
3c23a41a 417 * to run. If we cannot locate a thread spin in idle.
8a8d5d85
MD
418 */
419 u_int32_t rqmask = gd->gd_runqmask;
420 while (rqmask) {
421 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
422 if (ntd->td_mpcount == 0)
423 break;
424 }
425 if (ntd)
426 break;
427 rqmask &= ~(1 << nq);
428 nq = bsrl(rqmask);
429 }
430 if (ntd == NULL) {
a2a5ad0d
MD
431 ntd = &gd->gd_idlethread;
432 ntd->td_flags |= TDF_IDLE_NOHLT;
8a8d5d85
MD
433 } else {
434 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
435 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
436 }
437 } else {
438 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
439 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
440 }
441#else
4b5f931b
MD
442 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
443 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 444#endif
4b5f931b 445 } else {
3c23a41a
MD
446 /*
447 * Nothing to run but we may still need the BGL to deal with
448 * pending interrupts, spin in idle if so.
449 */
a2a5ad0d 450 ntd = &gd->gd_idlethread;
235957ed 451 if (gd->gd_reqflags)
3c23a41a 452 ntd->td_flags |= TDF_IDLE_NOHLT;
4b5f931b 453 }
f1d1c3fa 454 }
26a0694b
MD
455 KASSERT(ntd->td_pri >= TDPRI_CRIT,
456 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
457
458 /*
459 * Do the actual switch. If the new target does not need the MP lock
460 * and we are holding it, release the MP lock. If the new target requires
461 * the MP lock we have already acquired it for the target.
462 */
463#ifdef SMP
464 if (ntd->td_mpcount == 0 ) {
465 if (MP_LOCK_HELD())
466 cpu_rel_mplock();
467 } else {
468 ASSERT_MP_LOCK_HELD();
469 }
470#endif
8a8d5d85 471 if (td != ntd) {
f1d1c3fa 472 td->td_switch(ntd);
8a8d5d85 473 }
96728c05 474
f1d1c3fa 475 crit_exit();
8ad65e08
MD
476}
477
cb973d15
MD
478/*
479 * Switch if another thread has a higher priority. Do not switch to other
480 * threads at the same priority.
481 */
482void
483lwkt_maybe_switch()
484{
485 struct globaldata *gd = mycpu;
486 struct thread *td = gd->gd_curthread;
487
488 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
489 lwkt_switch();
490 }
491}
492
b68b7282 493/*
96728c05
MD
494 * Request that the target thread preempt the current thread. Preemption
495 * only works under a specific set of conditions:
b68b7282 496 *
96728c05
MD
497 * - We are not preempting ourselves
498 * - The target thread is owned by the current cpu
499 * - We are not currently being preempted
500 * - The target is not currently being preempted
501 * - We are able to satisfy the target's MP lock requirements (if any).
502 *
503 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
504 * this is called via lwkt_schedule() through the td_preemptable callback.
505 * critpri is the managed critical priority that we should ignore in order
506 * to determine whether preemption is possible (aka usually just the crit
507 * priority of lwkt_schedule() itself).
b68b7282 508 *
26a0694b
MD
509 * XXX at the moment we run the target thread in a critical section during
510 * the preemption in order to prevent the target from taking interrupts
511 * that *WE* can't. Preemption is strictly limited to interrupt threads
512 * and interrupt-like threads, outside of a critical section, and the
513 * preempted source thread will be resumed the instant the target blocks
514 * whether or not the source is scheduled (i.e. preemption is supposed to
515 * be as transparent as possible).
4b5f931b 516 *
8a8d5d85
MD
517 * The target thread inherits our MP count (added to its own) for the
518 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
519 * MP lock during the preemption. Therefore, any preempting targets must be
520 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
521 * out of sync with the physical mp_lock, but we do not have to preserve
522 * the original ownership of the lock if it was out of synch (that is, we
523 * can leave it synchronized on return).
b68b7282
MD
524 */
525void
96728c05 526lwkt_preempt(thread_t ntd, int critpri)
b68b7282 527{
46a3f46d
MD
528 struct globaldata *gd = mycpu;
529 thread_t td = gd->gd_curthread;
8a8d5d85
MD
530#ifdef SMP
531 int mpheld;
57c254db 532 int savecnt;
8a8d5d85 533#endif
b68b7282 534
26a0694b 535 /*
96728c05
MD
536 * The caller has put us in a critical section. We can only preempt
537 * if the caller of the caller was not in a critical section (basically
57c254db
MD
538 * a local interrupt), as determined by the 'critpri' parameter. If
539 * we are unable to preempt
96728c05
MD
540 *
541 * YYY The target thread must be in a critical section (else it must
542 * inherit our critical section? I dunno yet).
