alist debug mode - report unrounded byte usage.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
8c10bfcf 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
8c10bfcf 33 *
b402c633 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.109 2007/07/02 01:37:08 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
05220613
MD
44#ifdef _KERNEL
45
8ad65e08
MD
46#include <sys/param.h>
47#include <sys/systm.h>
48#include <sys/kernel.h>
49#include <sys/proc.h>
50#include <sys/rtprio.h>
51#include <sys/queue.h>
7d0bac62 52#include <sys/sysctl.h>
99df837e 53#include <sys/kthread.h>
f1d1c3fa 54#include <machine/cpu.h>
99df837e 55#include <sys/lock.h>
f6bf3af1 56#include <sys/caps.h>
9d265729 57#include <sys/spinlock.h>
57aa743c 58#include <sys/ktr.h>
9d265729
MD
59
60#include <sys/thread2.h>
61#include <sys/spinlock2.h>
f1d1c3fa 62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
71#include <vm/vm_zone.h>
72
99df837e 73#include <machine/stdarg.h>
96728c05 74#include <machine/smp.h>
99df837e 75
05220613
MD
76#else
77
78#include <sys/stdint.h>
fb04f4fd 79#include <libcaps/thread.h>
05220613
MD
80#include <sys/thread.h>
81#include <sys/msgport.h>
82#include <sys/errno.h>
fb04f4fd 83#include <libcaps/globaldata.h>
7e8303ad 84#include <machine/cpufunc.h>
05220613
MD
85#include <sys/thread2.h>
86#include <sys/msgport2.h>
709799ea 87#include <stdio.h>
05220613 88#include <stdlib.h>
709799ea 89#include <string.h>
709799ea 90#include <machine/lock.h>
b950d4b8
EN
91#include <machine/atomic.h>
92#include <machine/cpu.h>
05220613
MD
93
94#endif
95
7d0bac62 96static int untimely_switch = 0;
0f7a3396
MD
97#ifdef INVARIANTS
98static int panic_on_cscount = 0;
99#endif
05220613
MD
100static __int64_t switch_count = 0;
101static __int64_t preempt_hit = 0;
102static __int64_t preempt_miss = 0;
103static __int64_t preempt_weird = 0;
38717797
HP
104static __int64_t token_contention_count = 0;
105static __int64_t mplock_contention_count = 0;
fb0f29c4 106static int lwkt_use_spin_port;
05220613
MD
107
108#ifdef _KERNEL
109
fb0f29c4
MD
110/*
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
113 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
05220613 116SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
117#ifdef INVARIANTS
118SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
119#endif
4b5f931b 120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 121SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
124#ifdef INVARIANTS
125SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
126 &token_contention_count, 0, "spinning due to token contention");
127SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
128 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
129#endif
05220613
MD
130#endif
131
57aa743c
MD
132/*
133 * Kernel Trace
134 */
135#ifdef _KERNEL
136
137#if !defined(KTR_GIANT_CONTENTION)
138#define KTR_GIANT_CONTENTION KTR_ALL
139#endif
140
141KTR_INFO_MASTER(giant);
142KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
143KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
144
145#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
146
147#endif
148
4b5f931b
MD
149/*
150 * These helper procedures handle the runq, they can only be called from
151 * within a critical section.
75cdbe6c
MD
152 *
153 * WARNING! Prior to SMP being brought up it is possible to enqueue and
154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
155 * instead of 'mycpu' when referencing the globaldata structure. Once
156 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 157 */
f1d1c3fa
MD
158static __inline
159void
160_lwkt_dequeue(thread_t td)
161{
162 if (td->td_flags & TDF_RUNQ) {
4b5f931b 163 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 164 struct globaldata *gd = td->td_gd;
4b5f931b 165
f1d1c3fa 166 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
168 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
169 }
170}
171
172static __inline
173void
174_lwkt_enqueue(thread_t td)
175{
344ad853 176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 177 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 178 struct globaldata *gd = td->td_gd;
4b5f931b 179
f1d1c3fa 180 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
182 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
183 }
184}
8ad65e08 185
37af14fe
MD
186/*
187 * Schedule a thread to run. As the current thread we can always safely
188 * schedule ourselves, and a shortcut procedure is provided for that
189 * function.
190 *
191 * (non-blocking, self contained on a per cpu basis)
192 */
193void
194lwkt_schedule_self(thread_t td)
195{
196 crit_enter_quick(td);
37af14fe 197 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 198 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 199 _lwkt_enqueue(td);
37af14fe
MD
200 crit_exit_quick(td);
201}
202
203/*
204 * Deschedule a thread.
205 *
206 * (non-blocking, self contained on a per cpu basis)
207 */
208void
209lwkt_deschedule_self(thread_t td)
210{
211 crit_enter_quick(td);
37af14fe
MD
212 _lwkt_dequeue(td);
213 crit_exit_quick(td);
214}
215
2d93b37a
MD
216#ifdef _KERNEL
217
8ad65e08
MD
218/*
219 * LWKTs operate on a per-cpu basis
220 *
73e4f7b9 221 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
222 */
223void
224lwkt_gdinit(struct globaldata *gd)
225{
4b5f931b
MD
226 int i;
227
228 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
229 TAILQ_INIT(&gd->gd_tdrunq[i]);
230 gd->gd_runqmask = 0;
73e4f7b9 231 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
232}
233
2d93b37a
MD
234#endif /* _KERNEL */
235
7d0bac62
MD
236/*
237 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
238 * or LWKT start address before it can be scheduled. If the target cpu is
239 * -1 the thread will be created on the current cpu.
0cfcada1
MD
240 *
241 * If you intend to create a thread without a process context this function
242 * does everything except load the startup and switcher function.
