Bring in the remainder of the post-SoC amd64 enchilada.
[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 *
0aa16b5d 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe 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 42 */
1541028a 43#include "opt_ddb.h"
8ad65e08
MD
44
45#include <sys/param.h>
46#include <sys/systm.h>
47#include <sys/kernel.h>
48#include <sys/proc.h>
49#include <sys/rtprio.h>
50#include <sys/queue.h>
7d0bac62 51#include <sys/sysctl.h>
99df837e 52#include <sys/kthread.h>
f1d1c3fa 53#include <machine/cpu.h>
99df837e 54#include <sys/lock.h>
f6bf3af1 55#include <sys/caps.h>
9d265729 56#include <sys/spinlock.h>
57aa743c 57#include <sys/ktr.h>
9d265729
MD
58
59#include <sys/thread2.h>
60#include <sys/spinlock2.h>
f1d1c3fa 61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
1541028a
YT
74#ifdef DDB
75#include <ddb/ddb.h>
76#endif
77
40aaf5fc
NT
78static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
79
7d0bac62 80static int untimely_switch = 0;
0f7a3396
MD
81#ifdef INVARIANTS
82static int panic_on_cscount = 0;
83#endif
05220613
MD
84static __int64_t switch_count = 0;
85static __int64_t preempt_hit = 0;
86static __int64_t preempt_miss = 0;
87static __int64_t preempt_weird = 0;
38717797
HP
88static __int64_t token_contention_count = 0;
89static __int64_t mplock_contention_count = 0;
fb0f29c4 90static int lwkt_use_spin_port;
d2f86ad2 91#ifdef SMP
b9eb1c19 92static int chain_mplock = 0;
d2f86ad2 93#endif
40aaf5fc 94static struct objcache *thread_cache;
05220613 95
b9eb1c19
MD
96volatile cpumask_t mp_lock_contention_mask;
97
0855a2af
JG
98extern void cpu_heavy_restore(void);
99extern void cpu_lwkt_restore(void);
100extern void cpu_kthread_restore(void);
101extern void cpu_idle_restore(void);
102
103int
104jg_tos_ok(struct thread *td)
105{
106 if (td == NULL) {
107 return 1;
108 }
109 KKASSERT(td->td_sp != NULL);
110 unsigned long tos = ((unsigned long *)td->td_sp)[0];
111 int tos_ok = 0;
112 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore)
113 || (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
114 tos_ok = 1;
115 }
116 return tos_ok;
117}
118
fb0f29c4
MD
119/*
120 * We can make all thread ports use the spin backend instead of the thread
121 * backend. This should only be set to debug the spin backend.
122 */
123TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
124
05220613 125SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
126#ifdef INVARIANTS
127SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
128#endif
b9eb1c19
MD
129#ifdef SMP
130SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
131#endif
4b5f931b 132SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 133SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 134SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 135SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
136#ifdef INVARIANTS
137SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
138 &token_contention_count, 0, "spinning due to token contention");
139SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
140 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
141#endif
05220613 142
57aa743c
MD
143/*
144 * Kernel Trace
145 */
57aa743c
MD
146#if !defined(KTR_GIANT_CONTENTION)
147#define KTR_GIANT_CONTENTION KTR_ALL
148#endif
149
150KTR_INFO_MASTER(giant);
151KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
152KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
153
154#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
155
4b5f931b
MD
156/*
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
75cdbe6c
MD
159 *
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 164 */
f1d1c3fa
MD
165static __inline
166void
167_lwkt_dequeue(thread_t td)
168{
169 if (td->td_flags & TDF_RUNQ) {
4b5f931b 170 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 171 struct globaldata *gd = td->td_gd;
4b5f931b 172
f1d1c3fa 173 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
174 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
175 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
176 }
177}
178
179static __inline
180void
181_lwkt_enqueue(thread_t td)
182{
344ad853 183 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 184 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 185 struct globaldata *gd = td->td_gd;
4b5f931b 186
f1d1c3fa 187 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
188 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
189 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
190 }
191}
8ad65e08 192
40aaf5fc
NT
193static __boolean_t
194_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
195{
196 struct thread *td = (struct thread *)obj;
197
198 td->td_kstack = NULL;
199 td->td_kstack_size = 0;
200 td->td_flags = TDF_ALLOCATED_THREAD;
201 return (1);
202}
203
204static void
205_lwkt_thread_dtor(void *obj, void *privdata)
206{
207 struct thread *td = (struct thread *)obj;
208
209 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
210 ("_lwkt_thread_dtor: not allocated from objcache"));
211 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
212 td->td_kstack_size > 0,
213 ("_lwkt_thread_dtor: corrupted stack"));
214 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
215}
216
217/*
218 * Initialize the lwkt s/system.
219 */
220void
221lwkt_init(void)
222{
223 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
224 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
225 NULL, CACHE_NTHREADS/2,
226 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
227}
228
37af14fe
MD
229/*
230 * Schedule a thread to run. As the current thread we can always safely
231 * schedule ourselves, and a shortcut procedure is provided for that
232 * function.
233 *
234 * (non-blocking, self contained on a per cpu basis)
235 */
236void
237lwkt_schedule_self(thread_t td)
238{
239 crit_enter_quick(td);
37af14fe 240 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 241 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 242 _lwkt_enqueue(td);
37af14fe
MD
243 crit_exit_quick(td);
244}
245
246/*
247 * Deschedule a thread.
248 *
249 * (non-blocking, self contained on a per cpu basis)
250 */
251void
252lwkt_deschedule_self(thread_t td)
253{
254 crit_enter_quick(td);
37af14fe
MD
255 _lwkt_dequeue(td);
256 crit_exit_quick(td);
257}
258
8ad65e08
MD
259/*
260 * LWKTs operate on a per-cpu basis
261 *
73e4f7b9 262 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
263 */
264void
265lwkt_gdinit(struct globaldata *gd)
266{
4b5f931b
MD
267 int i;
268
269 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
270 TAILQ_INIT(&gd->gd_tdrunq[i]);
271 gd->gd_runqmask = 0;
73e4f7b9 272 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
273}
274
7d0bac62
MD
275/*
276 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
277 * or LWKT start address before it can be scheduled. If the target cpu is
278 * -1 the thread will be created on the current cpu.