26a0694b
MD
543 */
544 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 545
cb973d15 546 need_resched();
57c254db
MD
547 if (!_lwkt_wantresched(ntd, td)) {
548 ++preempt_miss;
549 return;
550 }
96728c05
MD
551 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
552 ++preempt_miss;
553 return;
554 }
555#ifdef SMP
46a3f46d 556 if (ntd->td_gd != gd) {
96728c05
MD
557 ++preempt_miss;
558 return;
559 }
560#endif
26a0694b
MD
561 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
562 ++preempt_weird;
563 return;
564 }
565 if (ntd->td_preempted) {
4b5f931b 566 ++preempt_hit;
26a0694b 567 return;
b68b7282 568 }
8a8d5d85 569#ifdef SMP
a2a5ad0d
MD
570 /*
571 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
572 * to or from the MP lock. In this case td_mpcount will be pre-disposed
573 * (non-zero) but not actually synchronized with the actual state of the
574 * lock. We can use it to imply an MP lock requirement for the
575 * preemption but we cannot use it to test whether we hold the MP lock
576 * or not.
a2a5ad0d 577 */
96728c05 578 savecnt = td->td_mpcount;
71ef2f5c 579 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
580 ntd->td_mpcount += td->td_mpcount;
581 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
582 ntd->td_mpcount -= td->td_mpcount;
583 ++preempt_miss;
584 return;
585 }
586#endif
26a0694b
MD
587
588 ++preempt_hit;
589 ntd->td_preempted = td;
590 td->td_flags |= TDF_PREEMPT_LOCK;
591 td->td_switch(ntd);
592 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
593#ifdef SMP
594 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
595 mpheld = MP_LOCK_HELD();
596 if (mpheld && td->td_mpcount == 0)
96728c05 597 cpu_rel_mplock();
71ef2f5c 598 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
599 panic("lwkt_preempt(): MP lock was not held through");
600#endif
26a0694b
MD
601 ntd->td_preempted = NULL;
602 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
603}
604
f1d1c3fa
MD
605/*
606 * Yield our thread while higher priority threads are pending. This is
607 * typically called when we leave a critical section but it can be safely
608 * called while we are in a critical section.
609 *
610 * This function will not generally yield to equal priority threads but it
611 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 612 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
613 * lwkt_yield_quick().
614 *
235957ed 615 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
616 * came along but was blocked and made pending.
617 *
f1d1c3fa
MD
618 * (self contained on a per cpu basis)
619 */
620void
621lwkt_yield_quick(void)
622{
7966cb69
MD
623 globaldata_t gd = mycpu;
624 thread_t td = gd->gd_curthread;
ef0fdad1 625
a2a5ad0d 626 /*
235957ed 627 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
628 * it with a non-zero cpl then we might not wind up calling splz after
629 * a task switch when the critical section is exited even though the
46a3f46d 630 * new task could accept the interrupt.
a2a5ad0d
MD
631 *
632 * XXX from crit_exit() only called after last crit section is released.
633 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
634 *
635 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
636 * except for this special case, we MUST call splz() here to handle any
637 * pending ints, particularly after we switch, or we might accidently
638 * halt the cpu with interrupts pending.
a2a5ad0d 639 */
46a3f46d 640 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 641 splz();
f1d1c3fa
MD
642
643 /*
644 * YYY enabling will cause wakeup() to task-switch, which really
645 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
646 * preemption and MP without actually doing preemption or MP, because a
647 * lot of code assumes that wakeup() does not block.
f1d1c3fa 648 */
46a3f46d
MD
649 if (untimely_switch && td->td_nest_count == 0 &&
650 gd->gd_intr_nesting_level == 0
651 ) {
f1d1c3fa
MD
652 crit_enter();
653 /*
654 * YYY temporary hacks until we disassociate the userland scheduler
655 * from the LWKT scheduler.
656 */
657 if (td->td_flags & TDF_RUNQ) {
658 lwkt_switch(); /* will not reenter yield function */
659 } else {
660 lwkt_schedule_self(); /* make sure we are scheduled */
661 lwkt_switch(); /* will not reenter yield function */
662 lwkt_deschedule_self(); /* make sure we are descheduled */
663 }
7966cb69 664 crit_exit_noyield(td);
f1d1c3fa 665 }
f1d1c3fa
MD
666}
667
8ad65e08 668/*
f1d1c3fa 669 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 670 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
671 * the crit_exit() call in lwkt_switch().
672 *
673 * (self contained on a per cpu basis)
8ad65e08
MD
674 */
675void
f1d1c3fa 676lwkt_yield(void)
8ad65e08 677{
f1d1c3fa
MD
678 lwkt_schedule_self();
679 lwkt_switch();
680}
681
682/*
683 * Schedule a thread to run. As the current thread we can always safely
684 * schedule ourselves, and a shortcut procedure is provided for that
685 * function.
