xref: /dflybsd-src/sys/kern/lwkt_thread.c (revision 6693db176654a0f25095ec64d0a74d58dcf0e47e)
1 /*
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  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
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
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
32  * SUCH DAMAGE.
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.
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>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
51 #include <sys/lock.h>
52 #include <sys/caps.h>
53 #include <sys/spinlock.h>
54 #include <sys/ktr.h>
55 
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
58 #include <sys/mplock2.h>
59 
60 #include <vm/vm.h>
61 #include <vm/vm_param.h>
62 #include <vm/vm_kern.h>
63 #include <vm/vm_object.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_pager.h>
67 #include <vm/vm_extern.h>
68 
69 #include <machine/stdarg.h>
70 #include <machine/smp.h>
71 
72 #if !defined(KTR_CTXSW)
73 #define KTR_CTXSW KTR_ALL
74 #endif
75 KTR_INFO_MASTER(ctxsw);
76 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "sw  %p > %p", 2 * sizeof(struct thread *));
77 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
78 
79 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
80 
81 #ifdef	INVARIANTS
82 static int panic_on_cscount = 0;
83 #endif
84 static __int64_t switch_count = 0;
85 static __int64_t preempt_hit = 0;
86 static __int64_t preempt_miss = 0;
87 static __int64_t preempt_weird = 0;
88 static __int64_t token_contention_count __debugvar = 0;
89 static int lwkt_use_spin_port;
90 static struct objcache *thread_cache;
91 
92 #ifdef SMP
93 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
94 #endif
95 
96 extern void cpu_heavy_restore(void);
97 extern void cpu_lwkt_restore(void);
98 extern void cpu_kthread_restore(void);
99 extern void cpu_idle_restore(void);
100 
101 #ifdef __x86_64__
102 
103 static int
104 jg_tos_ok(struct thread *td)
105 {
106 	void *tos;
107 	int tos_ok;
108 
109 	if (td == NULL) {
110 		return 1;
111 	}
112 	KKASSERT(td->td_sp != NULL);
113 	tos = ((void **)td->td_sp)[0];
114 	tos_ok = 0;
115 	if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
116 	    (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
117 		tos_ok = 1;
118 	}
119 	return tos_ok;
120 }
121 
122 #endif
123 
124 /*
125  * We can make all thread ports use the spin backend instead of the thread
126  * backend.  This should only be set to debug the spin backend.
127  */
128 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
129 
130 #ifdef	INVARIANTS
131 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
132 #endif
133 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
134 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
135 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
136 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
137 #ifdef	INVARIANTS
138 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
139 	&token_contention_count, 0, "spinning due to token contention");
140 #endif
141 
142 /*
143  * These helper procedures handle the runq, they can only be called from
144  * within a critical section.
145  *
146  * WARNING!  Prior to SMP being brought up it is possible to enqueue and
147  * dequeue threads belonging to other cpus, so be sure to use td->td_gd
148  * instead of 'mycpu' when referencing the globaldata structure.   Once
149  * SMP live enqueuing and dequeueing only occurs on the current cpu.
150  */
151 static __inline
152 void
153 _lwkt_dequeue(thread_t td)
154 {
155     if (td->td_flags & TDF_RUNQ) {
156 	int nq = td->td_pri & TDPRI_MASK;
157 	struct globaldata *gd = td->td_gd;
158 
159 	td->td_flags &= ~TDF_RUNQ;
160 	TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
161 	/* runqmask is passively cleaned up by the switcher */
162     }
163 }
164 
165 static __inline
166 void
167 _lwkt_enqueue(thread_t td)
168 {
169     if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
170 	int nq = td->td_pri & TDPRI_MASK;
171 	struct globaldata *gd = td->td_gd;
172 
173 	td->td_flags |= TDF_RUNQ;
174 	TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
175 	gd->gd_runqmask |= 1 << nq;
176     }
177 }
178 
179 static __boolean_t
180 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
181 {
182 	struct thread *td = (struct thread *)obj;
183 
184 	td->td_kstack = NULL;
185 	td->td_kstack_size = 0;
186 	td->td_flags = TDF_ALLOCATED_THREAD;
187 	return (1);
188 }
189 
190 static void
191 _lwkt_thread_dtor(void *obj, void *privdata)
192 {
193 	struct thread *td = (struct thread *)obj;
194 
195 	KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
196 	    ("_lwkt_thread_dtor: not allocated from objcache"));
197 	KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
198 		td->td_kstack_size > 0,
199 	    ("_lwkt_thread_dtor: corrupted stack"));
200 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
201 }
202 
203 /*
204  * Initialize the lwkt s/system.
205  */
206 void
207 lwkt_init(void)
208 {
209     /* An objcache has 2 magazines per CPU so divide cache size by 2. */
210     thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
211 			NULL, CACHE_NTHREADS/2,
212 			_lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
213 }
214 
215 /*
216  * Schedule a thread to run.  As the current thread we can always safely
217  * schedule ourselves, and a shortcut procedure is provided for that
218  * function.
219  *
220  * (non-blocking, self contained on a per cpu basis)
221  */
222 void
223 lwkt_schedule_self(thread_t td)
224 {
225     crit_enter_quick(td);
226     KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
227     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
228     _lwkt_enqueue(td);
229     crit_exit_quick(td);
230 }
231 
232 /*
233  * Deschedule a thread.
234  *
235  * (non-blocking, self contained on a per cpu basis)
236  */
237 void
238 lwkt_deschedule_self(thread_t td)
239 {
240     crit_enter_quick(td);
241     _lwkt_dequeue(td);
242     crit_exit_quick(td);
243 }
244 
245 /*
246  * LWKTs operate on a per-cpu basis
247  *
248  * WARNING!  Called from early boot, 'mycpu' may not work yet.
