kernel - Fix bug in usched_set()
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf 2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 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:
60f60350 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.
60f60350 20 *
8c10bfcf
MD
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.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
47#include <sys/queue.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f6bf3af1 52#include <sys/caps.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
7d0bac62 67
99df837e 68#include <machine/stdarg.h>
96728c05 69#include <machine/smp.h>
99df837e 70
1541028a 71
40aaf5fc
NT
72static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
73
3824f392
MD
74#ifdef SMP
75static int mplock_countx = 0;
76#endif
0f7a3396
MD
77#ifdef INVARIANTS
78static int panic_on_cscount = 0;
79#endif
05220613
MD
80static __int64_t switch_count = 0;
81static __int64_t preempt_hit = 0;
82static __int64_t preempt_miss = 0;
83static __int64_t preempt_weird = 0;
38717797
HP
84static __int64_t token_contention_count = 0;
85static __int64_t mplock_contention_count = 0;
fb0f29c4 86static int lwkt_use_spin_port;
d2f86ad2 87#ifdef SMP
b9eb1c19 88static int chain_mplock = 0;
d2f86ad2 89#endif
40aaf5fc 90static struct objcache *thread_cache;
05220613 91
b9eb1c19
MD
92volatile cpumask_t mp_lock_contention_mask;
93
0855a2af
JG
94extern void cpu_heavy_restore(void);
95extern void cpu_lwkt_restore(void);
96extern void cpu_kthread_restore(void);
97extern void cpu_idle_restore(void);
98
85514115
MD
99#ifdef __amd64__
100
101static int
0855a2af
JG
102jg_tos_ok(struct thread *td)
103{
85514115
MD
104 void *tos;
105 int tos_ok;
106
0855a2af
JG
107 if (td == NULL) {
108 return 1;
109 }
110 KKASSERT(td->td_sp != NULL);
85514115
MD
111 tos = ((void **)td->td_sp)[0];
112 tos_ok = 0;
113 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
114 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
115 tos_ok = 1;
116 }
117 return tos_ok;
118}
119
85514115
MD
120#endif
121
fb0f29c4
MD
122/*
123 * We can make all thread ports use the spin backend instead of the thread
124 * backend. This should only be set to debug the spin backend.
125 */
126TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
127
0f7a3396
MD
128#ifdef INVARIANTS
129SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
130#endif
b9eb1c19
MD
131#ifdef SMP
132SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
133#endif
4b5f931b 134SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 135SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 136SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 137SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
138#ifdef INVARIANTS
139SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
140 &token_contention_count, 0, "spinning due to token contention");
141SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
142 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
143#endif
05220613 144
57aa743c
MD
145/*
146 * Kernel Trace
147 */
57aa743c
MD
148#if !defined(KTR_GIANT_CONTENTION)
149#define KTR_GIANT_CONTENTION KTR_ALL
150#endif
151
152KTR_INFO_MASTER(giant);
153KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
154KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
155
156#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
157
4b5f931b
MD
158/*
159 * These helper procedures handle the runq, they can only be called from
160 * within a critical section.
75cdbe6c
MD
161 *
162 * WARNING! Prior to SMP being brought up it is possible to enqueue and
163 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
164 * instead of 'mycpu' when referencing the globaldata structure. Once
165 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 166 */
f1d1c3fa
MD
167static __inline
168void
169_lwkt_dequeue(thread_t td)
170{
171 if (td->td_flags & TDF_RUNQ) {
4b5f931b 172 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 173 struct globaldata *gd = td->td_gd;
4b5f931b 174
f1d1c3fa 175 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
176 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
177 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
178 }
179}
180
181static __inline
182void
183_lwkt_enqueue(thread_t td)
184{
7f5d7ed7 185 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 186 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 187 struct globaldata *gd = td->td_gd;
4b5f931b 188
f1d1c3fa 189 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
190 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
191 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
192 }
193}
8ad65e08 194
40aaf5fc
NT
195static __boolean_t
196_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
197{
198 struct thread *td = (struct thread *)obj;
199
200 td->td_kstack = NULL;
201 td->td_kstack_size = 0;
202 td->td_flags = TDF_ALLOCATED_THREAD;
203 return (1);
204}
205
206static void
207_lwkt_thread_dtor(void *obj, void *privdata)
208{
209 struct thread *td = (struct thread *)obj;
210
211 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
212 ("_lwkt_thread_dtor: not allocated from objcache"));
213 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
214 td->td_kstack_size > 0,
215 ("_lwkt_thread_dtor: corrupted stack"));
216 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
217}
218
219/*
220 * Initialize the lwkt s/system.
221 */
222void
223lwkt_init(void)
224{
225 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
226 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
227 NULL, CACHE_NTHREADS/2,
228 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
229}
230
37af14fe
MD
231/*
232 * Schedule a thread to run. As the current thread we can always safely
233 * schedule ourselves, and a shortcut procedure is provided for that
234 * function.
235 *
236 * (non-blocking, self contained on a per cpu basis)
237 */
238void
239lwkt_schedule_self(thread_t td)
240{
241 crit_enter_quick(td);
37af14fe 242 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 243 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 244 _lwkt_enqueue(td);
37af14fe
MD
245 crit_exit_quick(td);
246}
247
248/*
249 * Deschedule a thread.