686 *
687 * (non-blocking, self contained on a per cpu basis)
688 */
689void
690lwkt_schedule_self(void)
691{
692 thread_t td = curthread;
693
694 crit_enter();
695 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa 696 _lwkt_enqueue(td);
26a0694b
MD
697 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
698 panic("SCHED SELF PANIC");
f1d1c3fa 699 crit_exit();
8ad65e08 700}
8ad65e08
MD
701
702/*
f1d1c3fa
MD
703 * Generic schedule. Possibly schedule threads belonging to other cpus and
704 * deal with threads that might be blocked on a wait queue.
705 *
96728c05 706 * YYY this is one of the best places to implement load balancing code.
f1d1c3fa
MD
707 * Load balancing can be accomplished by requesting other sorts of actions
708 * for the thread in question.
8ad65e08
MD
709 */
710void
711lwkt_schedule(thread_t td)
712{
96728c05 713#ifdef INVARIANTS
26a0694b
MD
714 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
715 && td->td_proc->p_stat == SSLEEP
716 ) {
717 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
718 curthread,
719 curthread->td_proc ? curthread->td_proc->p_pid : -1,
720 curthread->td_proc ? curthread->td_proc->p_stat : -1,
721 td,
722 td->td_proc ? curthread->td_proc->p_pid : -1,
723 td->td_proc ? curthread->td_proc->p_stat : -1
724 );
725 panic("SCHED PANIC");
726 }
96728c05 727#endif
f1d1c3fa
MD
728 crit_enter();
729 if (td == curthread) {
730 _lwkt_enqueue(td);
731 } else {
732 lwkt_wait_t w;
733
734 /*
735 * If the thread is on a wait list we have to send our scheduling
736 * request to the owner of the wait structure. Otherwise we send
737 * the scheduling request to the cpu owning the thread. Races
738 * are ok, the target will forward the message as necessary (the
739 * message may chase the thread around before it finally gets
740 * acted upon).
741 *
742 * (remember, wait structures use stable storage)
743 */
744 if ((w = td->td_wait) != NULL) {
96728c05 745 if (lwkt_trytoken(&w->wa_token)) {
f1d1c3fa
MD
746 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
747 --w->wa_count;
748 td->td_wait = NULL;
75cdbe6c 749 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 750 _lwkt_enqueue(td);
57c254db 751 if (td->td_preemptable) {
96728c05 752 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
57c254db
MD
753 } else if (_lwkt_wantresched(td, curthread)) {
754 need_resched();
755 }
f1d1c3fa 756 } else {
a72187e9 757 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa 758 }
96728c05 759 lwkt_reltoken(&w->wa_token);
f1d1c3fa 760 } else {
96728c05 761 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
762 }
763 } else {
764 /*
765 * If the wait structure is NULL and we own the thread, there
766 * is no race (since we are in a critical section). If we
767 * do not own the thread there might be a race but the
768 * target cpu will deal with it.
769 */
75cdbe6c 770 if (smp_active == 0 || td->td_gd == mycpu) {
f1d1c3fa 771 _lwkt_enqueue(td);
57c254db 772 if (td->td_preemptable) {
96728c05 773 td->td_preemptable(td, TDPRI_CRIT);
57c254db
MD
774 } else if (_lwkt_wantresched(td, curthread)) {
775 need_resched();
776 }
f1d1c3fa 777 } else {
a72187e9 778 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
779 }
780 }
8ad65e08 781 }
f1d1c3fa 782 crit_exit();
8ad65e08
MD
783}
784
d9eea1a5
MD
785/*
786 * Managed acquisition. This code assumes that the MP lock is held for
787 * the tdallq operation and that the thread has been descheduled from its
788 * original cpu. We also have to wait for the thread to be entirely switched
789 * out on its original cpu (this is usually fast enough that we never loop)
790 * since the LWKT system does not have to hold the MP lock while switching
791 * and the target may have released it before switching.
792 */
a2a5ad0d
MD
793void
794lwkt_acquire(thread_t td)
795{
796 struct globaldata *gd;
797
798 gd = td->td_gd;
799 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
d9eea1a5
MD
800 while (td->td_flags & TDF_RUNNING) /* XXX spin */
801 ;
a2a5ad0d
MD
802 if (gd != mycpu) {
803 crit_enter();
804 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
805 gd = mycpu;
806 td->td_gd = gd;
a2a5ad0d
MD
807 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
808 crit_exit();
809 }
810}
811
8ad65e08 812/*
f1d1c3fa
MD
813 * Deschedule a thread.