249  */
250 void
251 lwkt_gdinit(struct globaldata *gd)
252 {
253     int i;
254 
255     for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
256 	TAILQ_INIT(&gd->gd_tdrunq[i]);
257     gd->gd_runqmask = 0;
258     TAILQ_INIT(&gd->gd_tdallq);
259 }
260 
261 /*
262  * Create a new thread.  The thread must be associated with a process context
263  * or LWKT start address before it can be scheduled.  If the target cpu is
264  * -1 the thread will be created on the current cpu.
265  *
266  * If you intend to create a thread without a process context this function
267  * does everything except load the startup and switcher function.
268  */
269 thread_t
270 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
271 {
272     globaldata_t gd = mycpu;
273     void *stack;
274 
275     /*
276      * If static thread storage is not supplied allocate a thread.  Reuse
277      * a cached free thread if possible.  gd_freetd is used to keep an exiting
278      * thread intact through the exit.
279      */
280     if (td == NULL) {
281 	if ((td = gd->gd_freetd) != NULL)
282 	    gd->gd_freetd = NULL;
283 	else
284 	    td = objcache_get(thread_cache, M_WAITOK);
285     	KASSERT((td->td_flags &
286 		 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
287 		("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
288     	flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
289     }
290 
291     /*
292      * Try to reuse cached stack.
293      */
294     if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
295 	if (flags & TDF_ALLOCATED_STACK) {
296 	    kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
297 	    stack = NULL;
298 	}
299     }
300     if (stack == NULL) {
301 	stack = (void *)kmem_alloc(&kernel_map, stksize);
302 	flags |= TDF_ALLOCATED_STACK;
303     }
304     if (cpu < 0)
305 	lwkt_init_thread(td, stack, stksize, flags, gd);
306     else
307 	lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
308     return(td);
309 }
310 
311 /*
312  * Initialize a preexisting thread structure.  This function is used by
313  * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
314  *
315  * All threads start out in a critical section at a priority of
316  * TDPRI_KERN_DAEMON.  Higher level code will modify the priority as
317  * appropriate.  This function may send an IPI message when the
318  * requested cpu is not the current cpu and consequently gd_tdallq may
319  * not be initialized synchronously from the point of view of the originating
320  * cpu.
321  *
322  * NOTE! we have to be careful in regards to creating threads for other cpus
323  * if SMP has not yet been activated.
324  */
325 #ifdef SMP
326 
327 static void
328 lwkt_init_thread_remote(void *arg)
329 {
330     thread_t td = arg;
331 
332     /*
333      * Protected by critical section held by IPI dispatch
334      */
335     TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
336 }
337 
338 #endif
339 
340 void
341 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
342 		struct globaldata *gd)
343 {
344     globaldata_t mygd = mycpu;
345 
346     bzero(td, sizeof(struct thread));
347     td->td_kstack = stack;
348     td->td_kstack_size = stksize;
349     td->td_flags = flags;
350     td->td_gd = gd;
351     td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
352 #ifdef SMP
353     if ((flags & TDF_MPSAFE) == 0)
354 	td->td_mpcount = 1;
355 #endif
356     if (lwkt_use_spin_port)
357 	lwkt_initport_spin(&td->td_msgport);
358     else
359 	lwkt_initport_thread(&td->td_msgport, td);
360     pmap_init_thread(td);
361 #ifdef SMP
362     /*
363      * Normally initializing a thread for a remote cpu requires sending an
364      * IPI.  However, the idlethread is setup before the other cpus are
365      * activated so we have to treat it as a special case.  XXX manipulation
366      * of gd_tdallq requires the BGL.
367      */
368     if (gd == mygd || td == &gd->gd_idlethread) {
369 	crit_enter_gd(mygd);
370 	TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
371 	crit_exit_gd(mygd);
372     } else {
373 	lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
374     }
375 #else
376     crit_enter_gd(mygd);
377     TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
378     crit_exit_gd(mygd);
379 #endif
380 }
381 
382 void
383 lwkt_set_comm(thread_t td, const char *ctl, ...)
384 {
385     __va_list va;
386 
387     __va_start(va, ctl);
388     kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
389     __va_end(va);
390 }
391 
392 void
393 lwkt_hold(thread_t td)
394 {
395     ++td->td_refs;
396 }
397 
398 void
399 lwkt_rele(thread_t td)
400 {
401     KKASSERT(td->td_refs > 0);
402     --td->td_refs;
403 }
404 
405 void
406 lwkt_wait_free(thread_t td)
407 {
408     while (td->td_refs)
409 	tsleep(td, 0, "tdreap", hz);
410 }
411 
412 void
413 lwkt_free_thread(thread_t td)
414 {
415     KASSERT((td->td_flags & TDF_RUNNING) == 0,
416 	("lwkt_free_thread: did not exit! %p", td));
417 
418     if (td->td_flags & TDF_ALLOCATED_THREAD) {
419     	objcache_put(thread_cache, td);
420     } else if (td->td_flags & TDF_ALLOCATED_STACK) {
421 	/* client-allocated struct with internally allocated stack */
422 	KASSERT(td->td_kstack && td->td_kstack_size > 0,
423 	    ("lwkt_free_thread: corrupted stack"));
424 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
425 	td->td_kstack = NULL;
426 	td->td_kstack_size = 0;
427     }
428 }
429 
430 
431 /*
432  * Switch to the next runnable lwkt.  If no LWKTs are runnable then
433  * switch to the idlethread.  Switching must occur within a critical
434  * section to avoid races with the scheduling queue.
435  *
436  * We always have full control over our cpu's run queue.  Other cpus
437  * that wish to manipulate our queue must use the cpu_*msg() calls to
438  * talk to our cpu, so a critical section is all that is needed and
439  * the result is very, very fast thread switching.