0cfcada1
MD
279 *
280 * If you intend to create a thread without a process context this function
281 * does everything except load the startup and switcher function.
7d0bac62
MD
282 */
283thread_t
d3d32139 284lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 285{
c070746a 286 globaldata_t gd = mycpu;
99df837e 287 void *stack;
7d0bac62 288
c070746a
MD
289 /*
290 * If static thread storage is not supplied allocate a thread. Reuse
291 * a cached free thread if possible. gd_freetd is used to keep an exiting
292 * thread intact through the exit.
293 */
ef0fdad1 294 if (td == NULL) {
c070746a
MD
295 if ((td = gd->gd_freetd) != NULL)
296 gd->gd_freetd = NULL;
297 else
298 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
299 KASSERT((td->td_flags &
300 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
301 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
302 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 303 }
c070746a
MD
304
305 /*
306 * Try to reuse cached stack.
307 */
f470d0c8
MD
308 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
309 if (flags & TDF_ALLOCATED_STACK) {
e4846942 310 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
311 stack = NULL;
312 }
313 }
314 if (stack == NULL) {
e4846942 315 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 316 flags |= TDF_ALLOCATED_STACK;
99df837e 317 }
75cdbe6c 318 if (cpu < 0)
c070746a 319 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 320 else
f470d0c8 321 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 322 return(td);
7d0bac62
MD
323}
324
325/*
326 * Initialize a preexisting thread structure. This function is used by
327 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
328 *
f8c3996b
MD
329 * All threads start out in a critical section at a priority of
330 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
331 * appropriate. This function may send an IPI message when the
332 * requested cpu is not the current cpu and consequently gd_tdallq may
333 * not be initialized synchronously from the point of view of the originating
334 * cpu.
335 *
336 * NOTE! we have to be careful in regards to creating threads for other cpus
337 * if SMP has not yet been activated.
7d0bac62 338 */
41a01a4d
MD
339#ifdef SMP
340
75cdbe6c
MD
341static void
342lwkt_init_thread_remote(void *arg)
343{
344 thread_t td = arg;
345
52eedfb5
MD
346 /*
347 * Protected by critical section held by IPI dispatch
348 */
75cdbe6c
MD
349 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
350}
351
41a01a4d
MD
352#endif
353
7d0bac62 354void
f470d0c8
MD
355lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
356 struct globaldata *gd)
7d0bac62 357{
37af14fe
MD
358 globaldata_t mygd = mycpu;
359
99df837e
MD
360 bzero(td, sizeof(struct thread));
361 td->td_kstack = stack;
f470d0c8 362 td->td_kstack_size = stksize;
d3d32139 363 td->td_flags = flags;
26a0694b 364 td->td_gd = gd;
f8c3996b 365 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
366#ifdef SMP
367 if ((flags & TDF_MPSAFE) == 0)
368 td->td_mpcount = 1;
369#endif
fb0f29c4
MD
370 if (lwkt_use_spin_port)
371 lwkt_initport_spin(&td->td_msgport);
372 else
373 lwkt_initport_thread(&td->td_msgport, td);
99df837e 374 pmap_init_thread(td);
0f7a3396 375#ifdef SMP
5d21b981
MD
376 /*
377 * Normally initializing a thread for a remote cpu requires sending an
378 * IPI. However, the idlethread is setup before the other cpus are
379 * activated so we have to treat it as a special case. XXX manipulation
380 * of gd_tdallq requires the BGL.
381 */
382 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 383 crit_enter_gd(mygd);
75cdbe6c 384 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 385 crit_exit_gd(mygd);
75cdbe6c 386 } else {
2db3b277 387 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 388 }
0f7a3396 389#else
37af14fe 390 crit_enter_gd(mygd);
0f7a3396 391 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 392 crit_exit_gd(mygd);
0f7a3396 393#endif
73e4f7b9
MD
394}
395
396void
397lwkt_set_comm(thread_t td, const char *ctl, ...)
398{
e2565a42 399 __va_list va;
73e4f7b9 400
e2565a42 401 __va_start(va, ctl);
379210cb 402 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 403 __va_end(va);
7d0bac62
MD
404}
405
99df837e 406void
73e4f7b9 407lwkt_hold(thread_t td)
99df837e 408{
73e4f7b9
MD
409 ++td->td_refs;
410}
411
412void
413lwkt_rele(thread_t td)
414{
415 KKASSERT(td->td_refs > 0);
416 --td->td_refs;
417}
418
419void
420lwkt_wait_free(thread_t td)
421{
422 while (td->td_refs)
377d4740 423 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
424}
425
426void
427lwkt_free_thread(thread_t td)
428{
d9eea1a5 429 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
430 ("lwkt_free_thread: did not exit! %p", td));
431
40aaf5fc
NT
432 if (td->td_flags & TDF_ALLOCATED_THREAD) {
433 objcache_put(thread_cache, td);
434 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
435 /* client-allocated struct with internally allocated stack */
436 KASSERT(td->td_kstack && td->td_kstack_size > 0,
437 ("lwkt_free_thread: corrupted stack"));
438 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
439 td->td_kstack = NULL;
440 td->td_kstack_size = 0;
99df837e
MD
441 }
442}
443
444
8ad65e08
MD
445/*
446 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
447 * switch to the idlethread. Switching must occur within a critical
448 * section to avoid races with the scheduling queue.
449 *
450 * We always have full control over our cpu's run queue. Other cpus
451 * that wish to manipulate our queue must use the cpu_*msg() calls to
452 * talk to our cpu, so a critical section is all that is needed and
453 * the result is very, very fast thread switching.