250 *
251 * (non-blocking, self contained on a per cpu basis)
252 */
253void
254lwkt_deschedule_self(thread_t td)
255{
256 crit_enter_quick(td);
37af14fe
MD
257 _lwkt_dequeue(td);
258 crit_exit_quick(td);
259}
260
8ad65e08
MD
261/*
262 * LWKTs operate on a per-cpu basis
263 *
73e4f7b9 264 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
265 */
266void
267lwkt_gdinit(struct globaldata *gd)
268{
4b5f931b
MD
269 int i;
270
271 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
272 TAILQ_INIT(&gd->gd_tdrunq[i]);
273 gd->gd_runqmask = 0;
73e4f7b9 274 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
275}
276
7d0bac62
MD
277/*
278 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
279 * or LWKT start address before it can be scheduled. If the target cpu is
280 * -1 the thread will be created on the current cpu.
0cfcada1
MD
281 *
282 * If you intend to create a thread without a process context this function
283 * does everything except load the startup and switcher function.
7d0bac62
MD
284 */
285thread_t
d3d32139 286lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 287{
c070746a 288 globaldata_t gd = mycpu;
99df837e 289 void *stack;
7d0bac62 290
c070746a
MD
291 /*
292 * If static thread storage is not supplied allocate a thread. Reuse
293 * a cached free thread if possible. gd_freetd is used to keep an exiting
294 * thread intact through the exit.
295 */
ef0fdad1 296 if (td == NULL) {
c070746a
MD
297 if ((td = gd->gd_freetd) != NULL)
298 gd->gd_freetd = NULL;
299 else
300 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
301 KASSERT((td->td_flags &
302 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
303 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
304 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 305 }
c070746a
MD
306
307 /*
308 * Try to reuse cached stack.
309 */
f470d0c8
MD
310 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
311 if (flags & TDF_ALLOCATED_STACK) {
e4846942 312 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
313 stack = NULL;
314 }
315 }
316 if (stack == NULL) {
e4846942 317 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 318 flags |= TDF_ALLOCATED_STACK;
99df837e 319 }
75cdbe6c 320 if (cpu < 0)
c070746a 321 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 322 else
f470d0c8 323 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 324 return(td);
7d0bac62
MD
325}
326
327/*
328 * Initialize a preexisting thread structure. This function is used by
329 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
330 *
f8c3996b
MD
331 * All threads start out in a critical section at a priority of
332 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
333 * appropriate. This function may send an IPI message when the
334 * requested cpu is not the current cpu and consequently gd_tdallq may
335 * not be initialized synchronously from the point of view of the originating
336 * cpu.
337 *
338 * NOTE! we have to be careful in regards to creating threads for other cpus
339 * if SMP has not yet been activated.
7d0bac62 340 */
41a01a4d
MD
341#ifdef SMP
342
75cdbe6c
MD
343static void
344lwkt_init_thread_remote(void *arg)
345{
346 thread_t td = arg;
347
52eedfb5
MD
348 /*
349 * Protected by critical section held by IPI dispatch
350 */
75cdbe6c
MD
351 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
352}
353
41a01a4d
MD
354#endif
355
7d0bac62 356void
f470d0c8
MD
357lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
358 struct globaldata *gd)
7d0bac62 359{
37af14fe
MD
360 globaldata_t mygd = mycpu;
361
99df837e
MD
362 bzero(td, sizeof(struct thread));
363 td->td_kstack = stack;
f470d0c8 364 td->td_kstack_size = stksize;
d3d32139 365 td->td_flags = flags;
26a0694b 366 td->td_gd = gd;
f8c3996b 367 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
368#ifdef SMP
369 if ((flags & TDF_MPSAFE) == 0)
370 td->td_mpcount = 1;
371#endif
fb0f29c4
MD
372 if (lwkt_use_spin_port)
373 lwkt_initport_spin(&td->td_msgport);
374 else
375 lwkt_initport_thread(&td->td_msgport, td);
99df837e 376 pmap_init_thread(td);
0f7a3396 377#ifdef SMP
5d21b981
MD
378 /*
379 * Normally initializing a thread for a remote cpu requires sending an
380 * IPI. However, the idlethread is setup before the other cpus are
381 * activated so we have to treat it as a special case. XXX manipulation
382 * of gd_tdallq requires the BGL.
383 */
384 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 385 crit_enter_gd(mygd);
75cdbe6c 386 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 387 crit_exit_gd(mygd);
75cdbe6c 388 } else {
2db3b277 389 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 390 }
0f7a3396 391#else
37af14fe 392 crit_enter_gd(mygd);
0f7a3396 393 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 394 crit_exit_gd(mygd);
0f7a3396 395#endif
73e4f7b9
MD
396}
397
398void
399lwkt_set_comm(thread_t td, const char *ctl, ...)