440  *
441  * The LWKT scheduler uses a fixed priority model and round-robins at
442  * each priority level.  User process scheduling is a totally
443  * different beast and LWKT priorities should not be confused with
444  * user process priorities.
445  *
446  * The MP lock may be out of sync with the thread's td_mpcount.  lwkt_switch()
447  * cleans it up.  Note that the td_switch() function cannot do anything that
448  * requires the MP lock since the MP lock will have already been setup for
449  * the target thread (not the current thread).  It's nice to have a scheduler
450  * that does not need the MP lock to work because it allows us to do some
451  * really cool high-performance MP lock optimizations.
452  *
453  * PREEMPTION NOTE: Preemption occurs via lwkt_preempt().  lwkt_switch()
454  * is not called by the current thread in the preemption case, only when
455  * the preempting thread blocks (in order to return to the original thread).
456  */
457 void
458 lwkt_switch(void)
459 {
460     globaldata_t gd = mycpu;
461     thread_t td = gd->gd_curthread;
462     thread_t ntd;
463 #ifdef SMP
464     int mpheld;
465 #endif
466 
467     /*
468      * Switching from within a 'fast' (non thread switched) interrupt or IPI
469      * is illegal.  However, we may have to do it anyway if we hit a fatal
470      * kernel trap or we have paniced.
471      *
472      * If this case occurs save and restore the interrupt nesting level.
473      */
474     if (gd->gd_intr_nesting_level) {
475 	int savegdnest;
476 	int savegdtrap;
477 
478 	if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
479 	    panic("lwkt_switch: cannot switch from within "
480 		  "a fast interrupt, yet, td %p\n", td);
481 	} else {
482 	    savegdnest = gd->gd_intr_nesting_level;
483 	    savegdtrap = gd->gd_trap_nesting_level;
484 	    gd->gd_intr_nesting_level = 0;
485 	    gd->gd_trap_nesting_level = 0;
486 	    if ((td->td_flags & TDF_PANICWARN) == 0) {
487 		td->td_flags |= TDF_PANICWARN;
488 		kprintf("Warning: thread switch from interrupt or IPI, "
489 			"thread %p (%s)\n", td, td->td_comm);
490 		print_backtrace();
491 	    }
492 	    lwkt_switch();
493 	    gd->gd_intr_nesting_level = savegdnest;
494 	    gd->gd_trap_nesting_level = savegdtrap;
495 	    return;
496 	}
497     }
498 
499     /*
500      * Passive release (used to transition from user to kernel mode
501      * when we block or switch rather then when we enter the kernel).
502      * This function is NOT called if we are switching into a preemption
503      * or returning from a preemption.  Typically this causes us to lose
504      * our current process designation (if we have one) and become a true
505      * LWKT thread, and may also hand the current process designation to
506      * another process and schedule thread.
507      */
508     if (td->td_release)
509 	    td->td_release(td);
510 
511     crit_enter_gd(gd);
512     if (td->td_toks)
513 	    lwkt_relalltokens(td);
514 
515     /*
516      * We had better not be holding any spin locks, but don't get into an
517      * endless panic loop.
518      */
519     KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
520 	    ("lwkt_switch: still holding a shared spinlock %p!",
521 	     gd->gd_spinlock_rd));
522     KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
523 	    ("lwkt_switch: still holding %d exclusive spinlocks!",
524 	     gd->gd_spinlocks_wr));
525 
526 
527 #ifdef SMP
528     /*
529      * td_mpcount cannot be used to determine if we currently hold the
530      * MP lock because get_mplock() will increment it prior to attempting
531      * to get the lock, and switch out if it can't.  Our ownership of
532      * the actual lock will remain stable while we are in a critical section
533      * (but, of course, another cpu may own or release the lock so the
534      * actual value of mp_lock is not stable).
535      */
536     mpheld = MP_LOCK_HELD();
537 #ifdef	INVARIANTS
538     if (td->td_cscount) {
539 	kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
540 		td);
541 	if (panic_on_cscount)
542 	    panic("switching while mastering cpusync");
543     }
544 #endif
545 #endif
546     if ((ntd = td->td_preempted) != NULL) {
547 	/*
548 	 * We had preempted another thread on this cpu, resume the preempted
549 	 * thread.  This occurs transparently, whether the preempted thread
550 	 * was scheduled or not (it may have been preempted after descheduling
551 	 * itself).
552 	 *
553 	 * We have to setup the MP lock for the original thread after backing
554 	 * out the adjustment that was made to curthread when the original
555 	 * was preempted.
556 	 */
557 	KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
558 #ifdef SMP
559 	if (ntd->td_mpcount && mpheld == 0) {
560 	    panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
561 	       td, ntd, td->td_mpcount, ntd->td_mpcount);
562 	}
563 	if (ntd->td_mpcount) {
564 	    td->td_mpcount -= ntd->td_mpcount;
565 	    KKASSERT(td->td_mpcount >= 0);
566 	}
567 #endif
568 	ntd->td_flags |= TDF_PREEMPT_DONE;
569 
570 	/*
571 	 * The interrupt may have woken a thread up, we need to properly
572 	 * set the reschedule flag if the originally interrupted thread is
573 	 * at a lower priority.
574 	 */
575 	if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
576 	    need_lwkt_resched();
577 	/* YYY release mp lock on switchback if original doesn't need it */
578     } else {
579 	/*
580 	 * Priority queue / round-robin at each priority.  Note that user
581 	 * processes run at a fixed, low priority and the user process
582 	 * scheduler deals with interactions between user processes
583 	 * by scheduling and descheduling them from the LWKT queue as
584 	 * necessary.
585 	 *
586 	 * We have to adjust the MP lock for the target thread.  If we
587 	 * need the MP lock and cannot obtain it we try to locate a
588 	 * thread that does not need the MP lock.  If we cannot, we spin
589 	 * instead of HLT.