454 *
96728c05
MD
455 * The LWKT scheduler uses a fixed priority model and round-robins at
456 * each priority level. User process scheduling is a totally
457 * different beast and LWKT priorities should not be confused with
458 * user process priorities.
f1d1c3fa 459 *
96728c05
MD
460 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
461 * cleans it up. Note that the td_switch() function cannot do anything that
462 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
463 * the target thread (not the current thread). It's nice to have a scheduler
464 * that does not need the MP lock to work because it allows us to do some
465 * really cool high-performance MP lock optimizations.
69d78e99
MD
466 *
467 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
468 * is not called by the current thread in the preemption case, only when
469 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
470 */
471void
472lwkt_switch(void)
473{
37af14fe
MD
474 globaldata_t gd = mycpu;
475 thread_t td = gd->gd_curthread;
8ad65e08 476 thread_t ntd;
8a8d5d85
MD
477#ifdef SMP
478 int mpheld;
479#endif
8ad65e08 480
46a3f46d 481 /*
27e88a6e
MD
482 * Switching from within a 'fast' (non thread switched) interrupt or IPI
483 * is illegal. However, we may have to do it anyway if we hit a fatal
484 * kernel trap or we have paniced.
485 *
486 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 487 */
27e88a6e
MD
488 if (gd->gd_intr_nesting_level) {
489 int savegdnest;
490 int savegdtrap;
491
492 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
493 panic("lwkt_switch: cannot switch from within "
494 "a fast interrupt, yet, td %p\n", td);
495 } else {
496 savegdnest = gd->gd_intr_nesting_level;
497 savegdtrap = gd->gd_trap_nesting_level;
498 gd->gd_intr_nesting_level = 0;
499 gd->gd_trap_nesting_level = 0;
a7422615
MD
500 if ((td->td_flags & TDF_PANICWARN) == 0) {
501 td->td_flags |= TDF_PANICWARN;
6ea70f76 502 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615
MD
503 "thread %p (%s)\n", td, td->td_comm);
504#ifdef DDB
505 db_print_backtrace();
506#endif
507 }
27e88a6e
MD
508 lwkt_switch();
509 gd->gd_intr_nesting_level = savegdnest;
510 gd->gd_trap_nesting_level = savegdtrap;
511 return;
512 }
96728c05 513 }
ef0fdad1 514
cb973d15
MD
515 /*
516 * Passive release (used to transition from user to kernel mode
517 * when we block or switch rather then when we enter the kernel).
518 * This function is NOT called if we are switching into a preemption
519 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
520 * our current process designation (if we have one) and become a true
521 * LWKT thread, and may also hand the current process designation to
522 * another process and schedule thread.
cb973d15
MD
523 */
524 if (td->td_release)
525 td->td_release(td);
526
37af14fe 527 crit_enter_gd(gd);
9d265729
MD
528 if (td->td_toks)
529 lwkt_relalltokens(td);
530
531 /*
b02926de
MD
532 * We had better not be holding any spin locks, but don't get into an
533 * endless panic loop.
9d265729 534 */
bbb31c5d
MD
535 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
536 ("lwkt_switch: still holding a shared spinlock %p!",
537 gd->gd_spinlock_rd));
d666840a
MD
538 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
539 ("lwkt_switch: still holding %d exclusive spinlocks!",
540 gd->gd_spinlocks_wr));
9d265729 541
8a8d5d85
MD
542
543#ifdef SMP
544 /*
545 * td_mpcount cannot be used to determine if we currently hold the
546 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
547 * to get the lock, and switch out if it can't. Our ownership of
548 * the actual lock will remain stable while we are in a critical section
549 * (but, of course, another cpu may own or release the lock so the
550 * actual value of mp_lock is not stable).
8a8d5d85
MD
551 */
552 mpheld = MP_LOCK_HELD();
0f7a3396
MD
553#ifdef INVARIANTS
554 if (td->td_cscount) {
6ea70f76 555 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
556 td);
557 if (panic_on_cscount)
558 panic("switching while mastering cpusync");
559 }
560#endif
8a8d5d85 561#endif
99df837e
MD
562 if ((ntd = td->td_preempted) != NULL) {
563 /*
564 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
565 * thread. This occurs transparently, whether the preempted thread
566 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
567 * itself).
568 *
569 * We have to setup the MP lock for the original thread after backing
570 * out the adjustment that was made to curthread when the original
571 * was preempted.
99df837e 572 */
26a0694b 573 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 574#ifdef SMP
96728c05 575 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 576 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
577 td, ntd, td->td_mpcount, ntd->td_mpcount);
578 }
8a8d5d85
MD
579 if (ntd->td_mpcount) {
580 td->td_mpcount -= ntd->td_mpcount;
581 KKASSERT(td->td_mpcount >= 0);
582 }
583#endif
26a0694b 584 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
585
586 /*
b9eb1c19
MD
587 * The interrupt may have woken a thread up, we need to properly
588 * set the reschedule flag if the originally interrupted thread is
589 * at a lower priority.
8ec60c3f
MD
590 */
591 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
592 need_lwkt_resched();
8a8d5d85 593 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 594 } else {
4b5f931b
MD
595 /*
596 * Priority queue / round-robin at each priority. Note that user
597 * processes run at a fixed, low priority and the user process
598 * scheduler deals with interactions between user processes
599 * by scheduling and descheduling them from the LWKT queue as
600 * necessary.
8a8d5d85
MD
601 *
602 * We have to adjust the MP lock for the target thread. If we
603 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
604 * thread that does not need the MP lock. If we cannot, we spin
605 * instead of HLT.
606 *
607 * A similar issue exists for the tokens held by the target thread.
608 * If we cannot obtain ownership of the tokens we cannot immediately
609 * schedule the thread.
610 */
611
8ec60c3f
MD
612 /*
613 * If an LWKT reschedule was requested, well that is what we are
614 * doing now so clear it.