7d0bac62
MD
243 */
244thread_t
d3d32139 245lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 246{
99df837e 247 void *stack;
37af14fe 248 globaldata_t gd = mycpu;
7d0bac62 249
ef0fdad1 250 if (td == NULL) {
37af14fe
MD
251 crit_enter_gd(gd);
252 if (gd->gd_tdfreecount > 0) {
253 --gd->gd_tdfreecount;
254 td = TAILQ_FIRST(&gd->gd_tdfreeq);
d9eea1a5 255 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
ef0fdad1 256 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
37af14fe
MD
257 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
258 crit_exit_gd(gd);
d3d32139 259 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
ef0fdad1 260 } else {
37af14fe 261 crit_exit_gd(gd);
05220613 262#ifdef _KERNEL
ef0fdad1 263 td = zalloc(thread_zone);
05220613
MD
264#else
265 td = malloc(sizeof(struct thread));
266#endif
ef0fdad1 267 td->td_kstack = NULL;
f470d0c8 268 td->td_kstack_size = 0;
ef0fdad1
MD
269 flags |= TDF_ALLOCATED_THREAD;
270 }
271 }
f470d0c8
MD
272 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
273 if (flags & TDF_ALLOCATED_STACK) {
a9e5ae82 274#ifdef _KERNEL
e4846942 275 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
a9e5ae82
JS
276#else
277 libcaps_free_stack(stack, td->td_kstack_size);
278#endif
f470d0c8
MD
279 stack = NULL;
280 }
281 }
282 if (stack == NULL) {
05220613 283#ifdef _KERNEL
e4846942 284 stack = (void *)kmem_alloc(&kernel_map, stksize);
05220613 285#else
f470d0c8 286 stack = libcaps_alloc_stack(stksize);
05220613 287#endif
ef0fdad1 288 flags |= TDF_ALLOCATED_STACK;
99df837e 289 }
75cdbe6c 290 if (cpu < 0)
f470d0c8 291 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 292 else
f470d0c8 293 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 294 return(td);
7d0bac62
MD
295}
296
709799ea
MD
297#ifdef _KERNEL
298
7d0bac62
MD
299/*
300 * Initialize a preexisting thread structure. This function is used by
301 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
302 *
f8c3996b
MD
303 * All threads start out in a critical section at a priority of
304 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
305 * appropriate. This function may send an IPI message when the
306 * requested cpu is not the current cpu and consequently gd_tdallq may
307 * not be initialized synchronously from the point of view of the originating
308 * cpu.
309 *
310 * NOTE! we have to be careful in regards to creating threads for other cpus
311 * if SMP has not yet been activated.
7d0bac62 312 */
41a01a4d
MD
313#ifdef SMP
314
75cdbe6c
MD
315static void
316lwkt_init_thread_remote(void *arg)
317{
318 thread_t td = arg;
319
52eedfb5
MD
320 /*
321 * Protected by critical section held by IPI dispatch
322 */
75cdbe6c
MD
323 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
324}
325
41a01a4d
MD
326#endif
327
7d0bac62 328void
f470d0c8
MD
329lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
330 struct globaldata *gd)
7d0bac62 331{
37af14fe
MD
332 globaldata_t mygd = mycpu;
333
99df837e
MD
334 bzero(td, sizeof(struct thread));
335 td->td_kstack = stack;
f470d0c8 336 td->td_kstack_size = stksize;
d3d32139 337 td->td_flags = flags;
26a0694b 338 td->td_gd = gd;
f8c3996b 339 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
340#ifdef SMP
341 if ((flags & TDF_MPSAFE) == 0)
342 td->td_mpcount = 1;
343#endif
fb0f29c4
MD
344 if (lwkt_use_spin_port)
345 lwkt_initport_spin(&td->td_msgport);
346 else
347 lwkt_initport_thread(&td->td_msgport, td);
99df837e 348 pmap_init_thread(td);
0f7a3396 349#ifdef SMP
5d21b981
MD
350 /*
351 * Normally initializing a thread for a remote cpu requires sending an
352 * IPI. However, the idlethread is setup before the other cpus are
353 * activated so we have to treat it as a special case. XXX manipulation
354 * of gd_tdallq requires the BGL.
355 */
356 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 357 crit_enter_gd(mygd);
75cdbe6c 358 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 359 crit_exit_gd(mygd);
75cdbe6c 360 } else {
2db3b277 361 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 362 }
0f7a3396 363#else
37af14fe 364 crit_enter_gd(mygd);
0f7a3396 365 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 366 crit_exit_gd(mygd);
0f7a3396 367#endif
73e4f7b9
MD
368}
369
2d93b37a
MD
370#endif /* _KERNEL */
371
73e4f7b9
MD
372void
373lwkt_set_comm(thread_t td, const char *ctl, ...)