590 	 *
591 	 * A similar issue exists for the tokens held by the target thread.
592 	 * If we cannot obtain ownership of the tokens we cannot immediately
593 	 * schedule the thread.
594 	 */
595 
596 	/*
597 	 * If an LWKT reschedule was requested, well that is what we are
598 	 * doing now so clear it.
599 	 */
600 	clear_lwkt_resched();
601 again:
602 	if (gd->gd_runqmask) {
603 	    int nq = bsrl(gd->gd_runqmask);
604 	    if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
605 		gd->gd_runqmask &= ~(1 << nq);
606 		goto again;
607 	    }
608 #ifdef SMP
609 	    /*
610 	     * THREAD SELECTION FOR AN SMP MACHINE BUILD
611 	     *
612 	     * If the target needs the MP lock and we couldn't get it,
613 	     * or if the target is holding tokens and we could not
614 	     * gain ownership of the tokens, continue looking for a
615 	     * thread to schedule and spin instead of HLT if we can't.
616 	     *
617 	     * NOTE: the mpheld variable invalid after this conditional, it
618 	     * can change due to both cpu_try_mplock() returning success
619 	     * AND interactions in lwkt_getalltokens() due to the fact that
620 	     * we are trying to check the mpcount of a thread other then
621 	     * the current thread.  Because of this, if the current thread
622 	     * is not holding td_mpcount, an IPI indirectly run via
623 	     * lwkt_getalltokens() can obtain and release the MP lock and
624 	     * cause the core MP lock to be released.
625 	     */
626 	    if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
627 		(ntd->td_toks && lwkt_getalltokens(ntd) == 0)
628 	    ) {
629 		u_int32_t rqmask = gd->gd_runqmask;
630 
631 		mpheld = MP_LOCK_HELD();
632 		ntd = NULL;
633 		while (rqmask) {
634 		    TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
635 			if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
636 			    /* spinning due to MP lock being held */
637 			    continue;
638 			}
639 
640 			/*
641 			 * mpheld state invalid after getalltokens call returns
642 			 * failure, but the variable is only needed for
643 			 * the loop.
644 			 */
645 			if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
646 			    /* spinning due to token contention */
647 #ifdef	INVARIANTS
648 			    ++token_contention_count;
649 #endif
650 			    mpheld = MP_LOCK_HELD();
651 			    continue;
652 			}
653 			break;
654 		    }
655 		    if (ntd)
656 			break;
657 		    rqmask &= ~(1 << nq);
658 		    nq = bsrl(rqmask);
659 
660 		    /*
661 		     * We have two choices. We can either refuse to run a
662 		     * user thread when a kernel thread needs the MP lock
663 		     * but could not get it, or we can allow it to run but
664 		     * then expect an IPI (hopefully) later on to force a
665 		     * reschedule when the MP lock might become available.
666 		     */
667 		    if (nq < TDPRI_KERN_LPSCHED) {
668 			break;	/* for now refuse to run */
669 #if 0
670 			if (chain_mplock == 0)
671 				break;
672 			/* continue loop, allow user threads to be scheduled */
673 #endif
674 		    }
675 		}
676 
677 		/*
678 		 * Case where a (kernel) thread needed the MP lock and could
679 		 * not get one, and we may or may not have found another
680 		 * thread which does not need the MP lock to run while
681 		 * we wait (ntd).
682 		 */
683 		if (ntd == NULL) {
684 		    ntd = &gd->gd_idlethread;
685 		    ntd->td_flags |= TDF_IDLE_NOHLT;
686 		    set_mplock_contention_mask(gd);
687 		    cpu_mplock_contested();
688 		    goto using_idle_thread;
689 		} else {
690 		    clr_mplock_contention_mask(gd);
691 		    ++gd->gd_cnt.v_swtch;
692 		    TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
693 		    TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
694 		}
695 	    } else {
696 		clr_mplock_contention_mask(gd);
697 		++gd->gd_cnt.v_swtch;
698 		TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
699 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
700 	    }
701 #else
702 	    /*
703 	     * THREAD SELECTION FOR A UP MACHINE BUILD.  We don't have to
704 	     * worry about tokens or the BGL.  However, we still have
705 	     * to call lwkt_getalltokens() in order to properly detect
706 	     * stale tokens.  This call cannot fail for a UP build!
707 	     */
708 	    lwkt_getalltokens(ntd);
709 	    ++gd->gd_cnt.v_swtch;
710 	    TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
711 	    TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 #endif
713 	} else {
714 	    /*
715 	     * We have nothing to run but only let the idle loop halt
716 	     * the cpu if there are no pending interrupts.
717 	     */
718 	    ntd = &gd->gd_idlethread;
719 	    if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
720 		ntd->td_flags |= TDF_IDLE_NOHLT;
721 #ifdef SMP
722 using_idle_thread:
723 	    /*
724 	     * The idle thread should not be holding the MP lock unless we
725 	     * are trapping in the kernel or in a panic.  Since we select the
726 	     * idle thread unconditionally when no other thread is available,
727 	     * if the MP lock is desired during a panic or kernel trap, we
728 	     * have to loop in the scheduler until we get it.
729 	     */
730 	    if (ntd->td_mpcount) {
731 		mpheld = MP_LOCK_HELD();
732 		if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
733 		    panic("Idle thread %p was holding the BGL!", ntd);
734 		if (mpheld == 0)
735 		    goto again;
736 	    }
737 #endif
738 	}
739     }
740     KASSERT(ntd->td_pri >= TDPRI_CRIT,
741 	("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
742 
743     /*
744      * Do the actual switch.  If the new target does not need the MP lock
745      * and we are holding it, release the MP lock.  If the new target requires
746      * the MP lock we have already acquired it for the target.