400{
e2565a42 401 __va_list va;
73e4f7b9 402
e2565a42 403 __va_start(va, ctl);
379210cb 404 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 405 __va_end(va);
7d0bac62
MD
406}
407
99df837e 408void
73e4f7b9 409lwkt_hold(thread_t td)
99df837e 410{
73e4f7b9
MD
411 ++td->td_refs;
412}
413
414void
415lwkt_rele(thread_t td)
416{
417 KKASSERT(td->td_refs > 0);
418 --td->td_refs;
419}
420
421void
422lwkt_wait_free(thread_t td)
423{
424 while (td->td_refs)
377d4740 425 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
426}
427
428void
429lwkt_free_thread(thread_t td)
430{
d9eea1a5 431 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
432 ("lwkt_free_thread: did not exit! %p", td));
433
40aaf5fc
NT
434 if (td->td_flags & TDF_ALLOCATED_THREAD) {
435 objcache_put(thread_cache, td);
436 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
437 /* client-allocated struct with internally allocated stack */
438 KASSERT(td->td_kstack && td->td_kstack_size > 0,
439 ("lwkt_free_thread: corrupted stack"));
440 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
441 td->td_kstack = NULL;
442 td->td_kstack_size = 0;
99df837e
MD
443 }
444}
445
446
8ad65e08
MD
447/*
448 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
449 * switch to the idlethread. Switching must occur within a critical
450 * section to avoid races with the scheduling queue.
451 *
452 * We always have full control over our cpu's run queue. Other cpus
453 * that wish to manipulate our queue must use the cpu_*msg() calls to
454 * talk to our cpu, so a critical section is all that is needed and
455 * the result is very, very fast thread switching.
456 *
96728c05
MD
457 * The LWKT scheduler uses a fixed priority model and round-robins at
458 * each priority level. User process scheduling is a totally
459 * different beast and LWKT priorities should not be confused with
460 * user process priorities.
f1d1c3fa 461 *
96728c05
MD
462 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
463 * cleans it up. Note that the td_switch() function cannot do anything that
464 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
465 * the target thread (not the current thread). It's nice to have a scheduler
466 * that does not need the MP lock to work because it allows us to do some
467 * really cool high-performance MP lock optimizations.
69d78e99
MD
468 *
469 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
470 * is not called by the current thread in the preemption case, only when
471 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
472 */
473void
474lwkt_switch(void)
475{
37af14fe
MD
476 globaldata_t gd = mycpu;
477 thread_t td = gd->gd_curthread;
8ad65e08 478 thread_t ntd;
8a8d5d85
MD
479#ifdef SMP
480 int mpheld;
481#endif
8ad65e08 482
46a3f46d 483 /*
27e88a6e
MD
484 * Switching from within a 'fast' (non thread switched) interrupt or IPI
485 * is illegal. However, we may have to do it anyway if we hit a fatal
486 * kernel trap or we have paniced.
487 *
488 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 489 */
27e88a6e
MD
490 if (gd->gd_intr_nesting_level) {
491 int savegdnest;
492 int savegdtrap;
493
494 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
495 panic("lwkt_switch: cannot switch from within "
496 "a fast interrupt, yet, td %p\n", td);
497 } else {
498 savegdnest = gd->gd_intr_nesting_level;
499 savegdtrap = gd->gd_trap_nesting_level;
500 gd->gd_intr_nesting_level = 0;
501 gd->gd_trap_nesting_level = 0;
a7422615
MD
502 if ((td->td_flags & TDF_PANICWARN) == 0) {
503 td->td_flags |= TDF_PANICWARN;
6ea70f76 504 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 505 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 506 print_backtrace();
a7422615 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 {
3824f392
MD
706 if (ntd->td_mpcount)
707 ++mplock_countx;
344ad853 708 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
709 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
711 }
712#else
df6b8ba0
MD
713 /*
714 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
715 * worry about tokens or the BGL. However, we still have
716 * to call lwkt_getalltokens() in order to properly detect
717 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 718 */
7eb611ef 719 lwkt_getalltokens(ntd);
344ad853 720 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
721 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
722 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 723#endif
4b5f931b 724 } else {
3c23a41a 725 /*
60f945af
MD
726 * We have nothing to run but only let the idle loop halt
727 * the cpu if there are no pending interrupts.
3c23a41a 728 */
a2a5ad0d 729 ntd = &gd->gd_idlethread;
60f945af 730 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 731 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 732#ifdef SMP
df6b8ba0
MD
733using_idle_thread:
734 /*
735 * The idle thread should not be holding the MP lock unless we
736 * are trapping in the kernel or in a panic. Since we select the
737 * idle thread unconditionally when no other thread is available,
738 * if the MP lock is desired during a panic or kernel trap, we
739 * have to loop in the scheduler until we get it.
740 */
741 if (ntd->td_mpcount) {
742 mpheld = MP_LOCK_HELD();
b402c633 743 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 744 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
745 } else if (mpheld == 0) {
746 cpu_mplock_contested();
df6b8ba0 747 goto again;
b402c633 748 }
df6b8ba0 749 }
a453459d 750#endif
4b5f931b 751 }
f1d1c3fa 752 }
26a0694b
MD
753 KASSERT(ntd->td_pri >= TDPRI_CRIT,
754 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
755
756 /*
757 * Do the actual switch. If the new target does not need the MP lock
758 * and we are holding it, release the MP lock. If the new target requires
759 * the MP lock we have already acquired it for the target.