374{
e2565a42 375 __va_list va;
73e4f7b9 376
e2565a42 377 __va_start(va, ctl);
379210cb 378 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 379 __va_end(va);
7d0bac62
MD
380}
381
99df837e 382void
73e4f7b9 383lwkt_hold(thread_t td)
99df837e 384{
73e4f7b9
MD
385 ++td->td_refs;
386}
387
388void
389lwkt_rele(thread_t td)
390{
391 KKASSERT(td->td_refs > 0);
392 --td->td_refs;
393}
394
c95cd171
MD
395#ifdef _KERNEL
396
73e4f7b9
MD
397void
398lwkt_wait_free(thread_t td)
399{
400 while (td->td_refs)
377d4740 401 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
402}
403
c95cd171
MD
404#endif
405
73e4f7b9
MD
406void
407lwkt_free_thread(thread_t td)
408{
409 struct globaldata *gd = mycpu;
410
d9eea1a5 411 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
412 ("lwkt_free_thread: did not exit! %p", td));
413
37af14fe 414 crit_enter_gd(gd);
73e4f7b9 415 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
99df837e
MD
416 (td->td_flags & TDF_ALLOCATED_THREAD)
417 ) {
73e4f7b9
MD
418 ++gd->gd_tdfreecount;
419 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
37af14fe 420 crit_exit_gd(gd);
99df837e 421 } else {
37af14fe 422 crit_exit_gd(gd);
99df837e 423 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
05220613 424#ifdef _KERNEL
e4846942 425 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
05220613 426#else
f470d0c8 427 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
05220613 428#endif
73e4f7b9 429 /* gd invalid */
99df837e 430 td->td_kstack = NULL;
f470d0c8 431 td->td_kstack_size = 0;
99df837e 432 }
05220613
MD
433 if (td->td_flags & TDF_ALLOCATED_THREAD) {
434#ifdef _KERNEL
99df837e 435 zfree(thread_zone, td);
05220613
MD
436#else
437 free(td);
438#endif
439 }
99df837e
MD
440 }
441}
442
443
8ad65e08
MD
444/*
445 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
446 * switch to the idlethread. Switching must occur within a critical
447 * section to avoid races with the scheduling queue.
448 *
449 * We always have full control over our cpu's run queue. Other cpus
450 * that wish to manipulate our queue must use the cpu_*msg() calls to
451 * talk to our cpu, so a critical section is all that is needed and
452 * the result is very, very fast thread switching.
453 *
96728c05
MD
454 * The LWKT scheduler uses a fixed priority model and round-robins at
455 * each priority level. User process scheduling is a totally
456 * different beast and LWKT priorities should not be confused with
457 * user process priorities.
f1d1c3fa 458 *
96728c05
MD
459 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
460 * cleans it up. Note that the td_switch() function cannot do anything that
461 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
462 * the target thread (not the current thread). It's nice to have a scheduler
463 * that does not need the MP lock to work because it allows us to do some
464 * really cool high-performance MP lock optimizations.
69d78e99
MD
465 *
466 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
467 * is not called by the current thread in the preemption case, only when
468 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
469 */
470void
471lwkt_switch(void)
472{
37af14fe
MD
473 globaldata_t gd = mycpu;
474 thread_t td = gd->gd_curthread;
8ad65e08 475 thread_t ntd;
8a8d5d85
MD
476#ifdef SMP
477 int mpheld;
478#endif
8ad65e08 479
46a3f46d 480 /*
27e88a6e
MD
481 * Switching from within a 'fast' (non thread switched) interrupt or IPI
482 * is illegal. However, we may have to do it anyway if we hit a fatal
483 * kernel trap or we have paniced.
484 *
485 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 486 */
27e88a6e
MD
487 if (gd->gd_intr_nesting_level) {
488 int savegdnest;
489 int savegdtrap;
490
491 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
492 panic("lwkt_switch: cannot switch from within "
493 "a fast interrupt, yet, td %p\n", td);
494 } else {
495 savegdnest = gd->gd_intr_nesting_level;
496 savegdtrap = gd->gd_trap_nesting_level;
497 gd->gd_intr_nesting_level = 0;
498 gd->gd_trap_nesting_level = 0;
a7422615
MD
499 if ((td->td_flags & TDF_PANICWARN) == 0) {
500 td->td_flags |= TDF_PANICWARN;
6ea70f76 501 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615
MD
502 "thread %p (%s)\n", td, td->td_comm);
503#ifdef DDB
504 db_print_backtrace();
505#endif
506 }
27e88a6e
MD
507 lwkt_switch();
508 gd->gd_intr_nesting_level = savegdnest;
509 gd->gd_trap_nesting_level = savegdtrap;
510 return;
511 }
96728c05 512 }
ef0fdad1 513
cb973d15
MD
514 /*
515 * Passive release (used to transition from user to kernel mode
516 * when we block or switch rather then when we enter the kernel).
517 * This function is NOT called if we are switching into a preemption
518 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
519 * our current process designation (if we have one) and become a true
520 * LWKT thread, and may also hand the current process designation to
521 * another process and schedule thread.
cb973d15
MD
522 */
523 if (td->td_release)
524 td->td_release(td);
525
37af14fe 526 crit_enter_gd(gd);
dd55d707 527#ifdef SMP
9d265729
MD
528 if (td->td_toks)
529 lwkt_relalltokens(td);
dd55d707 530#endif
9d265729
MD
531
532 /*
b02926de
MD
533 * We had better not be holding any spin locks, but don't get into an
534 * endless panic loop.
9d265729 535 */
bbb31c5d
MD
536 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
537 ("lwkt_switch: still holding a shared spinlock %p!",
538 gd->gd_spinlock_rd));
d666840a
MD
539 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
540 ("lwkt_switch: still holding %d exclusive spinlocks!",
541 gd->gd_spinlocks_wr));
9d265729 542
8a8d5d85
MD
543
544#ifdef SMP
545 /*
546 * td_mpcount cannot be used to determine if we currently hold the
547 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
548 * to get the lock, and switch out if it can't. Our ownership of
549 * the actual lock will remain stable while we are in a critical section
550 * (but, of course, another cpu may own or release the lock so the
551 * actual value of mp_lock is not stable).
8a8d5d85
MD
552 */
553 mpheld = MP_LOCK_HELD();
0f7a3396
MD
554#ifdef INVARIANTS
555 if (td->td_cscount) {
6ea70f76 556 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
557 td);
558 if (panic_on_cscount)
559 panic("switching while mastering cpusync");
560 }
561#endif
8a8d5d85 562#endif
99df837e
MD
563 if ((ntd = td->td_preempted) != NULL) {
564 /*
565 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
566 * thread. This occurs transparently, whether the preempted thread
567 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
568 * itself).