615 */
616 clear_lwkt_resched();
4b5f931b
MD
617again:
618 if (gd->gd_runqmask) {
619 int nq = bsrl(gd->gd_runqmask);
620 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
621 gd->gd_runqmask &= ~(1 << nq);
622 goto again;
623 }
8a8d5d85 624#ifdef SMP
41a01a4d 625 /*
df6b8ba0
MD
626 * THREAD SELECTION FOR AN SMP MACHINE BUILD
627 *
41a01a4d
MD
628 * If the target needs the MP lock and we couldn't get it,
629 * or if the target is holding tokens and we could not
630 * gain ownership of the tokens, continue looking for a
631 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
632 *
633 * NOTE: the mpheld variable invalid after this conditional, it
634 * can change due to both cpu_try_mplock() returning success
9d265729 635 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
636 * we are trying to check the mpcount of a thread other then
637 * the current thread. Because of this, if the current thread
638 * is not holding td_mpcount, an IPI indirectly run via
9d265729 639 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 640 * cause the core MP lock to be released.
41a01a4d
MD
641 */
642 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 643 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 644 ) {
8a8d5d85 645 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
646
647 mpheld = MP_LOCK_HELD();
648 ntd = NULL;
8a8d5d85
MD
649 while (rqmask) {
650 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 651 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 652 /* spinning due to MP lock being held */
38717797 653#ifdef INVARIANTS
a453459d 654 ++mplock_contention_count;
38717797 655#endif
a453459d 656 /* mplock still not held, 'mpheld' still valid */
41a01a4d 657 continue;
38717797 658 }
a453459d
MD
659
660 /*
9d265729 661 * mpheld state invalid after getalltokens call returns
a453459d
MD
662 * failure, but the variable is only needed for
663 * the loop.
664 */
9d265729 665 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 666 /* spinning due to token contention */
38717797 667#ifdef INVARIANTS
a453459d 668 ++token_contention_count;
38717797 669#endif
a453459d 670 mpheld = MP_LOCK_HELD();
41a01a4d 671 continue;
38717797 672 }
41a01a4d 673 break;
8a8d5d85
MD
674 }
675 if (ntd)
676 break;
677 rqmask &= ~(1 << nq);
678 nq = bsrl(rqmask);
b9eb1c19
MD
679
680 /*
681 * We have two choices. We can either refuse to run a
682 * user thread when a kernel thread needs the MP lock
683 * but could not get it, or we can allow it to run but
684 * then expect an IPI (hopefully) later on to force a
685 * reschedule when the MP lock might become available.
686 */
687 if (nq < TDPRI_KERN_LPSCHED) {
688 if (chain_mplock == 0)
689 break;
690 atomic_set_int(&mp_lock_contention_mask,
691 gd->gd_cpumask);
692 /* continue loop, allow user threads to be scheduled */
693 }
8a8d5d85
MD
694 }
695 if (ntd == NULL) {
b402c633 696 cpu_mplock_contested();
a2a5ad0d
MD
697 ntd = &gd->gd_idlethread;
698 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 699 goto using_idle_thread;
8a8d5d85 700 } else {
344ad853 701 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
702 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
704 }
705 } else {
344ad853 706 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
707 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
708 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
709 }
710#else
df6b8ba0
MD
711 /*
712 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
713 * worry about tokens or the BGL. However, we still have
714 * to call lwkt_getalltokens() in order to properly detect
715 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 716 */
7eb611ef 717 lwkt_getalltokens(ntd);
344ad853 718 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
719 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
720 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 721#endif
4b5f931b 722 } else {
3c23a41a 723 /*
60f945af
MD
724 * We have nothing to run but only let the idle loop halt
725 * the cpu if there are no pending interrupts.
3c23a41a 726 */
a2a5ad0d 727 ntd = &gd->gd_idlethread;
60f945af 728 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 729 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 730#ifdef SMP
df6b8ba0
MD
731using_idle_thread:
732 /*
733 * The idle thread should not be holding the MP lock unless we
734 * are trapping in the kernel or in a panic. Since we select the
735 * idle thread unconditionally when no other thread is available,
736 * if the MP lock is desired during a panic or kernel trap, we
737 * have to loop in the scheduler until we get it.
738 */
739 if (ntd->td_mpcount) {
740 mpheld = MP_LOCK_HELD();
b402c633 741 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 742 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
743 } else if (mpheld == 0) {
744 cpu_mplock_contested();
df6b8ba0 745 goto again;
b402c633 746 }
df6b8ba0 747 }
a453459d 748#endif
4b5f931b 749 }
f1d1c3fa 750 }
26a0694b
MD
751 KASSERT(ntd->td_pri >= TDPRI_CRIT,
752 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
753
754 /*
755 * Do the actual switch. If the new target does not need the MP lock
756 * and we are holding it, release the MP lock. If the new target requires
757 * the MP lock we have already acquired it for the target.
758 */
759#ifdef SMP
760 if (ntd->td_mpcount == 0 ) {
761 if (MP_LOCK_HELD())
762 cpu_rel_mplock();
763 } else {
a453459d 764 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
765 }
766#endif
94f6d86e
MD
767 if (td != ntd) {
768 ++switch_count;
0855a2af 769 KKASSERT(jg_tos_ok(ntd));
f1d1c3fa 770 td->td_switch(ntd);
94f6d86e 771 }
37af14fe
MD
772 /* NOTE: current cpu may have changed after switch */
773 crit_exit_quick(td);
8ad65e08
MD
774}
775
b68b7282 776/*
96728c05
MD
777 * Request that the target thread preempt the current thread. Preemption
778 * only works under a specific set of conditions:
b68b7282 779 *
96728c05
MD
780 * - We are not preempting ourselves
781 * - The target thread is owned by the current cpu
782 * - We are not currently being preempted
783 * - The target is not currently being preempted
d3d1cbc8
MD
784 * - We are not holding any spin locks
785 * - The target thread is not holding any tokens
96728c05
MD
786 * - We are able to satisfy the target's MP lock requirements (if any).
787 *
788 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
789 * this is called via lwkt_schedule() through the td_preemptable callback.