760 */
761#ifdef SMP
762 if (ntd->td_mpcount == 0 ) {
763 if (MP_LOCK_HELD())
764 cpu_rel_mplock();
765 } else {
a453459d 766 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
767 }
768#endif
94f6d86e
MD
769 if (td != ntd) {
770 ++switch_count;
85514115 771#ifdef __amd64__
0855a2af 772 KKASSERT(jg_tos_ok(ntd));
85514115 773#endif
f1d1c3fa 774 td->td_switch(ntd);
94f6d86e 775 }
37af14fe
MD
776 /* NOTE: current cpu may have changed after switch */
777 crit_exit_quick(td);
8ad65e08
MD
778}
779
b68b7282 780/*
96728c05
MD
781 * Request that the target thread preempt the current thread. Preemption
782 * only works under a specific set of conditions:
b68b7282 783 *
96728c05
MD
784 * - We are not preempting ourselves
785 * - The target thread is owned by the current cpu
786 * - We are not currently being preempted
787 * - The target is not currently being preempted
d3d1cbc8
MD
788 * - We are not holding any spin locks
789 * - The target thread is not holding any tokens
96728c05
MD
790 * - We are able to satisfy the target's MP lock requirements (if any).
791 *
792 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
793 * this is called via lwkt_schedule() through the td_preemptable callback.
794 * critpri is the managed critical priority that we should ignore in order
795 * to determine whether preemption is possible (aka usually just the crit
796 * priority of lwkt_schedule() itself).
b68b7282 797 *
26a0694b
MD
798 * XXX at the moment we run the target thread in a critical section during
799 * the preemption in order to prevent the target from taking interrupts
800 * that *WE* can't. Preemption is strictly limited to interrupt threads
801 * and interrupt-like threads, outside of a critical section, and the
802 * preempted source thread will be resumed the instant the target blocks
803 * whether or not the source is scheduled (i.e. preemption is supposed to
804 * be as transparent as possible).
4b5f931b 805 *
8a8d5d85
MD
806 * The target thread inherits our MP count (added to its own) for the
807 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
808 * MP lock during the preemption. Therefore, any preempting targets must be
809 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
810 * out of sync with the physical mp_lock, but we do not have to preserve
811 * the original ownership of the lock if it was out of synch (that is, we
812 * can leave it synchronized on return).
b68b7282
MD
813 */
814void
96728c05 815lwkt_preempt(thread_t ntd, int critpri)
b68b7282 816{
46a3f46d 817 struct globaldata *gd = mycpu;
0a3f9b47 818 thread_t td;
8a8d5d85
MD
819#ifdef SMP
820 int mpheld;
57c254db 821 int savecnt;
8a8d5d85 822#endif
b68b7282 823
26a0694b 824 /*
96728c05
MD
825 * The caller has put us in a critical section. We can only preempt
826 * if the caller of the caller was not in a critical section (basically
d666840a 827 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 828 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 829 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
830 *
831 * YYY The target thread must be in a critical section (else it must
832 * inherit our critical section? I dunno yet).
41a01a4d 833 *
0a3f9b47 834 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
835 */
836 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 837
0a3f9b47 838 td = gd->gd_curthread;
0a3f9b47 839 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
840 ++preempt_miss;
841 return;
842 }
96728c05
MD
843 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
844 ++preempt_miss;
8ec60c3f 845 need_lwkt_resched();
96728c05
MD
846 return;
847 }
848#ifdef SMP
46a3f46d 849 if (ntd->td_gd != gd) {
96728c05 850 ++preempt_miss;
8ec60c3f 851 need_lwkt_resched();
96728c05
MD
852 return;
853 }
854#endif
41a01a4d 855 /*
d3d1cbc8 856 * Take the easy way out and do not preempt if we are holding
d666840a 857 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
858 * preempted interlock against the target thread's tokens and whether
859 * we can get all the target thread's tokens, but this situation
860 * should not occur very often so its easier to simply not preempt.
d666840a
MD
861 * Also, plain spinlocks are impossible to figure out at this point so
862 * just don't preempt.
d3d1cbc8
MD
863 *
864 * Do not try to preempt if the target thread is holding any tokens.
865 * We could try to acquire the tokens but this case is so rare there
866 * is no need to support it.
41a01a4d 867 */
bbb31c5d 868 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 869 ++preempt_miss;
8ec60c3f 870 need_lwkt_resched();
41a01a4d
MD
871 return;
872 }
d3d1cbc8
MD
873 if (ntd->td_toks) {
874 ++preempt_miss;
875 need_lwkt_resched();
876 return;
877 }
26a0694b
MD
878 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
879 ++preempt_weird;
8ec60c3f 880 need_lwkt_resched();
26a0694b
MD
881 return;
882 }
883 if (ntd->td_preempted) {
4b5f931b 884 ++preempt_hit;
8ec60c3f 885 need_lwkt_resched();
26a0694b 886 return;
b68b7282 887 }
8a8d5d85 888#ifdef SMP
a2a5ad0d
MD
889 /*
890 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
891 * to or from the MP lock. In this case td_mpcount will be pre-disposed
892 * (non-zero) but not actually synchronized with the actual state of the
893 * lock. We can use it to imply an MP lock requirement for the
894 * preemption but we cannot use it to test whether we hold the MP lock
895 * or not.
a2a5ad0d 896 */
96728c05 897 savecnt = td->td_mpcount;
71ef2f5c 898 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
899 ntd->td_mpcount += td->td_mpcount;
900 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
901 ntd->td_mpcount -= td->td_mpcount;
902 ++preempt_miss;
8ec60c3f 903 need_lwkt_resched();
8a8d5d85
MD
904 return;
905 }
906#endif
26a0694b 907
8ec60c3f
MD
908 /*
909 * Since we are able to preempt the current thread, there is no need to
910 * call need_lwkt_resched().