814 *
815 * (non-blocking, self contained on a per cpu basis)
816 */
817void
818lwkt_deschedule_self(void)
819{
820 thread_t td = curthread;
821
822 crit_enter();
823 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
f1d1c3fa
MD
824 _lwkt_dequeue(td);
825 crit_exit();
826}
827
828/*
829 * Generic deschedule. Descheduling threads other then your own should be
830 * done only in carefully controlled circumstances. Descheduling is
831 * asynchronous.
832 *
833 * This function may block if the cpu has run out of messages.
8ad65e08
MD
834 */
835void
836lwkt_deschedule(thread_t td)
837{
f1d1c3fa
MD
838 crit_enter();
839 if (td == curthread) {
840 _lwkt_dequeue(td);
841 } else {
a72187e9 842 if (td->td_gd == mycpu) {
f1d1c3fa
MD
843 _lwkt_dequeue(td);
844 } else {
a72187e9 845 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
f1d1c3fa
MD
846 }
847 }
848 crit_exit();
849}
850
4b5f931b
MD
851/*
852 * Set the target thread's priority. This routine does not automatically
853 * switch to a higher priority thread, LWKT threads are not designed for
854 * continuous priority changes. Yield if you want to switch.
855 *
856 * We have to retain the critical section count which uses the high bits
26a0694b
MD
857 * of the td_pri field. The specified priority may also indicate zero or
858 * more critical sections by adding TDPRI_CRIT*N.
4b5f931b
MD
859 */
860void
861lwkt_setpri(thread_t td, int pri)
862{
26a0694b 863 KKASSERT(pri >= 0);
a72187e9 864 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
865 crit_enter();
866 if (td->td_flags & TDF_RUNQ) {
867 _lwkt_dequeue(td);
868 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
869 _lwkt_enqueue(td);
870 } else {
871 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
872 }
873 crit_exit();
874}
875
876void
877lwkt_setpri_self(int pri)
878{
879 thread_t td = curthread;
880
4b5f931b
MD
881 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
882 crit_enter();
883 if (td->td_flags & TDF_RUNQ) {
884 _lwkt_dequeue(td);
885 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
886 _lwkt_enqueue(td);
887 } else {
888 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
889 }
890 crit_exit();
891}
892
893struct proc *
894lwkt_preempted_proc(void)
895{
73e4f7b9 896 thread_t td = curthread;
4b5f931b
MD
897 while (td->td_preempted)
898 td = td->td_preempted;
899 return(td->td_proc);
900}
901
ece04fd0
MD
902typedef struct lwkt_gettoken_req {
903 lwkt_token_t tok;
904 int cpu;
905} lwkt_gettoken_req;
906
907#if 0
4b5f931b 908
f1d1c3fa
MD
909/*
910 * This function deschedules the current thread and blocks on the specified
911 * wait queue. We obtain ownership of the wait queue in order to block
912 * on it. A generation number is used to interlock the wait queue in case
913 * it gets signalled while we are blocked waiting on the token.
914 *
915 * Note: alternatively we could dequeue our thread and then message the
916 * target cpu owning the wait queue. YYY implement as sysctl.
917 *
918 * Note: wait queue signals normally ping-pong the cpu as an optimization.
919 */
96728c05 920
f1d1c3fa 921void
ae8050a4 922lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
f1d1c3fa
MD
923{
924 thread_t td = curthread;
f1d1c3fa 925
f1d1c3fa 926 lwkt_gettoken(&w->wa_token);
ae8050a4 927 if (w->wa_gen == *gen) {
f1d1c3fa
MD
928 _lwkt_dequeue(td);
929 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
930 ++w->wa_count;
931 td->td_wait = w;
ae8050a4 932 td->td_wmesg = wmesg;
ece04fd0 933again:
f1d1c3fa 934 lwkt_switch();
ece04fd0
MD
935 lwkt_regettoken(&w->wa_token);
936 if (td->td_wmesg != NULL) {
937 _lwkt_dequeue(td);
938 goto again;
939 }
8ad65e08 940 }
ae8050a4
MD
941 /* token might be lost, doesn't matter for gen update */
942 *gen = w->wa_gen;
f1d1c3fa
MD
943 lwkt_reltoken(&w->wa_token);
944}
945
946/*
947 * Signal a wait queue. We gain ownership of the wait queue in order to
948 * signal it. Once a thread is removed from the wait queue we have to
949 * deal with the cpu owning the thread.
950 *
951 * Note: alternatively we could message the target cpu owning the wait
952 * queue. YYY implement as sysctl.