747      */
748 #ifdef SMP
749     if (ntd->td_mpcount == 0 ) {
750 	if (MP_LOCK_HELD())
751 	    cpu_rel_mplock();
752     } else {
753 	ASSERT_MP_LOCK_HELD(ntd);
754     }
755 #endif
756     if (td != ntd) {
757 	++switch_count;
758 #ifdef __x86_64__
759     {
760 	int tos_ok __debugvar = jg_tos_ok(ntd);
761 	KKASSERT(tos_ok);
762     }
763 #endif
764 	KTR_LOG(ctxsw_sw, td, ntd);
765 	td->td_switch(ntd);
766     }
767     /* NOTE: current cpu may have changed after switch */
768     crit_exit_quick(td);
769 }
770 
771 /*
772  * Request that the target thread preempt the current thread.  Preemption
773  * only works under a specific set of conditions:
774  *
775  *	- We are not preempting ourselves
776  *	- The target thread is owned by the current cpu
777  *	- We are not currently being preempted
778  *	- The target is not currently being preempted
779  *	- We are not holding any spin locks
780  *	- The target thread is not holding any tokens
781  *	- We are able to satisfy the target's MP lock requirements (if any).
782  *
783  * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION.  Typically
784  * this is called via lwkt_schedule() through the td_preemptable callback.
785  * critpri is the managed critical priority that we should ignore in order
786  * to determine whether preemption is possible (aka usually just the crit
787  * priority of lwkt_schedule() itself).
788  *
789  * XXX at the moment we run the target thread in a critical section during
790  * the preemption in order to prevent the target from taking interrupts
791  * that *WE* can't.  Preemption is strictly limited to interrupt threads
792  * and interrupt-like threads, outside of a critical section, and the
793  * preempted source thread will be resumed the instant the target blocks
794  * whether or not the source is scheduled (i.e. preemption is supposed to
795  * be as transparent as possible).
796  *
797  * The target thread inherits our MP count (added to its own) for the
798  * duration of the preemption in order to preserve the atomicy of the
799  * MP lock during the preemption.  Therefore, any preempting targets must be
800  * careful in regards to MP assertions.  Note that the MP count may be
801  * out of sync with the physical mp_lock, but we do not have to preserve
802  * the original ownership of the lock if it was out of synch (that is, we
803  * can leave it synchronized on return).
804  */
805 void
806 lwkt_preempt(thread_t ntd, int critpri)
807 {
808     struct globaldata *gd = mycpu;
809     thread_t td;
810 #ifdef SMP
811     int mpheld;
812     int savecnt;
813 #endif
814 
815     /*
816      * The caller has put us in a critical section.  We can only preempt
817      * if the caller of the caller was not in a critical section (basically
818      * a local interrupt), as determined by the 'critpri' parameter.  We
819      * also can't preempt if the caller is holding any spinlocks (even if
820      * he isn't in a critical section).  This also handles the tokens test.
821      *
822      * YYY The target thread must be in a critical section (else it must
823      * inherit our critical section?  I dunno yet).
824      *
825      * Set need_lwkt_resched() unconditionally for now YYY.
826      */
827     KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
828 
829     td = gd->gd_curthread;
830     if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
831 	++preempt_miss;
832 	return;
833     }
834     if ((td->td_pri & ~TDPRI_MASK) > critpri) {
835 	++preempt_miss;
836 	need_lwkt_resched();
837 	return;
838     }
839 #ifdef SMP
840     if (ntd->td_gd != gd) {
841 	++preempt_miss;
842 	need_lwkt_resched();
843 	return;
844     }
845 #endif
846     /*
847      * Take the easy way out and do not preempt if we are holding
848      * any spinlocks.  We could test whether the thread(s) being
849      * preempted interlock against the target thread's tokens and whether
850      * we can get all the target thread's tokens, but this situation
851      * should not occur very often so its easier to simply not preempt.
852      * Also, plain spinlocks are impossible to figure out at this point so
853      * just don't preempt.
854      *
855      * Do not try to preempt if the target thread is holding any tokens.
856      * We could try to acquire the tokens but this case is so rare there
857      * is no need to support it.
858      */
859     if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
860 	++preempt_miss;
861 	need_lwkt_resched();
862 	return;
863     }
864     if (ntd->td_toks) {
865 	++preempt_miss;
866 	need_lwkt_resched();
867 	return;
868     }
869     if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
870 	++preempt_weird;
871 	need_lwkt_resched();
872 	return;
873     }
874     if (ntd->td_preempted) {
875 	++preempt_hit;
876 	need_lwkt_resched();
877 	return;
878     }
879 #ifdef SMP
880     /*
881      * note: an interrupt might have occured just as we were transitioning
882      * to or from the MP lock.  In this case td_mpcount will be pre-disposed
883      * (non-zero) but not actually synchronized with the actual state of the
884      * lock.  We can use it to imply an MP lock requirement for the
885      * preemption but we cannot use it to test whether we hold the MP lock
886      * or not.
887      */
888     savecnt = td->td_mpcount;
889     mpheld = MP_LOCK_HELD();
890     ntd->td_mpcount += td->td_mpcount;
891     if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
892 	ntd->td_mpcount -= td->td_mpcount;
893 	++preempt_miss;
894 	need_lwkt_resched();
895 	return;
896     }
897 #endif
898 
899     /*
900      * Since we are able to preempt the current thread, there is no need to
901      * call need_lwkt_resched().