911 */
26a0694b
MD
912 ++preempt_hit;
913 ntd->td_preempted = td;
914 td->td_flags |= TDF_PREEMPT_LOCK;
915 td->td_switch(ntd);
b9eb1c19 916
26a0694b 917 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
918#ifdef SMP
919 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
920 mpheld = MP_LOCK_HELD();
921 if (mpheld && td->td_mpcount == 0)
96728c05 922 cpu_rel_mplock();
71ef2f5c 923 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
924 panic("lwkt_preempt(): MP lock was not held through");
925#endif
26a0694b
MD
926 ntd->td_preempted = NULL;
927 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
928}
929
f1d1c3fa 930/*
faaeffac 931 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 932 *
faaeffac
MD
933 * td_nest_count prevents deep nesting via splz() or doreti() which
934 * might otherwise blow out the kernel stack. Note that except for
935 * this special case, we MUST call splz() here to handle any
936 * pending ints, particularly after we switch, or we might accidently
937 * halt the cpu with interrupts pending.
ef0fdad1 938 *
f1d1c3fa
MD
939 * (self contained on a per cpu basis)
940 */
941void
faaeffac 942splz_check(void)
f1d1c3fa 943{
7966cb69
MD
944 globaldata_t gd = mycpu;
945 thread_t td = gd->gd_curthread;
ef0fdad1 946
46a3f46d 947 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 948 splz();
f1d1c3fa
MD
949}
950
8ad65e08 951/*
faaeffac
MD
952 * This implements a normal yield which will yield to equal priority
953 * threads as well as higher priority threads. Note that gd_reqflags
954 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
955 *
956 * (self contained on a per cpu basis)
8ad65e08
MD
957 */
958void
f1d1c3fa 959lwkt_yield(void)
8ad65e08 960{
37af14fe 961 lwkt_schedule_self(curthread);
f1d1c3fa
MD
962 lwkt_switch();
963}
964
3824f392
MD
965/*
966 * This function is used along with the lwkt_passive_recover() inline
967 * by the trap code to negotiate a passive release of the current
968 * process/lwp designation with the user scheduler.
969 */
970void
971lwkt_passive_release(struct thread *td)
972{
973 struct lwp *lp = td->td_lwp;
974
975 td->td_release = NULL;
976 lwkt_setpri_self(TDPRI_KERN_USER);
977 lp->lwp_proc->p_usched->release_curproc(lp);
978}
979
980/*
981 * Make a kernel thread act as if it were in user mode with regards
982 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
983 * loops which may be potentially cpu-bound can call lwkt_user_yield().
984 *
985 * The lwkt_user_yield() function is designed to have very low overhead
986 * if no yield is determined to be needed.
987 */
988void
989lwkt_user_yield(void)
990{
991 thread_t td = curthread;
992 struct lwp *lp = td->td_lwp;
993
994#ifdef SMP
995 /*
996 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
997 * kernel can prevent other cpus from servicing interrupt threads
998 * which still require the MP lock (which is a lot of them). This
999 * has a chaining effect since if the interrupt is blocked, so is
1000 * the event, so normal scheduling will not pick up on the problem.
1001 */
1002 if (mplock_countx && td->td_mpcount) {
1003 int savecnt = td->td_mpcount;
1004
1005 td->td_mpcount = 1;
1006 rel_mplock();
1007 DELAY(10);
1008 get_mplock();
1009 td->td_mpcount = savecnt;
1010 mplock_countx = 0;
1011 }
1012#endif
1013
1014 /*
1015 * Another kernel thread wants the cpu
1016 */
1017 if (lwkt_resched_wanted())
1018 lwkt_switch();
1019
1020 /*
1021 * If the user scheduler has asynchronously determined that the current
1022 * process (when running in user mode) needs to lose the cpu then make
1023 * sure we are released.
1024 */
1025 if (user_resched_wanted()) {
1026 if (td->td_release)
1027 td->td_release(td);
1028 }
1029
1030 /*
1031 * If we are released reduce our priority
1032 */
1033 if (td->td_release == NULL) {
1034 if (lwkt_check_resched(td) > 0)
1035 lwkt_switch();
1036 lp->lwp_proc->p_usched->acquire_curproc(lp);
1037 td->td_release = lwkt_passive_release;
1038 lwkt_setpri_self(TDPRI_USER_NORM);
1039 }
1040}
1041
b9eb1c19
MD
1042/*
1043 * Return 0 if no runnable threads are pending at the same or higher
1044 * priority as the passed thread.
1045 *
1046 * Return 1 if runnable threads are pending at the same priority.
1047 *
1048 * Return 2 if runnable threads are pending at a higher priority.
1049 */
1050int
1051lwkt_check_resched(thread_t td)
1052{
1053 int pri = td->td_pri & TDPRI_MASK;
1054
1055 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1056 return(2);
1057 if (TAILQ_NEXT(td, td_threadq))
1058 return(1);
1059 return(0);
1060}
1061
8ad65e08 1062/*
f1d1c3fa
MD
1063 * Generic schedule. Possibly schedule threads belonging to other cpus and
1064 * deal with threads that might be blocked on a wait queue.
1065 *
0a3f9b47
MD
1066 * We have a little helper inline function which does additional work after
1067 * the thread has been enqueued, including dealing with preemption and
1068 * setting need_lwkt_resched() (which prevents the kernel from returning
1069 * to userland until it has processed higher priority threads).