569 *
570 * We have to setup the MP lock for the original thread after backing
571 * out the adjustment that was made to curthread when the original
572 * was preempted.
99df837e 573 */
26a0694b 574 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 575#ifdef SMP
96728c05 576 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 577 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
578 td, ntd, td->td_mpcount, ntd->td_mpcount);
579 }
8a8d5d85
MD
580 if (ntd->td_mpcount) {
581 td->td_mpcount -= ntd->td_mpcount;
582 KKASSERT(td->td_mpcount >= 0);
583 }
584#endif
26a0694b 585 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
586
587 /*
588 * XXX. The interrupt may have woken a thread up, we need to properly
589 * set the reschedule flag if the originally interrupted thread is at
590 * a lower priority.
591 */
592 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
593 need_lwkt_resched();
8a8d5d85 594 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 595 } else {
4b5f931b
MD
596 /*
597 * Priority queue / round-robin at each priority. Note that user
598 * processes run at a fixed, low priority and the user process
599 * scheduler deals with interactions between user processes
600 * by scheduling and descheduling them from the LWKT queue as
601 * necessary.
8a8d5d85
MD
602 *
603 * We have to adjust the MP lock for the target thread. If we
604 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
605 * thread that does not need the MP lock. If we cannot, we spin
606 * instead of HLT.
607 *
608 * A similar issue exists for the tokens held by the target thread.
609 * If we cannot obtain ownership of the tokens we cannot immediately
610 * schedule the thread.
611 */
612
8ec60c3f
MD
613 /*
614 * If an LWKT reschedule was requested, well that is what we are
615 * doing now so clear it.
616 */
617 clear_lwkt_resched();
4b5f931b
MD
618again:
619 if (gd->gd_runqmask) {
620 int nq = bsrl(gd->gd_runqmask);
621 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
622 gd->gd_runqmask &= ~(1 << nq);
623 goto again;
624 }
8a8d5d85 625#ifdef SMP
41a01a4d 626 /*
df6b8ba0
MD
627 * THREAD SELECTION FOR AN SMP MACHINE BUILD
628 *
41a01a4d
MD
629 * If the target needs the MP lock and we couldn't get it,
630 * or if the target is holding tokens and we could not
631 * gain ownership of the tokens, continue looking for a
632 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
633 *
634 * NOTE: the mpheld variable invalid after this conditional, it
635 * can change due to both cpu_try_mplock() returning success
9d265729 636 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
637 * we are trying to check the mpcount of a thread other then
638 * the current thread. Because of this, if the current thread
639 * is not holding td_mpcount, an IPI indirectly run via
9d265729 640 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 641 * cause the core MP lock to be released.
41a01a4d
MD
642 */
643 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 644 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 645 ) {
8a8d5d85 646 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
647
648 mpheld = MP_LOCK_HELD();
649 ntd = NULL;
8a8d5d85
MD
650 while (rqmask) {
651 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 652 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 653 /* spinning due to MP lock being held */
38717797 654#ifdef INVARIANTS
a453459d 655 ++mplock_contention_count;
38717797 656#endif
a453459d 657 /* mplock still not held, 'mpheld' still valid */
41a01a4d 658 continue;
38717797 659 }
a453459d
MD
660
661 /*
9d265729 662 * mpheld state invalid after getalltokens call returns
a453459d
MD
663 * failure, but the variable is only needed for
664 * the loop.
665 */
9d265729 666 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 667 /* spinning due to token contention */
38717797 668#ifdef INVARIANTS
a453459d 669 ++token_contention_count;
38717797 670#endif
a453459d 671 mpheld = MP_LOCK_HELD();
41a01a4d 672 continue;
38717797 673 }
41a01a4d 674 break;
8a8d5d85
MD
675 }
676 if (ntd)
677 break;
678 rqmask &= ~(1 << nq);
679 nq = bsrl(rqmask);
680 }
681 if (ntd == NULL) {
b402c633 682 cpu_mplock_contested();
a2a5ad0d
MD
683 ntd = &gd->gd_idlethread;
684 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 685 goto using_idle_thread;
8a8d5d85 686 } else {
344ad853 687 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
688 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
689 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
690 }
691 } else {
344ad853 692 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
693 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
694 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
695 }
696#else
df6b8ba0
MD
697 /*
698 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
699 * worry about tokens or the BGL.
700 */
344ad853 701 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
702 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 704#endif
4b5f931b 705 } else {
3c23a41a 706 /*
60f945af
MD
707 * We have nothing to run but only let the idle loop halt
708 * the cpu if there are no pending interrupts.
3c23a41a 709 */
a2a5ad0d 710 ntd = &gd->gd_idlethread;
60f945af 711 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 712 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 713#ifdef SMP
df6b8ba0
MD
714using_idle_thread:
715 /*
716 * The idle thread should not be holding the MP lock unless we
717 * are trapping in the kernel or in a panic. Since we select the
718 * idle thread unconditionally when no other thread is available,
719 * if the MP lock is desired during a panic or kernel trap, we
720 * have to loop in the scheduler until we get it.
721 */
722 if (ntd->td_mpcount) {
723 mpheld = MP_LOCK_HELD();
b402c633 724 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 725 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
726 } else if (mpheld == 0) {
727 cpu_mplock_contested();
df6b8ba0 728 goto again;
b402c633 729 }
df6b8ba0 730 }
a453459d 731#endif
4b5f931b 732 }
f1d1c3fa 733 }
26a0694b
MD
734 KASSERT(ntd->td_pri >= TDPRI_CRIT,
735 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
736
737 /*
738 * Do the actual switch. If the new target does not need the MP lock
739 * and we are holding it, release the MP lock. If the new target requires
740 * the MP lock we have already acquired it for the target.