902      */
903     ++preempt_hit;
904     ntd->td_preempted = td;
905     td->td_flags |= TDF_PREEMPT_LOCK;
906     KTR_LOG(ctxsw_pre, td, ntd);
907     td->td_switch(ntd);
908 
909     KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
910 #ifdef SMP
911     KKASSERT(savecnt == td->td_mpcount);
912     mpheld = MP_LOCK_HELD();
913     if (mpheld && td->td_mpcount == 0)
914 	cpu_rel_mplock();
915     else if (mpheld == 0 && td->td_mpcount)
916 	panic("lwkt_preempt(): MP lock was not held through");
917 #endif
918     ntd->td_preempted = NULL;
919     td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
920 }
921 
922 /*
923  * Conditionally call splz() if gd_reqflags indicates work is pending.
924  *
925  * td_nest_count prevents deep nesting via splz() or doreti() which
926  * might otherwise blow out the kernel stack.  Note that except for
927  * this special case, we MUST call splz() here to handle any
928  * pending ints, particularly after we switch, or we might accidently
929  * halt the cpu with interrupts pending.
930  *
931  * (self contained on a per cpu basis)
932  */
933 void
934 splz_check(void)
935 {
936     globaldata_t gd = mycpu;
937     thread_t td = gd->gd_curthread;
938 
939     if (gd->gd_reqflags && td->td_nest_count < 2)
940 	splz();
941 }
942 
943 /*
944  * This implements a normal yield which will yield to equal priority
945  * threads as well as higher priority threads.  Note that gd_reqflags
946  * tests will be handled by the crit_exit() call in lwkt_switch().
947  *
948  * (self contained on a per cpu basis)
949  */
950 void
951 lwkt_yield(void)
952 {
953     lwkt_schedule_self(curthread);
954     lwkt_switch();
955 }
956 
957 /*
958  * This function is used along with the lwkt_passive_recover() inline
959  * by the trap code to negotiate a passive release of the current
960  * process/lwp designation with the user scheduler.
961  */
962 void
963 lwkt_passive_release(struct thread *td)
964 {
965     struct lwp *lp = td->td_lwp;
966 
967     td->td_release = NULL;
968     lwkt_setpri_self(TDPRI_KERN_USER);
969     lp->lwp_proc->p_usched->release_curproc(lp);
970 }
971 
972 /*
973  * Make a kernel thread act as if it were in user mode with regards
974  * to scheduling, to avoid becoming cpu-bound in the kernel.  Kernel
975  * loops which may be potentially cpu-bound can call lwkt_user_yield().
976  *
977  * The lwkt_user_yield() function is designed to have very low overhead
978  * if no yield is determined to be needed.
979  */
980 void
981 lwkt_user_yield(void)
982 {
983     thread_t td = curthread;
984     struct lwp *lp = td->td_lwp;
985 
986 #ifdef SMP
987     /*
988      * XXX SEVERE TEMPORARY HACK.  A cpu-bound operation running in the
989      * kernel can prevent other cpus from servicing interrupt threads
990      * which still require the MP lock (which is a lot of them).  This
991      * has a chaining effect since if the interrupt is blocked, so is
992      * the event, so normal scheduling will not pick up on the problem.
993      */
994     if (mp_lock_contention_mask && td->td_mpcount) {
995 	yield_mplock(td);
996     }
997 #endif
998 
999     /*
1000      * Another kernel thread wants the cpu
1001      */
1002     if (lwkt_resched_wanted())
1003 	lwkt_switch();
1004 
1005     /*
1006      * If the user scheduler has asynchronously determined that the current
1007      * process (when running in user mode) needs to lose the cpu then make
1008      * sure we are released.
1009      */
1010     if (user_resched_wanted()) {
1011 	if (td->td_release)
1012 	    td->td_release(td);
1013     }
1014 
1015     /*
1016      * If we are released reduce our priority
1017      */
1018     if (td->td_release == NULL) {
1019 	if (lwkt_check_resched(td) > 0)
1020 		lwkt_switch();
1021 	if (lp) {
1022 		lp->lwp_proc->p_usched->acquire_curproc(lp);
1023 		td->td_release = lwkt_passive_release;
1024 		lwkt_setpri_self(TDPRI_USER_NORM);
1025 	}
1026     }
1027 }
1028 
1029 /*
1030  * Return 0 if no runnable threads are pending at the same or higher
1031  * priority as the passed thread.
1032  *
1033  * Return 1 if runnable threads are pending at the same priority.
1034  *
1035  * Return 2 if runnable threads are pending at a higher priority.
1036  */
1037 int
1038 lwkt_check_resched(thread_t td)
1039 {
1040 	int pri = td->td_pri & TDPRI_MASK;
1041 
1042 	if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1043 		return(2);
1044 	if (TAILQ_NEXT(td, td_threadq))
1045 		return(1);
1046 	return(0);
1047 }
1048 
1049 /*
1050  * Generic schedule.  Possibly schedule threads belonging to other cpus and
1051  * deal with threads that might be blocked on a wait queue.
1052  *
1053  * We have a little helper inline function which does additional work after
1054  * the thread has been enqueued, including dealing with preemption and
1055  * setting need_lwkt_resched() (which prevents the kernel from returning
1056  * to userland until it has processed higher priority threads).
1057  *
1058  * It is possible for this routine to be called after a failed _enqueue
1059  * (due to the target thread migrating, sleeping, or otherwise blocked).
1060  * We have to check that the thread is actually on the run queue!
1061  *
1062  * reschedok is an optimized constant propagated from lwkt_schedule() or
1063  * lwkt_schedule_noresched().  By default it is non-zero, causing a
1064  * reschedule to be requested if the target thread has a higher priority.
1065  * The port messaging code will set MSG_NORESCHED and cause reschedok to
1066  * be 0, prevented undesired reschedules.