6330a558
MD
1070 *
1071 * It is possible for this routine to be called after a failed _enqueue
1072 * (due to the target thread migrating, sleeping, or otherwise blocked).
1073 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1074 *
1075 * reschedok is an optimized constant propagated from lwkt_schedule() or
1076 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1077 * reschedule to be requested if the target thread has a higher priority.
1078 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1079 * be 0, prevented undesired reschedules.
8ad65e08 1080 */
0a3f9b47
MD
1081static __inline
1082void
361d01dd 1083_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1084{
b9eb1c19 1085 thread_t otd;
c730be20 1086
6330a558 1087 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1088 if (ntd->td_preemptable && reschedok) {
6330a558 1089 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1090 } else if (reschedok) {
b9eb1c19
MD
1091 otd = curthread;
1092 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1093 need_lwkt_resched();
6330a558 1094 }
0a3f9b47
MD
1095 }
1096}
1097
361d01dd 1098static __inline
8ad65e08 1099void
361d01dd 1100_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1101{
37af14fe
MD
1102 globaldata_t mygd = mycpu;
1103
41a01a4d 1104 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1105 crit_enter_gd(mygd);
9388413d 1106 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1107 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1108 _lwkt_enqueue(td);
1109 } else {
f1d1c3fa 1110 /*
7cd8d145
MD
1111 * If we own the thread, there is no race (since we are in a
1112 * critical section). If we do not own the thread there might
1113 * be a race but the target cpu will deal with it.
f1d1c3fa 1114 */
0f7a3396 1115#ifdef SMP
7cd8d145 1116 if (td->td_gd == mygd) {
9d265729 1117 _lwkt_enqueue(td);
361d01dd 1118 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1119 } else {
7cd8d145
MD
1120 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1121 }
0f7a3396 1122#else
7cd8d145 1123 _lwkt_enqueue(td);
361d01dd 1124 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1125#endif
8ad65e08 1126 }
37af14fe 1127 crit_exit_gd(mygd);
8ad65e08
MD
1128}
1129
361d01dd
MD
1130void
1131lwkt_schedule(thread_t td)
1132{
1133 _lwkt_schedule(td, 1);
1134}
1135
1136void
1137lwkt_schedule_noresched(thread_t td)
1138{
1139 _lwkt_schedule(td, 0);
1140}
1141
52eedfb5
MD
1142#ifdef SMP
1143
d9eea1a5 1144/*
52eedfb5
MD
1145 * Thread migration using a 'Pull' method. The thread may or may not be
1146 * the current thread. It MUST be descheduled and in a stable state.
1147 * lwkt_giveaway() must be called on the cpu owning the thread.
1148 *
1149 * At any point after lwkt_giveaway() is called, the target cpu may
1150 * 'pull' the thread by calling lwkt_acquire().
1151 *
ae8e83e6
MD
1152 * We have to make sure the thread is not sitting on a per-cpu tsleep
1153 * queue or it will blow up when it moves to another cpu.
1154 *
52eedfb5 1155 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1156 */
52eedfb5
MD
1157void
1158lwkt_giveaway(thread_t td)
1159{
3b4192fb 1160 globaldata_t gd = mycpu;
52eedfb5 1161
3b4192fb
MD
1162 crit_enter_gd(gd);
1163 if (td->td_flags & TDF_TSLEEPQ)
1164 tsleep_remove(td);
1165 KKASSERT(td->td_gd == gd);
1166 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1167 td->td_flags |= TDF_MIGRATING;
1168 crit_exit_gd(gd);
52eedfb5
MD
1169}
1170
a2a5ad0d
MD
1171void
1172lwkt_acquire(thread_t td)
1173{
37af14fe
MD
1174 globaldata_t gd;
1175 globaldata_t mygd;
a2a5ad0d 1176
52eedfb5 1177 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1178 gd = td->td_gd;
37af14fe 1179 mygd = mycpu;
52eedfb5 1180 if (gd != mycpu) {
35238fa5 1181 cpu_lfence();
52eedfb5 1182 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1183 crit_enter_gd(mygd);
df910c23
MD
1184 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1185#ifdef SMP
1186 lwkt_process_ipiq();
1187#endif
52eedfb5 1188 cpu_lfence();
df910c23 1189 }
37af14fe 1190 td->td_gd = mygd;
52eedfb5
MD
1191 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1192 td->td_flags &= ~TDF_MIGRATING;
1193 crit_exit_gd(mygd);
1194 } else {
1195 crit_enter_gd(mygd);
1196 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1197 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1198 crit_exit_gd(mygd);
a2a5ad0d
MD
1199 }
1200}
1201
52eedfb5
MD
1202#endif
1203
f1d1c3fa
MD
1204/*
1205 * Generic deschedule. Descheduling threads other then your own should be
1206 * done only in carefully controlled circumstances. Descheduling is
1207 * asynchronous.
1208 *
1209 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1210 */
1211void
1212lwkt_deschedule(thread_t td)
1213{
f1d1c3fa 1214 crit_enter();
b8a98473 1215#ifdef SMP
f1d1c3fa
MD
1216 if (td == curthread) {
1217 _lwkt_dequeue(td);
1218 } else {
a72187e9 1219 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1220 _lwkt_dequeue(td);
1221 } else {
b8a98473 1222 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1223 }
1224 }
b8a98473
MD
1225#else
1226 _lwkt_dequeue(td);
1227#endif
f1d1c3fa
MD
1228 crit_exit();
1229}
1230
4b5f931b
MD
1231/*
1232 * Set the target thread's priority. This routine does not automatically
1233 * switch to a higher priority thread, LWKT threads are not designed for
1234 * continuous priority changes. Yield if you want to switch.