953 */
954void
ece04fd0 955lwkt_signal(lwkt_wait_t w, int count)
f1d1c3fa
MD
956{
957 thread_t td;
958 int count;
959
960 lwkt_gettoken(&w->wa_token);
961 ++w->wa_gen;
ece04fd0
MD
962 if (count < 0)
963 count = w->wa_count;
f1d1c3fa
MD
964 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
965 --count;
966 --w->wa_count;
967 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
968 td->td_wait = NULL;
ae8050a4 969 td->td_wmesg = NULL;
a72187e9 970 if (td->td_gd == mycpu) {
f1d1c3fa
MD
971 _lwkt_enqueue(td);
972 } else {
a72187e9 973 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
f1d1c3fa
MD
974 }
975 lwkt_regettoken(&w->wa_token);
976 }
977 lwkt_reltoken(&w->wa_token);
978}
979
ece04fd0
MD
980#endif
981
f1d1c3fa 982/*
96728c05 983 * Acquire ownership of a token
f1d1c3fa 984 *
96728c05 985 * Acquire ownership of a token. The token may have spl and/or critical
f1d1c3fa
MD
986 * section side effects, depending on its purpose. These side effects
987 * guarentee that you will maintain ownership of the token as long as you
988 * do not block. If you block you may lose access to the token (but you
989 * must still release it even if you lose your access to it).
990 *
96728c05 991 * YYY for now we use a critical section to prevent IPIs from taking away
a2a5ad0d 992 * a token, but do we really only need to disable IPIs ?
96728c05
MD
993 *
994 * YYY certain tokens could be made to act like mutexes when performance
995 * would be better (e.g. t_cpu == -1). This is not yet implemented.
996 *
a2a5ad0d
MD
997 * YYY the tokens replace 4.x's simplelocks for the most part, but this
998 * means that 4.x does not expect a switch so for now we cannot switch
999 * when waiting for an IPI to be returned.
1000 *
1001 * YYY If the token is owned by another cpu we may have to send an IPI to
96728c05
MD
1002 * it and then block. The IPI causes the token to be given away to the
1003 * requesting cpu, unless it has already changed hands. Since only the
1004 * current cpu can give away a token it owns we do not need a memory barrier.
a2a5ad0d 1005 * This needs serious optimization.
f1d1c3fa 1006 */
57c254db
MD
1007
1008#ifdef SMP
1009
96728c05
MD
1010static
1011void
1012lwkt_gettoken_remote(void *arg)
1013{
1014 lwkt_gettoken_req *req = arg;
1015 if (req->tok->t_cpu == mycpu->gd_cpuid) {
634081ff 1016#ifdef INVARIANTS
a2a5ad0d
MD
1017 if (token_debug)
1018 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
634081ff 1019#endif
96728c05 1020 req->tok->t_cpu = req->cpu;
a2a5ad0d
MD
1021 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1022 /* else set reqcpu to point to current cpu for release */
96728c05
MD
1023 }
1024}
1025
57c254db
MD
1026#endif
1027
8a8d5d85 1028int
f1d1c3fa
MD
1029lwkt_gettoken(lwkt_token_t tok)
1030{
1031 /*
1032 * Prevent preemption so the token can't be taken away from us once
1033 * we gain ownership of it. Use a synchronous request which might
1034 * block. The request will be forwarded as necessary playing catchup
1035 * to the token.
1036 */
96728c05 1037
f1d1c3fa 1038 crit_enter();
57c254db 1039#ifdef INVARIANTS
1d4c9574
MD
1040 if (curthread->td_pri > 1800) {
1041 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1042 tok, ((int **)&tok)[-1]);
1043 }
a2a5ad0d
MD
1044 if (curthread->td_pri > 2000) {
1045 curthread->td_pri = 1000;
1046 panic("too HIGH!");
57c254db
MD
1047 }
1048#endif
96728c05 1049#ifdef SMP
d0e06f83 1050 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1051 struct lwkt_gettoken_req req;
1052 int seq;
96728c05
MD
1053 int dcpu;
1054
1055 req.cpu = mycpu->gd_cpuid;
1056 req.tok = tok;
1057 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1058 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1059#ifdef INVARIANTS
a2a5ad0d
MD
1060 if (token_debug)
1061 printf("REQT%d ", dcpu);
634081ff 1062#endif
96728c05
MD
1063 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1064 lwkt_wait_ipiq(dcpu, seq);
634081ff 1065#ifdef INVARIANTS
a2a5ad0d
MD
1066 if (token_debug)
1067 printf("REQR%d ", tok->t_cpu);
634081ff 1068#endif
f1d1c3fa
MD
1069 }
1070#endif
1071 /*
1072 * leave us in a critical section on return. This will be undone
8a8d5d85 1073 * by lwkt_reltoken(). Bump the generation number.
f1d1c3fa 1074 */
8a8d5d85 1075 return(++tok->t_gen);
f1d1c3fa
MD
1076}
1077
96728c05
MD
1078/*
1079 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1080 * failure.