790 * critpri is the managed critical priority that we should ignore in order
791 * to determine whether preemption is possible (aka usually just the crit
792 * priority of lwkt_schedule() itself).
b68b7282 793 *
26a0694b
MD
794 * XXX at the moment we run the target thread in a critical section during
795 * the preemption in order to prevent the target from taking interrupts
796 * that *WE* can't. Preemption is strictly limited to interrupt threads
797 * and interrupt-like threads, outside of a critical section, and the
798 * preempted source thread will be resumed the instant the target blocks
799 * whether or not the source is scheduled (i.e. preemption is supposed to
800 * be as transparent as possible).
4b5f931b 801 *
8a8d5d85
MD
802 * The target thread inherits our MP count (added to its own) for the
803 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
804 * MP lock during the preemption. Therefore, any preempting targets must be
805 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
806 * out of sync with the physical mp_lock, but we do not have to preserve
807 * the original ownership of the lock if it was out of synch (that is, we
808 * can leave it synchronized on return).
b68b7282
MD
809 */
810void
96728c05 811lwkt_preempt(thread_t ntd, int critpri)
b68b7282 812{
46a3f46d 813 struct globaldata *gd = mycpu;
0a3f9b47 814 thread_t td;
8a8d5d85
MD
815#ifdef SMP
816 int mpheld;
57c254db 817 int savecnt;
8a8d5d85 818#endif
b68b7282 819
26a0694b 820 /*
96728c05
MD
821 * The caller has put us in a critical section. We can only preempt
822 * if the caller of the caller was not in a critical section (basically
d666840a 823 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 824 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 825 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
826 *
827 * YYY The target thread must be in a critical section (else it must
828 * inherit our critical section? I dunno yet).
41a01a4d 829 *
0a3f9b47 830 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
831 */
832 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 833
0a3f9b47 834 td = gd->gd_curthread;
0a3f9b47 835 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
836 ++preempt_miss;
837 return;
838 }
96728c05
MD
839 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
840 ++preempt_miss;
8ec60c3f 841 need_lwkt_resched();
96728c05
MD
842 return;
843 }
844#ifdef SMP
46a3f46d 845 if (ntd->td_gd != gd) {
96728c05 846 ++preempt_miss;
8ec60c3f 847 need_lwkt_resched();
96728c05
MD
848 return;
849 }
850#endif
41a01a4d 851 /*
d3d1cbc8 852 * Take the easy way out and do not preempt if we are holding
d666840a 853 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
854 * preempted interlock against the target thread's tokens and whether
855 * we can get all the target thread's tokens, but this situation
856 * should not occur very often so its easier to simply not preempt.
d666840a
MD
857 * Also, plain spinlocks are impossible to figure out at this point so
858 * just don't preempt.
d3d1cbc8
MD
859 *
860 * Do not try to preempt if the target thread is holding any tokens.
861 * We could try to acquire the tokens but this case is so rare there
862 * is no need to support it.
41a01a4d 863 */
bbb31c5d 864 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 865 ++preempt_miss;
8ec60c3f 866 need_lwkt_resched();
41a01a4d
MD
867 return;
868 }
d3d1cbc8
MD
869 if (ntd->td_toks) {
870 ++preempt_miss;
871 need_lwkt_resched();
872 return;
873 }
26a0694b
MD
874 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
875 ++preempt_weird;
8ec60c3f 876 need_lwkt_resched();
26a0694b
MD
877 return;
878 }
879 if (ntd->td_preempted) {
4b5f931b 880 ++preempt_hit;
8ec60c3f 881 need_lwkt_resched();
26a0694b 882 return;
b68b7282 883 }
8a8d5d85 884#ifdef SMP
a2a5ad0d
MD
885 /*
886 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
887 * to or from the MP lock. In this case td_mpcount will be pre-disposed
888 * (non-zero) but not actually synchronized with the actual state of the
889 * lock. We can use it to imply an MP lock requirement for the
890 * preemption but we cannot use it to test whether we hold the MP lock
891 * or not.
a2a5ad0d 892 */
96728c05 893 savecnt = td->td_mpcount;
71ef2f5c 894 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
895 ntd->td_mpcount += td->td_mpcount;
896 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
897 ntd->td_mpcount -= td->td_mpcount;
898 ++preempt_miss;
8ec60c3f 899 need_lwkt_resched();
8a8d5d85
MD
900 return;
901 }
902#endif
26a0694b 903
8ec60c3f
MD
904 /*
905 * Since we are able to preempt the current thread, there is no need to
906 * call need_lwkt_resched().
907 */
26a0694b
MD
908 ++preempt_hit;
909 ntd->td_preempted = td;
910 td->td_flags |= TDF_PREEMPT_LOCK;
911 td->td_switch(ntd);
b9eb1c19 912
26a0694b 913 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
914#ifdef SMP
915 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
916 mpheld = MP_LOCK_HELD();
917 if (mpheld && td->td_mpcount == 0)
96728c05 918 cpu_rel_mplock();
71ef2f5c 919 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
920 panic("lwkt_preempt(): MP lock was not held through");
921#endif
26a0694b
MD
922 ntd->td_preempted = NULL;
923 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
924}
925
f1d1c3fa
MD
926/*
927 * Yield our thread while higher priority threads are pending. This is
928 * typically called when we leave a critical section but it can be safely
929 * called while we are in a critical section.
930 *
931 * This function will not generally yield to equal priority threads but it
932 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 933 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
934 * lwkt_yield_quick().
935 *
235957ed 936 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
937 * came along but was blocked and made pending.
938 *
f1d1c3fa
MD
939 * (self contained on a per cpu basis)
940 */
941void
942lwkt_yield_quick(void)
943{
7966cb69
MD
944 globaldata_t gd = mycpu;
945 thread_t td = gd->gd_curthread;
ef0fdad1 946
a2a5ad0d 947 /*
235957ed 948 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
949 * it with a non-zero cpl then we might not wind up calling splz after
950 * a task switch when the critical section is exited even though the
46a3f46d 951 * new task could accept the interrupt.
a2a5ad0d
MD
952 *
953 * XXX from crit_exit() only called after last crit section is released.