1235 *
1236 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1237 * of the td_pri field. The specified priority may also indicate zero or
1238 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1239 *
1240 * Note that we requeue the thread whether it winds up on a different runq
1241 * or not. uio_yield() depends on this and the routine is not normally
1242 * called with the same priority otherwise.
4b5f931b
MD
1243 */
1244void
1245lwkt_setpri(thread_t td, int pri)
1246{
26a0694b 1247 KKASSERT(pri >= 0);
a72187e9 1248 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1249 crit_enter();
1250 if (td->td_flags & TDF_RUNQ) {
1251 _lwkt_dequeue(td);
1252 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1253 _lwkt_enqueue(td);
1254 } else {
1255 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1256 }
1257 crit_exit();
1258}
1259
1260void
1261lwkt_setpri_self(int pri)
1262{
1263 thread_t td = curthread;
1264
4b5f931b
MD
1265 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1266 crit_enter();
1267 if (td->td_flags & TDF_RUNQ) {
1268 _lwkt_dequeue(td);
1269 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1270 _lwkt_enqueue(td);
1271 } else {
1272 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1273 }
1274 crit_exit();
1275}
1276
5d21b981 1277/*
52eedfb5
MD
1278 * Migrate the current thread to the specified cpu.
1279 *
1280 * This is accomplished by descheduling ourselves from the current cpu,
1281 * moving our thread to the tdallq of the target cpu, IPI messaging the
1282 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1283 * races while the thread is being migrated.
ae8e83e6
MD
1284 *
1285 * We must be sure to remove ourselves from the current cpu's tsleepq
1286 * before potentially moving to another queue. The thread can be on
1287 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1288 */
3d28ff59 1289#ifdef SMP
5d21b981 1290static void lwkt_setcpu_remote(void *arg);
3d28ff59 1291#endif
5d21b981
MD
1292
1293void
1294lwkt_setcpu_self(globaldata_t rgd)
1295{
1296#ifdef SMP
1297 thread_t td = curthread;
1298
1299 if (td->td_gd != rgd) {
1300 crit_enter_quick(td);
ae8e83e6 1301 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1302 tsleep_remove(td);
5d21b981
MD
1303 td->td_flags |= TDF_MIGRATING;
1304 lwkt_deschedule_self(td);
52eedfb5 1305 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1306 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1307 lwkt_switch();
1308 /* we are now on the target cpu */
52eedfb5 1309 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1310 crit_exit_quick(td);
1311 }
1312#endif
1313}
1314
ecdefdda
MD
1315void
1316lwkt_migratecpu(int cpuid)
1317{
1318#ifdef SMP
1319 globaldata_t rgd;
1320
1321 rgd = globaldata_find(cpuid);
1322 lwkt_setcpu_self(rgd);
1323#endif
1324}
1325
5d21b981
MD
1326/*
1327 * Remote IPI for cpu migration (called while in a critical section so we
1328 * do not have to enter another one). The thread has already been moved to
1329 * our cpu's allq, but we must wait for the thread to be completely switched
1330 * out on the originating cpu before we schedule it on ours or the stack
1331 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1332 * change to main memory.
1333 *
1334 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1335 * against wakeups. It is best if this interface is used only when there
1336 * are no pending events that might try to schedule the thread.
1337 */
3d28ff59 1338#ifdef SMP
5d21b981
MD
1339static void
1340lwkt_setcpu_remote(void *arg)
1341{
1342 thread_t td = arg;
1343 globaldata_t gd = mycpu;
1344
df910c23
MD
1345 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1346#ifdef SMP
1347 lwkt_process_ipiq();
1348#endif
35238fa5 1349 cpu_lfence();
df910c23 1350 }
5d21b981 1351 td->td_gd = gd;
35238fa5 1352 cpu_sfence();
5d21b981 1353 td->td_flags &= ~TDF_MIGRATING;
9388413d 1354 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1355 _lwkt_enqueue(td);
1356}
3d28ff59 1357#endif
5d21b981 1358
553ea3c8 1359struct lwp *
4b5f931b
MD
1360lwkt_preempted_proc(void)
1361{
73e4f7b9 1362 thread_t td = curthread;
4b5f931b
MD
1363 while (td->td_preempted)
1364 td = td->td_preempted;
553ea3c8 1365 return(td->td_lwp);
4b5f931b
MD
1366}
1367
99df837e
MD
1368/*
1369 * Create a kernel process/thread/whatever. It shares it's address space
1370 * with proc0 - ie: kernel only.
1371 *
365fa13f
MD
1372 * NOTE! By default new threads are created with the MP lock held. A
1373 * thread which does not require the MP lock should release it by calling
1374 * rel_mplock() at the start of the new thread.
99df837e
MD
1375 */
1376int
1377lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1378 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1379 const char *fmt, ...)