1081 */
1082int
1083lwkt_trytoken(lwkt_token_t tok)
1084{
1085 crit_enter();
1086#ifdef SMP
1087 if (tok->t_cpu != mycpu->gd_cpuid) {
a015262c 1088 crit_exit();
96728c05
MD
1089 return(0);
1090 }
1091#endif
1092 /* leave us in the critical section */
1093 ++tok->t_gen;
1094 return(1);
1095}
1096
f1d1c3fa
MD
1097/*
1098 * Release your ownership of a token. Releases must occur in reverse
1099 * order to aquisitions, eventually so priorities can be unwound properly
1100 * like SPLs. At the moment the actual implemention doesn't care.
1101 *
1102 * We can safely hand a token that we own to another cpu without notifying
1103 * it, but once we do we can't get it back without requesting it (unless
1104 * the other cpu hands it back to us before we check).
1105 *
1106 * We might have lost the token, so check that.
7ba9c17c
MD
1107 *
1108 * Return the token's generation number. The number is useful to callers
1109 * who may want to know if the token was stolen during potential blockages.
f1d1c3fa 1110 */
7ba9c17c 1111int
f1d1c3fa
MD
1112lwkt_reltoken(lwkt_token_t tok)
1113{
7ba9c17c
MD
1114 int gen;
1115
d0e06f83 1116 if (tok->t_cpu == mycpu->gd_cpuid) {
f1d1c3fa
MD
1117 tok->t_cpu = tok->t_reqcpu;
1118 }
7ba9c17c 1119 gen = tok->t_gen;
f1d1c3fa 1120 crit_exit();
7ba9c17c 1121 return(gen);
f1d1c3fa
MD
1122}
1123
1124/*
7ba9c17c
MD
1125 * Reacquire a token that might have been lost. 0 is returned if the
1126 * generation has not changed (nobody stole the token from us), -1 is
1127 * returned otherwise. The token is reacquired regardless but the
1128 * generation number is not bumped further if we already own the token.
8a8d5d85 1129 *
7ba9c17c
MD
1130 * For efficiency we inline the best-case situation for lwkt_regettoken()
1131 * (i.e .we still own the token).
8a8d5d85
MD
1132 */
1133int
1134lwkt_gentoken(lwkt_token_t tok, int *gen)
1135{
7ba9c17c 1136 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
8a8d5d85 1137 return(0);
7ba9c17c
MD
1138 *gen = lwkt_regettoken(tok);
1139 return(-1);
8a8d5d85
MD
1140}
1141
8a8d5d85 1142/*
435ff993
MD
1143 * Re-acquire a token that might have been lost. The generation number
1144 * is bumped and returned regardless of whether the token had been lost
1145 * or not (because we only have cpu granularity we have to bump the token
1146 * either way).
f1d1c3fa
MD
1147 */
1148int
1149lwkt_regettoken(lwkt_token_t tok)
1150{
96728c05 1151 /* assert we are in a critical section */
d0e06f83 1152 if (tok->t_cpu != mycpu->gd_cpuid) {
96728c05 1153#ifdef SMP
d0e06f83 1154 while (tok->t_cpu != mycpu->gd_cpuid) {
57c254db
MD
1155 struct lwkt_gettoken_req req;
1156 int seq;
96728c05 1157 int dcpu;
57c254db 1158
96728c05
MD
1159 req.cpu = mycpu->gd_cpuid;
1160 req.tok = tok;
1161 dcpu = (volatile int)tok->t_cpu;
a2a5ad0d 1162 KKASSERT(dcpu >= 0 && dcpu < ncpus);
634081ff 1163#ifdef INVARIANTS
cb973d15
MD
1164 if (token_debug)
1165 printf("REQT%d ", dcpu);
634081ff 1166#endif
96728c05
MD
1167 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1168 lwkt_wait_ipiq(dcpu, seq);
634081ff 1169#ifdef INVARIATNS
cb973d15
MD
1170 if (token_debug)
1171 printf("REQR%d ", tok->t_cpu);
634081ff 1172#endif
f1d1c3fa 1173 }
f1d1c3fa 1174#endif
96728c05 1175 }
435ff993 1176 ++tok->t_gen;
8a8d5d85 1177 return(tok->t_gen);
8ad65e08
MD
1178}
1179
72740893
MD
1180void
1181lwkt_inittoken(lwkt_token_t tok)
1182{
1183 /*
1184 * Zero structure and set cpu owner and reqcpu to cpu 0.
1185 */
1186 bzero(tok, sizeof(*tok));
1187}
1188
99df837e
MD
1189/*
1190 * Create a kernel process/thread/whatever. It shares it's address space
1191 * with proc0 - ie: kernel only.
1192 *
365fa13f
MD
1193 * NOTE! By default new threads are created with the MP lock held. A
1194 * thread which does not require the MP lock should release it by calling
1195 * rel_mplock() at the start of the new thread.
99df837e
MD
1196 */
1197int
1198lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1199 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1200 const char *fmt, ...)