741 */
742#ifdef SMP
743 if (ntd->td_mpcount == 0 ) {
744 if (MP_LOCK_HELD())
745 cpu_rel_mplock();
746 } else {
a453459d 747 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
748 }
749#endif
94f6d86e
MD
750 if (td != ntd) {
751 ++switch_count;
f1d1c3fa 752 td->td_switch(ntd);
94f6d86e 753 }
37af14fe
MD
754 /* NOTE: current cpu may have changed after switch */
755 crit_exit_quick(td);
8ad65e08
MD
756}
757
b68b7282 758/*
96728c05
MD
759 * Request that the target thread preempt the current thread. Preemption
760 * only works under a specific set of conditions:
b68b7282 761 *
96728c05
MD
762 * - We are not preempting ourselves
763 * - The target thread is owned by the current cpu
764 * - We are not currently being preempted
765 * - The target is not currently being preempted
766 * - We are able to satisfy the target's MP lock requirements (if any).
767 *
768 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
769 * this is called via lwkt_schedule() through the td_preemptable callback.
770 * critpri is the managed critical priority that we should ignore in order
771 * to determine whether preemption is possible (aka usually just the crit
772 * priority of lwkt_schedule() itself).
b68b7282 773 *
26a0694b
MD
774 * XXX at the moment we run the target thread in a critical section during
775 * the preemption in order to prevent the target from taking interrupts
776 * that *WE* can't. Preemption is strictly limited to interrupt threads
777 * and interrupt-like threads, outside of a critical section, and the
778 * preempted source thread will be resumed the instant the target blocks
779 * whether or not the source is scheduled (i.e. preemption is supposed to
780 * be as transparent as possible).
4b5f931b 781 *
8a8d5d85
MD
782 * The target thread inherits our MP count (added to its own) for the
783 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
784 * MP lock during the preemption. Therefore, any preempting targets must be
785 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
786 * out of sync with the physical mp_lock, but we do not have to preserve
787 * the original ownership of the lock if it was out of synch (that is, we
788 * can leave it synchronized on return).
b68b7282
MD
789 */
790void
96728c05 791lwkt_preempt(thread_t ntd, int critpri)
b68b7282 792{
46a3f46d 793 struct globaldata *gd = mycpu;
0a3f9b47 794 thread_t td;
8a8d5d85
MD
795#ifdef SMP
796 int mpheld;
57c254db 797 int savecnt;
8a8d5d85 798#endif
b68b7282 799
26a0694b 800 /*
96728c05
MD
801 * The caller has put us in a critical section. We can only preempt
802 * if the caller of the caller was not in a critical section (basically
d666840a
MD
803 * a local interrupt), as determined by the 'critpri' parameter. We
804 * also acn't preempt if the caller is holding any spinlocks (even if
805 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
806 *
807 * YYY The target thread must be in a critical section (else it must
808 * inherit our critical section? I dunno yet).
41a01a4d 809 *
0a3f9b47 810 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
811 */
812 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 813
0a3f9b47 814 td = gd->gd_curthread;
0a3f9b47 815 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
816 ++preempt_miss;
817 return;
818 }
96728c05
MD
819 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
820 ++preempt_miss;
8ec60c3f 821 need_lwkt_resched();
96728c05
MD
822 return;
823 }
824#ifdef SMP
46a3f46d 825 if (ntd->td_gd != gd) {
96728c05 826 ++preempt_miss;
8ec60c3f 827 need_lwkt_resched();
96728c05
MD
828 return;
829 }
830#endif
41a01a4d
MD
831 /*
832 * Take the easy way out and do not preempt if the target is holding
d666840a 833 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
834 * preempted interlock against the target thread's tokens and whether
835 * we can get all the target thread's tokens, but this situation
836 * should not occur very often so its easier to simply not preempt.
d666840a
MD
837 * Also, plain spinlocks are impossible to figure out at this point so
838 * just don't preempt.
41a01a4d 839 */
bbb31c5d 840 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 841 ++preempt_miss;
8ec60c3f 842 need_lwkt_resched();
41a01a4d
MD
843 return;
844 }
26a0694b
MD
845 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
846 ++preempt_weird;
8ec60c3f 847 need_lwkt_resched();
26a0694b
MD
848 return;
849 }
850 if (ntd->td_preempted) {
4b5f931b 851 ++preempt_hit;
8ec60c3f 852 need_lwkt_resched();
26a0694b 853 return;
b68b7282 854 }
8a8d5d85 855#ifdef SMP
a2a5ad0d
MD
856 /*
857 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
858 * to or from the MP lock. In this case td_mpcount will be pre-disposed
859 * (non-zero) but not actually synchronized with the actual state of the
860 * lock. We can use it to imply an MP lock requirement for the
861 * preemption but we cannot use it to test whether we hold the MP lock
862 * or not.
a2a5ad0d 863 */
96728c05 864 savecnt = td->td_mpcount;
71ef2f5c 865 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
866 ntd->td_mpcount += td->td_mpcount;
867 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
868 ntd->td_mpcount -= td->td_mpcount;
869 ++preempt_miss;
8ec60c3f 870 need_lwkt_resched();
8a8d5d85
MD
871 return;
872 }
873#endif
26a0694b 874
8ec60c3f
MD
875 /*
876 * Since we are able to preempt the current thread, there is no need to
877 * call need_lwkt_resched().