99df837e 1380{
73e4f7b9 1381 thread_t td;
e2565a42 1382 __va_list ap;
99df837e 1383
d3d32139 1384 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1385 tdflags);
a2a5ad0d
MD
1386 if (tdp)
1387 *tdp = td;
709799ea 1388 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1389
1390 /*
1391 * Set up arg0 for 'ps' etc
1392 */
e2565a42 1393 __va_start(ap, fmt);
379210cb 1394 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1395 __va_end(ap);
99df837e
MD
1396
1397 /*
1398 * Schedule the thread to run
1399 */
ef0fdad1
MD
1400 if ((td->td_flags & TDF_STOPREQ) == 0)
1401 lwkt_schedule(td);
1402 else
1403 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1404 return 0;
1405}
1406
1407/*
1408 * Destroy an LWKT thread. Warning! This function is not called when
1409 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1410 * uses a different reaping mechanism.
1411 */
1412void
1413lwkt_exit(void)
1414{
1415 thread_t td = curthread;
c070746a 1416 thread_t std;
8826f33a 1417 globaldata_t gd;
99df837e
MD
1418
1419 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1420 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1421 caps_exit(td);
c070746a
MD
1422
1423 /*
1424 * Get us into a critical section to interlock gd_freetd and loop
1425 * until we can get it freed.
1426 *
1427 * We have to cache the current td in gd_freetd because objcache_put()ing
1428 * it would rip it out from under us while our thread is still active.
1429 */
1430 gd = mycpu;
37af14fe 1431 crit_enter_quick(td);
c070746a
MD
1432 while ((std = gd->gd_freetd) != NULL) {
1433 gd->gd_freetd = NULL;
1434 objcache_put(thread_cache, std);
1435 }
3b4192fb
MD
1436
1437 /*
1438 * Remove thread resources from kernel lists and deschedule us for
1439 * the last time.
1440 */
1441 if (td->td_flags & TDF_TSLEEPQ)
1442 tsleep_remove(td);
37af14fe 1443 lwkt_deschedule_self(td);
e56e4dea 1444 lwkt_remove_tdallq(td);
c070746a
MD
1445 if (td->td_flags & TDF_ALLOCATED_THREAD)
1446 gd->gd_freetd = td;
99df837e
MD
1447 cpu_thread_exit();
1448}
1449
e56e4dea
MD
1450void
1451lwkt_remove_tdallq(thread_t td)
1452{
1453 KKASSERT(td->td_gd == mycpu);
1454 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1455}
1456
2d93b37a
MD
1457void
1458crit_panic(void)
1459{
1460 thread_t td = curthread;
1461 int lpri = td->td_pri;
1462
1463 td->td_pri = 0;
1464 panic("td_pri is/would-go negative! %p %d", td, lpri);
1465}
1466
d165e668
MD
1467#ifdef SMP
1468
bd8015ca
MD
1469/*
1470 * Called from debugger/panic on cpus which have been stopped. We must still
1471 * process the IPIQ while stopped, even if we were stopped while in a critical
1472 * section (XXX).
1473 *
1474 * If we are dumping also try to process any pending interrupts. This may
1475 * or may not work depending on the state of the cpu at the point it was
1476 * stopped.
1477 */
1478void
1479lwkt_smp_stopped(void)
1480{
1481 globaldata_t gd = mycpu;
1482
1483 crit_enter_gd(gd);
1484 if (dumping) {
1485 lwkt_process_ipiq();
1486 splz();
1487 } else {
1488 lwkt_process_ipiq();
1489 }
1490 crit_exit_gd(gd);
1491}
1492
57aa743c
MD
1493/*
1494 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1495 * get_mplock() has already incremented td_mpcount. We must block and
1496 * not return until giant is held.
1497 *
1498 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1499 * reschedule the thread until it can obtain the giant lock for it.
1500 */
1501void
1502lwkt_mp_lock_contested(void)
1503{
3824f392 1504 ++mplock_countx;
57aa743c 1505 loggiant(beg);
57aa743c 1506 lwkt_switch();
57aa743c 1507 loggiant(end);
57aa743c
MD
1508}
1509
b9eb1c19
MD
1510/*
1511 * The rel_mplock() code will call this function after releasing the
1512 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1513 *
1514 * We then chain an IPI to a single other cpu potentially needing the
1515 * lock. This is a bit heuristical and we can wind up with IPIs flying
1516 * all over the place.
1517 */
1518static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1519
1520void
1521lwkt_mp_lock_uncontested(void)
1522{
1523 globaldata_t gd;
1524 globaldata_t dgd;
1525 cpumask_t mask;
1526 cpumask_t tmpmask;
1527 int cpuid;
1528
1529 if (chain_mplock) {
1530 gd = mycpu;
1531 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1532 mask = mp_lock_contention_mask;
1533 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1534
1535 if (mask) {
1536 if (mask & tmpmask)
1537 cpuid = bsfl(mask & tmpmask);
1538 else
1539 cpuid = bsfl(mask);
1540 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1541 dgd = globaldata_find(cpuid);
1542 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1543 }
1544 }
1545}
1546
1547/*
1548 * The idea is for this IPI to interrupt a potentially lower priority
1549 * thread, such as a user thread, to allow the scheduler to reschedule
1550 * a higher priority kernel thread that needs the MP lock.
1551 *
1552 * For now we set the LWKT reschedule flag which generates an AST in
1553 * doreti, though theoretically it is also possible to possibly preempt
1554 * here if the underlying thread was operating in user mode. Nah.
1555 */
1556static void
1557lwkt_mp_lock_uncontested_remote(void *arg __unused)
1558{
1559 need_lwkt_resched();
1560}
1561
d165e668 1562#endif