99df837e 1201{
73e4f7b9 1202 thread_t td;
e2565a42 1203 __va_list ap;
99df837e 1204
75cdbe6c 1205 td = lwkt_alloc_thread(template, cpu);
a2a5ad0d
MD
1206 if (tdp)
1207 *tdp = td;
99df837e 1208 cpu_set_thread_handler(td, kthread_exit, func, arg);
ef0fdad1 1209 td->td_flags |= TDF_VERBOSE | tdflags;
8a8d5d85
MD
1210#ifdef SMP
1211 td->td_mpcount = 1;
1212#endif
99df837e
MD
1213
1214 /*
1215 * Set up arg0 for 'ps' etc
1216 */
e2565a42 1217 __va_start(ap, fmt);
99df837e 1218 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1219 __va_end(ap);
99df837e
MD
1220
1221 /*
1222 * Schedule the thread to run
1223 */
ef0fdad1
MD
1224 if ((td->td_flags & TDF_STOPREQ) == 0)
1225 lwkt_schedule(td);
1226 else
1227 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1228 return 0;
1229}
1230
1231/*
1232 * Destroy an LWKT thread. Warning! This function is not called when
1233 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1234 * uses a different reaping mechanism.
1235 */
1236void
1237lwkt_exit(void)
1238{
1239 thread_t td = curthread;
1240
1241 if (td->td_flags & TDF_VERBOSE)
1242 printf("kthread %p %s has exited\n", td, td->td_comm);
1243 crit_enter();
1244 lwkt_deschedule_self();
1245 ++mycpu->gd_tdfreecount;
1246 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1247 cpu_thread_exit();
1248}
1249
1250/*
1251 * Create a kernel process/thread/whatever. It shares it's address space
ef0fdad1 1252 * with proc0 - ie: kernel only. 5.x compatible.
365fa13f
MD
1253 *
1254 * NOTE! By default kthreads are created with the MP lock held. A
1255 * thread which does not require the MP lock should release it by calling
1256 * rel_mplock() at the start of the new thread.
99df837e
MD
1257 */
1258int
1259kthread_create(void (*func)(void *), void *arg,
1260 struct thread **tdp, const char *fmt, ...)
1261{
73e4f7b9 1262 thread_t td;
e2565a42 1263 __va_list ap;
99df837e 1264
75cdbe6c 1265 td = lwkt_alloc_thread(NULL, -1);
a2a5ad0d
MD
1266 if (tdp)
1267 *tdp = td;
99df837e
MD
1268 cpu_set_thread_handler(td, kthread_exit, func, arg);
1269 td->td_flags |= TDF_VERBOSE;
8a8d5d85
MD
1270#ifdef SMP
1271 td->td_mpcount = 1;
1272#endif
99df837e
MD
1273
1274 /*
1275 * Set up arg0 for 'ps' etc
1276 */
e2565a42 1277 __va_start(ap, fmt);
99df837e 1278 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1279 __va_end(ap);
99df837e
MD
1280
1281 /*
1282 * Schedule the thread to run
1283 */
1284 lwkt_schedule(td);
1285 return 0;
1286}
1287
26a0694b
MD
1288void
1289crit_panic(void)
1290{
73e4f7b9 1291 thread_t td = curthread;
26a0694b
MD
1292 int lpri = td->td_pri;
1293
1294 td->td_pri = 0;
1295 panic("td_pri is/would-go negative! %p %d", td, lpri);
1296}
1297
99df837e
MD
1298/*
1299 * Destroy an LWKT thread. Warning! This function is not called when
1300 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1301 * uses a different reaping mechanism.
1302 *
1303 * XXX duplicates lwkt_exit()
1304 */
1305void
1306kthread_exit(void)
1307{
1308 lwkt_exit();
1309}
1310
96728c05
MD
1311#ifdef SMP
1312
1313/*
1314 * Send a function execution request to another cpu. The request is queued
1315 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1316 * possible target cpu. The FIFO can be written.
1317 *
1318 * YYY If the FIFO fills up we have to enable interrupts and process the
1319 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1320 * Create a CPU_*() function to do this!
1321 *
46a3f46d
MD
1322 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1323 * end will take care of any pending interrupts.
1324 *
96728c05
MD
1325 * Must be called from a critical section.
1326 */
1327int
1328lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1329{
1330 lwkt_ipiq_t ip;
1331 int windex;
a2a5ad0d 1332 struct globaldata *gd = mycpu;
96728c05 1333
a2a5ad0d 1334 if (dcpu == gd->gd_cpuid) {
96728c05
MD
1335 func(arg);
1336 return(0);
1337 }
cb973d15 1338 crit_enter();
a2a5ad0d
MD
1339 ++gd->gd_intr_nesting_level;
1340#ifdef INVARIANTS
1341 if (gd->gd_intr_nesting_level > 20)
1342 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1343#endif
96728c05
MD
1344 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1345 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1346 ++ipiq_count;
a2a5ad0d 1347 ip = &gd->gd_ipiq[dcpu];
cb973d15
MD
1348
1349 /*
1350 * We always drain before the FIFO becomes full so it should never
1351 * become full. We need to leave enough entries to deal with
1352 * reentrancy.