878 */
26a0694b
MD
879 ++preempt_hit;
880 ntd->td_preempted = td;
881 td->td_flags |= TDF_PREEMPT_LOCK;
882 td->td_switch(ntd);
883 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
884#ifdef SMP
885 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
886 mpheld = MP_LOCK_HELD();
887 if (mpheld && td->td_mpcount == 0)
96728c05 888 cpu_rel_mplock();
71ef2f5c 889 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
890 panic("lwkt_preempt(): MP lock was not held through");
891#endif
26a0694b
MD
892 ntd->td_preempted = NULL;
893 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
894}
895
f1d1c3fa
MD
896/*
897 * Yield our thread while higher priority threads are pending. This is
898 * typically called when we leave a critical section but it can be safely
899 * called while we are in a critical section.
900 *
901 * This function will not generally yield to equal priority threads but it
902 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 903 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
904 * lwkt_yield_quick().
905 *
235957ed 906 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
907 * came along but was blocked and made pending.
908 *
f1d1c3fa
MD
909 * (self contained on a per cpu basis)
910 */
911void
912lwkt_yield_quick(void)
913{
7966cb69
MD
914 globaldata_t gd = mycpu;
915 thread_t td = gd->gd_curthread;
ef0fdad1 916
a2a5ad0d 917 /*
235957ed 918 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
919 * it with a non-zero cpl then we might not wind up calling splz after
920 * a task switch when the critical section is exited even though the
46a3f46d 921 * new task could accept the interrupt.
a2a5ad0d
MD
922 *
923 * XXX from crit_exit() only called after last crit section is released.
924 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
925 *
926 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
927 * except for this special case, we MUST call splz() here to handle any
928 * pending ints, particularly after we switch, or we might accidently
929 * halt the cpu with interrupts pending.
a2a5ad0d 930 */
46a3f46d 931 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 932 splz();
f1d1c3fa
MD
933
934 /*
935 * YYY enabling will cause wakeup() to task-switch, which really
936 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
937 * preemption and MP without actually doing preemption or MP, because a
938 * lot of code assumes that wakeup() does not block.
f1d1c3fa 939 */
46a3f46d
MD
940 if (untimely_switch && td->td_nest_count == 0 &&
941 gd->gd_intr_nesting_level == 0
942 ) {
37af14fe 943 crit_enter_quick(td);
f1d1c3fa
MD
944 /*
945 * YYY temporary hacks until we disassociate the userland scheduler
946 * from the LWKT scheduler.
947 */
948 if (td->td_flags & TDF_RUNQ) {
949 lwkt_switch(); /* will not reenter yield function */
950 } else {
37af14fe 951 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 952 lwkt_switch(); /* will not reenter yield function */
37af14fe 953 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 954 }
7966cb69 955 crit_exit_noyield(td);
f1d1c3fa 956 }
f1d1c3fa
MD
957}
958
8ad65e08 959/*
f1d1c3fa 960 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 961 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
962 * the crit_exit() call in lwkt_switch().
963 *
964 * (self contained on a per cpu basis)
8ad65e08
MD
965 */
966void
f1d1c3fa 967lwkt_yield(void)
8ad65e08 968{
37af14fe 969 lwkt_schedule_self(curthread);
f1d1c3fa
MD
970 lwkt_switch();
971}
972
8ad65e08 973/*
f1d1c3fa
MD
974 * Generic schedule. Possibly schedule threads belonging to other cpus and
975 * deal with threads that might be blocked on a wait queue.
976 *
0a3f9b47
MD
977 * We have a little helper inline function which does additional work after
978 * the thread has been enqueued, including dealing with preemption and
979 * setting need_lwkt_resched() (which prevents the kernel from returning
980 * to userland until it has processed higher priority threads).
6330a558
MD
981 *
982 * It is possible for this routine to be called after a failed _enqueue
983 * (due to the target thread migrating, sleeping, or otherwise blocked).
984 * We have to check that the thread is actually on the run queue!
8ad65e08 985 */
0a3f9b47
MD
986static __inline
987void
8ec60c3f 988_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 989{
6330a558
MD
990 if (ntd->td_flags & TDF_RUNQ) {
991 if (ntd->td_preemptable) {
992 ntd->td_preemptable(ntd, cpri); /* YYY +token */
993 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
994 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
995 ) {
996 need_lwkt_resched();
997 }
0a3f9b47
MD
998 }
999}
1000
8ad65e08
MD
1001void
1002lwkt_schedule(thread_t td)
1003{
37af14fe
MD
1004 globaldata_t mygd = mycpu;
1005
41a01a4d 1006 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1007 crit_enter_gd(mygd);
9388413d 1008 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1009 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1010 _lwkt_enqueue(td);
1011 } else {
f1d1c3fa 1012 /*
7cd8d145
MD
1013 * If we own the thread, there is no race (since we are in a
1014 * critical section). If we do not own the thread there might
1015 * be a race but the target cpu will deal with it.
f1d1c3fa 1016 */
0f7a3396 1017#ifdef SMP
7cd8d145 1018 if (td->td_gd == mygd) {
9d265729
MD
1019 _lwkt_enqueue(td);
1020 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1021 } else {
7cd8d145
MD
1022 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1023 }
0f7a3396 1024#else
7cd8d145
MD
1025 _lwkt_enqueue(td);
1026 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1027#endif
8ad65e08 1028 }
37af14fe 1029 crit_exit_gd(mygd);
8ad65e08
MD
1030}
1031
52eedfb5
MD
1032#ifdef SMP
1033
d9eea1a5 1034/*
52eedfb5
MD
1035 * Thread migration using a 'Pull' method. The thread may or may not be
1036 * the current thread. It MUST be descheduled and in a stable state.