954 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
955 *
956 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
957 * except for this special case, we MUST call splz() here to handle any
958 * pending ints, particularly after we switch, or we might accidently
959 * halt the cpu with interrupts pending.
a2a5ad0d 960 */
46a3f46d 961 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 962 splz();
f1d1c3fa
MD
963
964 /*
965 * YYY enabling will cause wakeup() to task-switch, which really
966 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
967 * preemption and MP without actually doing preemption or MP, because a
968 * lot of code assumes that wakeup() does not block.
f1d1c3fa 969 */
46a3f46d
MD
970 if (untimely_switch && td->td_nest_count == 0 &&
971 gd->gd_intr_nesting_level == 0
972 ) {
37af14fe 973 crit_enter_quick(td);
f1d1c3fa
MD
974 /*
975 * YYY temporary hacks until we disassociate the userland scheduler
976 * from the LWKT scheduler.
977 */
978 if (td->td_flags & TDF_RUNQ) {
979 lwkt_switch(); /* will not reenter yield function */
980 } else {
37af14fe 981 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 982 lwkt_switch(); /* will not reenter yield function */
37af14fe 983 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 984 }
7966cb69 985 crit_exit_noyield(td);
f1d1c3fa 986 }
f1d1c3fa
MD
987}
988
8ad65e08 989/*
f1d1c3fa 990 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 991 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
992 * the crit_exit() call in lwkt_switch().
993 *
994 * (self contained on a per cpu basis)
8ad65e08
MD
995 */
996void
f1d1c3fa 997lwkt_yield(void)
8ad65e08 998{
37af14fe 999 lwkt_schedule_self(curthread);
f1d1c3fa
MD
1000 lwkt_switch();
1001}
1002
b9eb1c19
MD
1003/*
1004 * Return 0 if no runnable threads are pending at the same or higher
1005 * priority as the passed thread.
1006 *
1007 * Return 1 if runnable threads are pending at the same priority.
1008 *
1009 * Return 2 if runnable threads are pending at a higher priority.
1010 */
1011int
1012lwkt_check_resched(thread_t td)
1013{
1014 int pri = td->td_pri & TDPRI_MASK;
1015
1016 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1017 return(2);
1018 if (TAILQ_NEXT(td, td_threadq))
1019 return(1);
1020 return(0);
1021}
1022
8ad65e08 1023/*
f1d1c3fa
MD
1024 * Generic schedule. Possibly schedule threads belonging to other cpus and
1025 * deal with threads that might be blocked on a wait queue.
1026 *
0a3f9b47
MD
1027 * We have a little helper inline function which does additional work after
1028 * the thread has been enqueued, including dealing with preemption and
1029 * setting need_lwkt_resched() (which prevents the kernel from returning
1030 * to userland until it has processed higher priority threads).
6330a558
MD
1031 *
1032 * It is possible for this routine to be called after a failed _enqueue
1033 * (due to the target thread migrating, sleeping, or otherwise blocked).
1034 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1035 *
1036 * reschedok is an optimized constant propagated from lwkt_schedule() or
1037 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1038 * reschedule to be requested if the target thread has a higher priority.
1039 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1040 * be 0, prevented undesired reschedules.
8ad65e08 1041 */
0a3f9b47
MD
1042static __inline
1043void
361d01dd 1044_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1045{
b9eb1c19 1046 thread_t otd;
c730be20 1047
6330a558 1048 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1049 if (ntd->td_preemptable && reschedok) {
6330a558 1050 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1051 } else if (reschedok) {
b9eb1c19
MD
1052 otd = curthread;
1053 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1054 need_lwkt_resched();
6330a558 1055 }
0a3f9b47
MD
1056 }
1057}
1058
361d01dd 1059static __inline
8ad65e08 1060void
361d01dd 1061_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1062{
37af14fe
MD
1063 globaldata_t mygd = mycpu;
1064
41a01a4d 1065 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1066 crit_enter_gd(mygd);
9388413d 1067 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1068 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1069 _lwkt_enqueue(td);
1070 } else {
f1d1c3fa 1071 /*
7cd8d145
MD
1072 * If we own the thread, there is no race (since we are in a
1073 * critical section). If we do not own the thread there might
1074 * be a race but the target cpu will deal with it.
f1d1c3fa 1075 */
0f7a3396 1076#ifdef SMP
7cd8d145 1077 if (td->td_gd == mygd) {
9d265729 1078 _lwkt_enqueue(td);
361d01dd 1079 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1080 } else {
7cd8d145
MD
1081 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1082 }
0f7a3396 1083#else
7cd8d145 1084 _lwkt_enqueue(td);
361d01dd 1085 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1086#endif
8ad65e08 1087 }
37af14fe 1088 crit_exit_gd(mygd);
8ad65e08
MD
1089}
1090
361d01dd
MD
1091void
1092lwkt_schedule(thread_t td)
1093{
1094 _lwkt_schedule(td, 1);
1095}
1096
1097void
1098lwkt_schedule_noresched(thread_t td)
1099{
1100 _lwkt_schedule(td, 0);
1101}
1102
52eedfb5
MD
1103#ifdef SMP
1104
d9eea1a5 1105/*
52eedfb5
MD
1106 * Thread migration using a 'Pull' method. The thread may or may not be
1107 * the current thread. It MUST be descheduled and in a stable state.
1108 * lwkt_giveaway() must be called on the cpu owning the thread.
1109 *
1110 * At any point after lwkt_giveaway() is called, the target cpu may
1111 * 'pull' the thread by calling lwkt_acquire().