1353 */
1354 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1355 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1356 ip->ip_func[windex] = func;
1357 ip->ip_arg[windex] = arg;
1358 /* YYY memory barrier */
1359 ++ip->ip_windex;
96728c05
MD
1360 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1361 unsigned int eflags = read_eflags();
1362 cpu_enable_intr();
1363 ++ipiq_fifofull;
cb973d15 1364 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
96728c05
MD
1365 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1366 lwkt_process_ipiq();
1367 }
1368 write_eflags(eflags);
1369 }
a2a5ad0d 1370 --gd->gd_intr_nesting_level;
96728c05 1371 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
cb973d15 1372 crit_exit();
96728c05
MD
1373 return(ip->ip_windex);
1374}
1375
cb973d15
MD
1376/*
1377 * Send a message to several target cpus. Typically used for scheduling.
435ff993 1378 * The message will not be sent to stopped cpus.
cb973d15
MD
1379 */
1380void
1381lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1382{
1383 int cpuid;
1384
435ff993 1385 mask &= ~stopped_cpus;
cb973d15
MD
1386 while (mask) {
1387 cpuid = bsfl(mask);
1388 lwkt_send_ipiq(cpuid, func, arg);
1389 mask &= ~(1 << cpuid);
1390 }
1391}
1392
96728c05
MD
1393/*
1394 * Wait for the remote cpu to finish processing a function.
1395 *
1396 * YYY we have to enable interrupts and process the IPIQ while waiting
1397 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1398 * function to do this! YYY we really should 'block' here.
1399 *
1400 * Must be called from a critical section. Thsi routine may be called
1401 * from an interrupt (for example, if an interrupt wakes a foreign thread
1402 * up).
1403 */
1404void
1405lwkt_wait_ipiq(int dcpu, int seq)
1406{
1407 lwkt_ipiq_t ip;
a2a5ad0d 1408 int maxc = 100000000;
96728c05
MD
1409
1410 if (dcpu != mycpu->gd_cpuid) {
1411 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1412 ip = &mycpu->gd_ipiq[dcpu];
cb973d15 1413 if ((int)(ip->ip_xindex - seq) < 0) {
96728c05
MD
1414 unsigned int eflags = read_eflags();
1415 cpu_enable_intr();
cb973d15 1416 while ((int)(ip->ip_xindex - seq) < 0) {
96728c05 1417 lwkt_process_ipiq();
a2a5ad0d 1418 if (--maxc == 0)
cb973d15 1419 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
a2a5ad0d
MD
1420 if (maxc < -1000000)
1421 panic("LWKT_WAIT_IPIQ");
96728c05
MD
1422 }
1423 write_eflags(eflags);
1424 }
1425 }
1426}
1427
1428/*
1429 * Called from IPI interrupt (like a fast interrupt), which has placed
1430 * us in a critical section. The MP lock may or may not be held.
cb973d15
MD
1431 * May also be called from doreti or splz, or be reentrantly called
1432 * indirectly through the ip_func[] we run.
96728c05
MD
1433 */
1434void
1435lwkt_process_ipiq(void)
1436{
1437 int n;
1438 int cpuid = mycpu->gd_cpuid;
1439
1440 for (n = 0; n < ncpus; ++n) {
1441 lwkt_ipiq_t ip;
1442 int ri;
1443
1444 if (n == cpuid)
1445 continue;
1446 ip = globaldata_find(n)->gd_ipiq;
1447 if (ip == NULL)
1448 continue;
1449 ip = &ip[cpuid];
cb973d15
MD
1450
1451 /*
1452 * Note: xindex is only updated after we are sure the function has
1453 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1454 * function may send an IPI which may block/drain.
1455 */
96728c05
MD
1456 while (ip->ip_rindex != ip->ip_windex) {
1457 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
96728c05 1458 ++ip->ip_rindex;
cb973d15
MD
1459 ip->ip_func[ri](ip->ip_arg[ri]);
1460 /* YYY memory barrier */
1461 ip->ip_xindex = ip->ip_rindex;
96728c05
MD
1462 }
1463 }
1464}
1465
1466#else
1467
1468int
1469lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1470{
1471 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1472 return(0); /* NOT REACHED */
1473}
1474
1475void
1476lwkt_wait_ipiq(int dcpu, int seq)
1477{
1478 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);
1479}
1480
1481#endif