1037 * lwkt_giveaway() must be called on the cpu owning the thread.
1038 *
1039 * At any point after lwkt_giveaway() is called, the target cpu may
1040 * 'pull' the thread by calling lwkt_acquire().
1041 *
1042 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1043 */
52eedfb5
MD
1044void
1045lwkt_giveaway(thread_t td)
1046{
1047 globaldata_t gd = mycpu;
1048
1049 crit_enter_gd(gd);
1050 KKASSERT(td->td_gd == gd);
1051 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1052 td->td_flags |= TDF_MIGRATING;
1053 crit_exit_gd(gd);
1054}
1055
a2a5ad0d
MD
1056void
1057lwkt_acquire(thread_t td)
1058{
37af14fe
MD
1059 globaldata_t gd;
1060 globaldata_t mygd;
a2a5ad0d 1061
52eedfb5 1062 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1063 gd = td->td_gd;
37af14fe 1064 mygd = mycpu;
52eedfb5 1065 if (gd != mycpu) {
35238fa5 1066 cpu_lfence();
52eedfb5 1067 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1068 crit_enter_gd(mygd);
52eedfb5
MD
1069 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1070 cpu_lfence();
37af14fe 1071 td->td_gd = mygd;
52eedfb5
MD
1072 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1073 td->td_flags &= ~TDF_MIGRATING;
1074 crit_exit_gd(mygd);
1075 } else {
1076 crit_enter_gd(mygd);
1077 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1078 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1079 crit_exit_gd(mygd);
a2a5ad0d
MD
1080 }
1081}
1082
52eedfb5
MD
1083#endif
1084
f1d1c3fa
MD
1085/*
1086 * Generic deschedule. Descheduling threads other then your own should be
1087 * done only in carefully controlled circumstances. Descheduling is
1088 * asynchronous.
1089 *
1090 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1091 */
1092void
1093lwkt_deschedule(thread_t td)
1094{
f1d1c3fa 1095 crit_enter();
b8a98473 1096#ifdef SMP
f1d1c3fa
MD
1097 if (td == curthread) {
1098 _lwkt_dequeue(td);
1099 } else {
a72187e9 1100 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1101 _lwkt_dequeue(td);
1102 } else {
b8a98473 1103 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1104 }
1105 }
b8a98473
MD
1106#else
1107 _lwkt_dequeue(td);
1108#endif
f1d1c3fa
MD
1109 crit_exit();
1110}
1111
4b5f931b
MD
1112/*
1113 * Set the target thread's priority. This routine does not automatically
1114 * switch to a higher priority thread, LWKT threads are not designed for
1115 * continuous priority changes. Yield if you want to switch.
1116 *
1117 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1118 * of the td_pri field. The specified priority may also indicate zero or
1119 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1120 *
1121 * Note that we requeue the thread whether it winds up on a different runq
1122 * or not. uio_yield() depends on this and the routine is not normally
1123 * called with the same priority otherwise.
4b5f931b
MD
1124 */
1125void
1126lwkt_setpri(thread_t td, int pri)
1127{
26a0694b 1128 KKASSERT(pri >= 0);
a72187e9 1129 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1130 crit_enter();
1131 if (td->td_flags & TDF_RUNQ) {
1132 _lwkt_dequeue(td);
1133 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1134 _lwkt_enqueue(td);
1135 } else {
1136 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1137 }
1138 crit_exit();
1139}
1140
1141void
1142lwkt_setpri_self(int pri)
1143{
1144 thread_t td = curthread;
1145
4b5f931b
MD
1146 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1147 crit_enter();
1148 if (td->td_flags & TDF_RUNQ) {
1149 _lwkt_dequeue(td);
1150 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1151 _lwkt_enqueue(td);
1152 } else {
1153 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1154 }
1155 crit_exit();
1156}
1157
8ec60c3f
MD
1158/*
1159 * Determine if there is a runnable thread at a higher priority then
1160 * the current thread. lwkt_setpri() does not check this automatically.
1161 * Return 1 if there is, 0 if there isn't.
1162 *
1163 * Example: if bit 31 of runqmask is set and the current thread is priority
1164 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1165 *
1166 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1167 * up comparing against 0xffffffff, a comparison that will always be false.
1168 */
1169int
1170lwkt_checkpri_self(void)
1171{
1172 globaldata_t gd = mycpu;
1173 thread_t td = gd->gd_curthread;
1174 int nq = td->td_pri & TDPRI_MASK;
1175
1176 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1177 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1178 return(1);
1179 ++nq;
1180 }
1181 return(0);
1182}
1183
5d21b981 1184/*
52eedfb5
MD
1185 * Migrate the current thread to the specified cpu.
1186 *
1187 * This is accomplished by descheduling ourselves from the current cpu,
1188 * moving our thread to the tdallq of the target cpu, IPI messaging the
1189 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1190 * races while the thread is being migrated.
5d21b981 1191 */
3d28ff59 1192#ifdef SMP
5d21b981 1193static void lwkt_setcpu_remote(void *arg);
3d28ff59 1194#endif
5d21b981
MD
1195
1196void
1197lwkt_setcpu_self(globaldata_t rgd)
1198{
1199#ifdef SMP
1200 thread_t td = curthread;
1201
1202 if (td->td_gd != rgd) {
1203 crit_enter_quick(td);
1204 td->td_flags |= TDF_MIGRATING;
1205 lwkt_deschedule_self(td);
52eedfb5 1206 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1207 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1208 lwkt_switch();
1209 /* we are now on the target cpu */
52eedfb5 1210 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1211 crit_exit_quick(td);
1212 }
1213#endif
1214}
1215
ecdefdda
MD
1216void
1217lwkt_migratecpu(int cpuid)
1218{
1219#ifdef SMP
1220 globaldata_t rgd;
1221
1222 rgd = globaldata_find(cpuid);
1223 lwkt_setcpu_self(rgd);
1224#endif
1225}
1226
5d21b981
MD
1227/*
1228 * Remote IPI for cpu migration (called while in a critical section so we
1229 * do not have to enter another one). The thread has already been moved to
1230 * our cpu's allq, but we must wait for the thread to be completely switched
1231 * out on the originating cpu before we schedule it on ours or the stack
1232 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1233 * change to main memory.