1067  */
1068 static __inline
1069 void
1070 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1071 {
1072     thread_t otd;
1073 
1074     if (ntd->td_flags & TDF_RUNQ) {
1075 	if (ntd->td_preemptable && reschedok) {
1076 	    ntd->td_preemptable(ntd, cpri);	/* YYY +token */
1077 	} else if (reschedok) {
1078 	    otd = curthread;
1079 	    if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1080 		need_lwkt_resched();
1081 	}
1082     }
1083 }
1084 
1085 static __inline
1086 void
1087 _lwkt_schedule(thread_t td, int reschedok)
1088 {
1089     globaldata_t mygd = mycpu;
1090 
1091     KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1092     crit_enter_gd(mygd);
1093     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1094     if (td == mygd->gd_curthread) {
1095 	_lwkt_enqueue(td);
1096     } else {
1097 	/*
1098 	 * If we own the thread, there is no race (since we are in a
1099 	 * critical section).  If we do not own the thread there might
1100 	 * be a race but the target cpu will deal with it.
1101 	 */
1102 #ifdef SMP
1103 	if (td->td_gd == mygd) {
1104 	    _lwkt_enqueue(td);
1105 	    _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1106 	} else {
1107 	    lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1108 	}
1109 #else
1110 	_lwkt_enqueue(td);
1111 	_lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1112 #endif
1113     }
1114     crit_exit_gd(mygd);
1115 }
1116 
1117 void
1118 lwkt_schedule(thread_t td)
1119 {
1120     _lwkt_schedule(td, 1);
1121 }
1122 
1123 void
1124 lwkt_schedule_noresched(thread_t td)
1125 {
1126     _lwkt_schedule(td, 0);
1127 }
1128 
1129 #ifdef SMP
1130 
1131 /*
1132  * When scheduled remotely if frame != NULL the IPIQ is being
1133  * run via doreti or an interrupt then preemption can be allowed.
1134  *
1135  * To allow preemption we have to drop the critical section so only
1136  * one is present in _lwkt_schedule_post.
1137  */
1138 static void
1139 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1140 {
1141     thread_t td = curthread;
1142     thread_t ntd = arg;
1143 
1144     if (frame && ntd->td_preemptable) {
1145 	crit_exit_noyield(td);
1146 	_lwkt_schedule(ntd, 1);
1147 	crit_enter_quick(td);
1148     } else {
1149 	_lwkt_schedule(ntd, 1);
1150     }
1151 }
1152 
1153 /*
1154  * Thread migration using a 'Pull' method.  The thread may or may not be
1155  * the current thread.  It MUST be descheduled and in a stable state.
1156  * lwkt_giveaway() must be called on the cpu owning the thread.
1157  *
1158  * At any point after lwkt_giveaway() is called, the target cpu may
1159  * 'pull' the thread by calling lwkt_acquire().
1160  *
1161  * We have to make sure the thread is not sitting on a per-cpu tsleep
1162  * queue or it will blow up when it moves to another cpu.
1163  *
1164  * MPSAFE - must be called under very specific conditions.
1165  */
1166 void
1167 lwkt_giveaway(thread_t td)
1168 {
1169     globaldata_t gd = mycpu;
1170 
1171     crit_enter_gd(gd);
1172     if (td->td_flags & TDF_TSLEEPQ)
1173 	tsleep_remove(td);
1174     KKASSERT(td->td_gd == gd);
1175     TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1176     td->td_flags |= TDF_MIGRATING;
1177     crit_exit_gd(gd);
1178 }
1179 
1180 void
1181 lwkt_acquire(thread_t td)
1182 {
1183     globaldata_t gd;
1184     globaldata_t mygd;
1185 
1186     KKASSERT(td->td_flags & TDF_MIGRATING);
1187     gd = td->td_gd;
1188     mygd = mycpu;
1189     if (gd != mycpu) {
1190 	cpu_lfence();
1191 	KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1192 	crit_enter_gd(mygd);
1193 	while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1194 #ifdef SMP
1195 	    lwkt_process_ipiq();
1196 #endif
1197 	    cpu_lfence();
1198 	}
1199 	td->td_gd = mygd;
1200 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1201 	td->td_flags &= ~TDF_MIGRATING;
1202 	crit_exit_gd(mygd);
1203     } else {
1204 	crit_enter_gd(mygd);
1205 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1206 	td->td_flags &= ~TDF_MIGRATING;
1207 	crit_exit_gd(mygd);
1208     }
1209 }
1210 
1211 #endif
1212 
1213 /*
1214  * Generic deschedule.  Descheduling threads other then your own should be
1215  * done only in carefully controlled circumstances.  Descheduling is
1216  * asynchronous.
1217  *
1218  * This function may block if the cpu has run out of messages.
1219  */
1220 void
1221 lwkt_deschedule(thread_t td)
1222 {
1223     crit_enter();
1224 #ifdef SMP
1225     if (td == curthread) {
1226 	_lwkt_dequeue(td);
1227     } else {
1228 	if (td->td_gd == mycpu) {
1229 	    _lwkt_dequeue(td);
1230 	} else {
1231 	    lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1232 	}
1233     }
1234 #else
1235     _lwkt_dequeue(td);
1236 #endif
1237     crit_exit();
1238 }
1239 
1240 /*
1241  * Set the target thread's priority.  This routine does not automatically
1242  * switch to a higher priority thread, LWKT threads are not designed for
1243  * continuous priority changes.  Yield if you want to switch.
1244  *
1245  * We have to retain the critical section count which uses the high bits
1246  * of the td_pri field.  The specified priority may also indicate zero or
1247  * more critical sections by adding TDPRI_CRIT*N.
1248  *
1249  * Note that we requeue the thread whether it winds up on a different runq
1250  * or not.  uio_yield() depends on this and the routine is not normally
1251  * called with the same priority otherwise.