1112 *
1113 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1114 */
52eedfb5
MD
1115void
1116lwkt_giveaway(thread_t td)
1117{
1118 globaldata_t gd = mycpu;
1119
1120 crit_enter_gd(gd);
1121 KKASSERT(td->td_gd == gd);
1122 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1123 td->td_flags |= TDF_MIGRATING;
1124 crit_exit_gd(gd);
1125}
1126
a2a5ad0d
MD
1127void
1128lwkt_acquire(thread_t td)
1129{
37af14fe
MD
1130 globaldata_t gd;
1131 globaldata_t mygd;
a2a5ad0d 1132
52eedfb5 1133 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1134 gd = td->td_gd;
37af14fe 1135 mygd = mycpu;
52eedfb5 1136 if (gd != mycpu) {
35238fa5 1137 cpu_lfence();
52eedfb5 1138 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1139 crit_enter_gd(mygd);
df910c23
MD
1140 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1141#ifdef SMP
1142 lwkt_process_ipiq();
1143#endif
52eedfb5 1144 cpu_lfence();
df910c23 1145 }
37af14fe 1146 td->td_gd = mygd;
52eedfb5
MD
1147 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1148 td->td_flags &= ~TDF_MIGRATING;
1149 crit_exit_gd(mygd);
1150 } else {
1151 crit_enter_gd(mygd);
1152 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1153 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1154 crit_exit_gd(mygd);
a2a5ad0d
MD
1155 }
1156}
1157
52eedfb5
MD
1158#endif
1159
f1d1c3fa
MD
1160/*
1161 * Generic deschedule. Descheduling threads other then your own should be
1162 * done only in carefully controlled circumstances. Descheduling is
1163 * asynchronous.
1164 *
1165 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1166 */
1167void
1168lwkt_deschedule(thread_t td)
1169{
f1d1c3fa 1170 crit_enter();
b8a98473 1171#ifdef SMP
f1d1c3fa
MD
1172 if (td == curthread) {
1173 _lwkt_dequeue(td);
1174 } else {
a72187e9 1175 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1176 _lwkt_dequeue(td);
1177 } else {
b8a98473 1178 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1179 }
1180 }
b8a98473
MD
1181#else
1182 _lwkt_dequeue(td);
1183#endif
f1d1c3fa
MD
1184 crit_exit();
1185}
1186
4b5f931b
MD
1187/*
1188 * Set the target thread's priority. This routine does not automatically
1189 * switch to a higher priority thread, LWKT threads are not designed for
1190 * continuous priority changes. Yield if you want to switch.
1191 *
1192 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1193 * of the td_pri field. The specified priority may also indicate zero or
1194 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1195 *
1196 * Note that we requeue the thread whether it winds up on a different runq
1197 * or not. uio_yield() depends on this and the routine is not normally
1198 * called with the same priority otherwise.
4b5f931b
MD
1199 */
1200void
1201lwkt_setpri(thread_t td, int pri)
1202{
26a0694b 1203 KKASSERT(pri >= 0);
a72187e9 1204 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1205 crit_enter();
1206 if (td->td_flags & TDF_RUNQ) {
1207 _lwkt_dequeue(td);
1208 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1209 _lwkt_enqueue(td);
1210 } else {
1211 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1212 }
1213 crit_exit();
1214}
1215
1216void
1217lwkt_setpri_self(int pri)
1218{
1219 thread_t td = curthread;
1220
4b5f931b
MD
1221 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1222 crit_enter();
1223 if (td->td_flags & TDF_RUNQ) {
1224 _lwkt_dequeue(td);
1225 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1226 _lwkt_enqueue(td);
1227 } else {
1228 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1229 }
1230 crit_exit();
1231}
1232
5d21b981 1233/*
52eedfb5
MD
1234 * Migrate the current thread to the specified cpu.
1235 *
1236 * This is accomplished by descheduling ourselves from the current cpu,
1237 * moving our thread to the tdallq of the target cpu, IPI messaging the
1238 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1239 * races while the thread is being migrated.
5d21b981 1240 */
3d28ff59 1241#ifdef SMP
5d21b981 1242static void lwkt_setcpu_remote(void *arg);
3d28ff59 1243#endif
5d21b981
MD
1244
1245void
1246lwkt_setcpu_self(globaldata_t rgd)
1247{
1248#ifdef SMP
1249 thread_t td = curthread;
1250
1251 if (td->td_gd != rgd) {
1252 crit_enter_quick(td);
1253 td->td_flags |= TDF_MIGRATING;
1254 lwkt_deschedule_self(td);
52eedfb5 1255 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1256 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1257 lwkt_switch();
1258 /* we are now on the target cpu */
52eedfb5 1259 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1260 crit_exit_quick(td);
1261 }
1262#endif
1263}
1264
ecdefdda
MD
1265void
1266lwkt_migratecpu(int cpuid)
1267{
1268#ifdef SMP
1269 globaldata_t rgd;
1270
1271 rgd = globaldata_find(cpuid);
1272 lwkt_setcpu_self(rgd);
1273#endif
1274}
1275
5d21b981
MD
1276/*
1277 * Remote IPI for cpu migration (called while in a critical section so we
1278 * do not have to enter another one). The thread has already been moved to
1279 * our cpu's allq, but we must wait for the thread to be completely switched
1280 * out on the originating cpu before we schedule it on ours or the stack
1281 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1282 * change to main memory.
1283 *
1284 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1285 * against wakeups. It is best if this interface is used only when there
1286 * are no pending events that might try to schedule the thread.
1287 */
3d28ff59 1288#ifdef SMP
5d21b981
MD
1289static void
1290lwkt_setcpu_remote(void *arg)
1291{
1292 thread_t td = arg;
1293 globaldata_t gd = mycpu;
1294
df910c23
MD
1295 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1296#ifdef SMP
1297 lwkt_process_ipiq();
1298#endif
35238fa5 1299 cpu_lfence();
df910c23 1300 }
5d21b981 1301 td->td_gd = gd;
35238fa5 1302 cpu_sfence();
5d21b981 1303 td->td_flags &= ~TDF_MIGRATING;
9388413d 1304 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1305 _lwkt_enqueue(td);
1306}
3d28ff59 1307#endif
5d21b981 1308
553ea3c8 1309struct lwp *
4b5f931b
MD
1310lwkt_preempted_proc(void)
1311{
73e4f7b9 1312 thread_t td = curthread;
4b5f931b
MD
1313 while (td->td_preempted)
1314 td = td->td_preempted;
553ea3c8 1315 return(td->td_lwp);
4b5f931b
MD
1316}
1317
99df837e
MD
1318/*
1319 * Create a kernel process/thread/whatever. It shares it's address space
1320 * with proc0 - ie: kernel only.