1234 *
1235 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1236 * against wakeups. It is best if this interface is used only when there
1237 * are no pending events that might try to schedule the thread.
1238 */
3d28ff59 1239#ifdef SMP
5d21b981
MD
1240static void
1241lwkt_setcpu_remote(void *arg)
1242{
1243 thread_t td = arg;
1244 globaldata_t gd = mycpu;
1245
c1102e9f 1246 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1247 cpu_lfence();
5d21b981 1248 td->td_gd = gd;
35238fa5 1249 cpu_sfence();
5d21b981 1250 td->td_flags &= ~TDF_MIGRATING;
9388413d 1251 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1252 _lwkt_enqueue(td);
1253}
3d28ff59 1254#endif
5d21b981 1255
553ea3c8 1256struct lwp *
4b5f931b
MD
1257lwkt_preempted_proc(void)
1258{
73e4f7b9 1259 thread_t td = curthread;
4b5f931b
MD
1260 while (td->td_preempted)
1261 td = td->td_preempted;
553ea3c8 1262 return(td->td_lwp);
4b5f931b
MD
1263}
1264
99df837e
MD
1265/*
1266 * Create a kernel process/thread/whatever. It shares it's address space
1267 * with proc0 - ie: kernel only.
1268 *
365fa13f
MD
1269 * NOTE! By default new threads are created with the MP lock held. A
1270 * thread which does not require the MP lock should release it by calling
1271 * rel_mplock() at the start of the new thread.
99df837e
MD
1272 */
1273int
1274lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1275 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1276 const char *fmt, ...)
99df837e 1277{
73e4f7b9 1278 thread_t td;
e2565a42 1279 __va_list ap;
99df837e 1280
d3d32139 1281 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1282 tdflags);
a2a5ad0d
MD
1283 if (tdp)
1284 *tdp = td;
709799ea 1285 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1286
1287 /*
1288 * Set up arg0 for 'ps' etc
1289 */
e2565a42 1290 __va_start(ap, fmt);
379210cb 1291 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1292 __va_end(ap);
99df837e
MD
1293
1294 /*
1295 * Schedule the thread to run
1296 */
ef0fdad1
MD
1297 if ((td->td_flags & TDF_STOPREQ) == 0)
1298 lwkt_schedule(td);
1299 else
1300 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1301 return 0;
1302}
1303
2d93b37a 1304/*
2d93b37a
MD
1305 * kthread_* is specific to the kernel and is not needed by userland.
1306 */
1307#ifdef _KERNEL
1308
99df837e
MD
1309/*
1310 * Destroy an LWKT thread. Warning! This function is not called when
1311 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1312 * uses a different reaping mechanism.
1313 */
1314void
1315lwkt_exit(void)
1316{
1317 thread_t td = curthread;
8826f33a 1318 globaldata_t gd;
99df837e
MD
1319
1320 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1321 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1322 caps_exit(td);
37af14fe
MD
1323 crit_enter_quick(td);
1324 lwkt_deschedule_self(td);
8826f33a 1325 gd = mycpu;
e56e4dea 1326 lwkt_remove_tdallq(td);
8826f33a
MD
1327 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1328 ++gd->gd_tdfreecount;
1329 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1330 }
99df837e
MD
1331 cpu_thread_exit();
1332}
1333
e56e4dea
MD
1334void
1335lwkt_remove_tdallq(thread_t td)
1336{
1337 KKASSERT(td->td_gd == mycpu);
1338 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1339}
1340
2d93b37a
MD
1341#endif /* _KERNEL */
1342
1343void
1344crit_panic(void)
1345{
1346 thread_t td = curthread;
1347 int lpri = td->td_pri;
1348
1349 td->td_pri = 0;
1350 panic("td_pri is/would-go negative! %p %d", td, lpri);
1351}
1352
d165e668
MD
1353#ifdef SMP
1354
bd8015ca
MD
1355/*
1356 * Called from debugger/panic on cpus which have been stopped. We must still
1357 * process the IPIQ while stopped, even if we were stopped while in a critical
1358 * section (XXX).
1359 *
1360 * If we are dumping also try to process any pending interrupts. This may
1361 * or may not work depending on the state of the cpu at the point it was
1362 * stopped.
1363 */
1364void
1365lwkt_smp_stopped(void)
1366{
1367 globaldata_t gd = mycpu;
1368
1369 crit_enter_gd(gd);
1370 if (dumping) {
1371 lwkt_process_ipiq();
1372 splz();
1373 } else {
1374 lwkt_process_ipiq();
1375 }
1376 crit_exit_gd(gd);
1377}
1378
57aa743c
MD
1379/*
1380 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1381 * get_mplock() has already incremented td_mpcount. We must block and
1382 * not return until giant is held.
1383 *
1384 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1385 * reschedule the thread until it can obtain the giant lock for it.
1386 */
1387void
1388lwkt_mp_lock_contested(void)
1389{
1390#ifdef _KERNEL
1391 loggiant(beg);
1392#endif
1393 lwkt_switch();
1394#ifdef _KERNEL
1395 loggiant(end);
1396#endif
1397}
1398
d165e668 1399#endif