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