1321 *
365fa13f
MD
1322 * NOTE! By default new threads are created with the MP lock held. A
1323 * thread which does not require the MP lock should release it by calling
1324 * rel_mplock() at the start of the new thread.
99df837e
MD
1325 */
1326int
1327lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1328 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1329 const char *fmt, ...)
99df837e 1330{
73e4f7b9 1331 thread_t td;
e2565a42 1332 __va_list ap;
99df837e 1333
d3d32139 1334 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1335 tdflags);
a2a5ad0d
MD
1336 if (tdp)
1337 *tdp = td;
709799ea 1338 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1339
1340 /*
1341 * Set up arg0 for 'ps' etc
1342 */
e2565a42 1343 __va_start(ap, fmt);
379210cb 1344 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1345 __va_end(ap);
99df837e
MD
1346
1347 /*
1348 * Schedule the thread to run
1349 */
ef0fdad1
MD
1350 if ((td->td_flags & TDF_STOPREQ) == 0)
1351 lwkt_schedule(td);
1352 else
1353 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1354 return 0;
1355}
1356
1357/*
1358 * Destroy an LWKT thread. Warning! This function is not called when
1359 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1360 * uses a different reaping mechanism.
1361 */
1362void
1363lwkt_exit(void)
1364{
1365 thread_t td = curthread;
c070746a 1366 thread_t std;
8826f33a 1367 globaldata_t gd;
99df837e
MD
1368
1369 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1370 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1371 caps_exit(td);
c070746a
MD
1372
1373 /*
1374 * Get us into a critical section to interlock gd_freetd and loop
1375 * until we can get it freed.
1376 *
1377 * We have to cache the current td in gd_freetd because objcache_put()ing
1378 * it would rip it out from under us while our thread is still active.
1379 */
1380 gd = mycpu;
37af14fe 1381 crit_enter_quick(td);
c070746a
MD
1382 while ((std = gd->gd_freetd) != NULL) {
1383 gd->gd_freetd = NULL;
1384 objcache_put(thread_cache, std);
1385 }
37af14fe 1386 lwkt_deschedule_self(td);
e56e4dea 1387 lwkt_remove_tdallq(td);
c070746a
MD
1388 if (td->td_flags & TDF_ALLOCATED_THREAD)
1389 gd->gd_freetd = td;
99df837e
MD
1390 cpu_thread_exit();
1391}
1392
e56e4dea
MD
1393void
1394lwkt_remove_tdallq(thread_t td)
1395{
1396 KKASSERT(td->td_gd == mycpu);
1397 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1398}
1399
2d93b37a
MD
1400void
1401crit_panic(void)
1402{
1403 thread_t td = curthread;
1404 int lpri = td->td_pri;
1405
1406 td->td_pri = 0;
1407 panic("td_pri is/would-go negative! %p %d", td, lpri);
1408}
1409
d165e668
MD
1410#ifdef SMP
1411
bd8015ca
MD
1412/*
1413 * Called from debugger/panic on cpus which have been stopped. We must still
1414 * process the IPIQ while stopped, even if we were stopped while in a critical
1415 * section (XXX).
1416 *
1417 * If we are dumping also try to process any pending interrupts. This may
1418 * or may not work depending on the state of the cpu at the point it was
1419 * stopped.
1420 */
1421void
1422lwkt_smp_stopped(void)
1423{
1424 globaldata_t gd = mycpu;
1425
1426 crit_enter_gd(gd);
1427 if (dumping) {
1428 lwkt_process_ipiq();
1429 splz();
1430 } else {
1431 lwkt_process_ipiq();
1432 }
1433 crit_exit_gd(gd);
1434}
1435
57aa743c
MD
1436/*
1437 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1438 * get_mplock() has already incremented td_mpcount. We must block and
1439 * not return until giant is held.
1440 *
1441 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1442 * reschedule the thread until it can obtain the giant lock for it.
1443 */
1444void
1445lwkt_mp_lock_contested(void)
1446{
57aa743c 1447 loggiant(beg);
57aa743c 1448 lwkt_switch();
57aa743c 1449 loggiant(end);
57aa743c
MD
1450}
1451
b9eb1c19
MD
1452/*
1453 * The rel_mplock() code will call this function after releasing the
1454 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1455 *
1456 * We then chain an IPI to a single other cpu potentially needing the
1457 * lock. This is a bit heuristical and we can wind up with IPIs flying
1458 * all over the place.
1459 */
1460static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1461
1462void
1463lwkt_mp_lock_uncontested(void)
1464{
1465 globaldata_t gd;
1466 globaldata_t dgd;
1467 cpumask_t mask;
1468 cpumask_t tmpmask;
1469 int cpuid;
1470
1471 if (chain_mplock) {
1472 gd = mycpu;
1473 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1474 mask = mp_lock_contention_mask;
1475 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1476
1477 if (mask) {
1478 if (mask & tmpmask)
1479 cpuid = bsfl(mask & tmpmask);
1480 else
1481 cpuid = bsfl(mask);
1482 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1483 dgd = globaldata_find(cpuid);
1484 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1485 }
1486 }
1487}
1488
1489/*
1490 * The idea is for this IPI to interrupt a potentially lower priority
1491 * thread, such as a user thread, to allow the scheduler to reschedule
1492 * a higher priority kernel thread that needs the MP lock.
1493 *
1494 * For now we set the LWKT reschedule flag which generates an AST in
1495 * doreti, though theoretically it is also possible to possibly preempt
1496 * here if the underlying thread was operating in user mode. Nah.
1497 */
1498static void
1499lwkt_mp_lock_uncontested_remote(void *arg __unused)
1500{
1501 need_lwkt_resched();
1502}
1503
d165e668 1504#endif