1252  */
1253 void
1254 lwkt_setpri(thread_t td, int pri)
1255 {
1256     KKASSERT(pri >= 0);
1257     KKASSERT(td->td_gd == mycpu);
1258     crit_enter();
1259     if (td->td_flags & TDF_RUNQ) {
1260 	_lwkt_dequeue(td);
1261 	td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1262 	_lwkt_enqueue(td);
1263     } else {
1264 	td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1265     }
1266     crit_exit();
1267 }
1268 
1269 /*
1270  * Set the initial priority for a thread prior to it being scheduled for
1271  * the first time.  The thread MUST NOT be scheduled before or during
1272  * this call.  The thread may be assigned to a cpu other then the current
1273  * cpu.
1274  *
1275  * Typically used after a thread has been created with TDF_STOPPREQ,
1276  * and before the thread is initially scheduled.
1277  */
1278 void
1279 lwkt_setpri_initial(thread_t td, int pri)
1280 {
1281     KKASSERT(pri >= 0);
1282     KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1283     td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1284 }
1285 
1286 void
1287 lwkt_setpri_self(int pri)
1288 {
1289     thread_t td = curthread;
1290 
1291     KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1292     crit_enter();
1293     if (td->td_flags & TDF_RUNQ) {
1294 	_lwkt_dequeue(td);
1295 	td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1296 	_lwkt_enqueue(td);
1297     } else {
1298 	td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1299     }
1300     crit_exit();
1301 }
1302 
1303 /*
1304  * Migrate the current thread to the specified cpu.
1305  *
1306  * This is accomplished by descheduling ourselves from the current cpu,
1307  * moving our thread to the tdallq of the target cpu, IPI messaging the
1308  * target cpu, and switching out.  TDF_MIGRATING prevents scheduling
1309  * races while the thread is being migrated.
1310  *
1311  * We must be sure to remove ourselves from the current cpu's tsleepq
1312  * before potentially moving to another queue.  The thread can be on
1313  * a tsleepq due to a left-over tsleep_interlock().
1314  */
1315 #ifdef SMP
1316 static void lwkt_setcpu_remote(void *arg);
1317 #endif
1318 
1319 void
1320 lwkt_setcpu_self(globaldata_t rgd)
1321 {
1322 #ifdef SMP
1323     thread_t td = curthread;
1324 
1325     if (td->td_gd != rgd) {
1326 	crit_enter_quick(td);
1327 	if (td->td_flags & TDF_TSLEEPQ)
1328 	    tsleep_remove(td);
1329 	td->td_flags |= TDF_MIGRATING;
1330 	lwkt_deschedule_self(td);
1331 	TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1332 	lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1333 	lwkt_switch();
1334 	/* we are now on the target cpu */
1335 	TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1336 	crit_exit_quick(td);
1337     }
1338 #endif
1339 }
1340 
1341 void
1342 lwkt_migratecpu(int cpuid)
1343 {
1344 #ifdef SMP
1345 	globaldata_t rgd;
1346 
1347 	rgd = globaldata_find(cpuid);
1348 	lwkt_setcpu_self(rgd);
1349 #endif
1350 }
1351 
1352 /*
1353  * Remote IPI for cpu migration (called while in a critical section so we
1354  * do not have to enter another one).  The thread has already been moved to
1355  * our cpu's allq, but we must wait for the thread to be completely switched
1356  * out on the originating cpu before we schedule it on ours or the stack
1357  * state may be corrupt.  We clear TDF_MIGRATING after flushing the GD
1358  * change to main memory.
1359  *
1360  * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1361  * against wakeups.  It is best if this interface is used only when there
1362  * are no pending events that might try to schedule the thread.
1363  */
1364 #ifdef SMP
1365 static void
1366 lwkt_setcpu_remote(void *arg)
1367 {
1368     thread_t td = arg;
1369     globaldata_t gd = mycpu;
1370 
1371     while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1372 #ifdef SMP
1373 	lwkt_process_ipiq();
1374 #endif
1375 	cpu_lfence();
1376     }
1377     td->td_gd = gd;
1378     cpu_sfence();
1379     td->td_flags &= ~TDF_MIGRATING;
1380     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1381     _lwkt_enqueue(td);
1382 }
1383 #endif
1384 
1385 struct lwp *
1386 lwkt_preempted_proc(void)
1387 {
1388     thread_t td = curthread;
1389     while (td->td_preempted)
1390 	td = td->td_preempted;
1391     return(td->td_lwp);
1392 }
1393 
1394 /*
1395  * Create a kernel process/thread/whatever.  It shares it's address space
1396  * with proc0 - ie: kernel only.
1397  *
1398  * NOTE!  By default new threads are created with the MP lock held.  A
1399  * thread which does not require the MP lock should release it by calling
1400  * rel_mplock() at the start of the new thread.
1401  */
1402 int
1403 lwkt_create(void (*func)(void *), void *arg,
1404     struct thread **tdp, thread_t template, int tdflags, int cpu,
1405     const char *fmt, ...)
1406 {
1407     thread_t td;
1408     __va_list ap;
1409 
1410     td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1411 			   tdflags);
1412     if (tdp)
1413 	*tdp = td;
1414     cpu_set_thread_handler(td, lwkt_exit, func, arg);
1415 
1416     /*
1417      * Set up arg0 for 'ps' etc
1418      */
1419     __va_start(ap, fmt);
1420     kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1421     __va_end(ap);
1422 
1423     /*
1424      * Schedule the thread to run
1425      */
1426     if ((td->td_flags & TDF_STOPREQ) == 0)
1427 	lwkt_schedule(td);
1428     else
1429 	td->td_flags &= ~TDF_STOPREQ;
1430     return 0;
1431 }
1432 
1433 /*
1434  * Destroy an LWKT thread.   Warning!  This function is not called when
1435  * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1436  * uses a different reaping mechanism.
1437  */
1438 void
1439 lwkt_exit(void)
1440 {
1441     thread_t td = curthread;
1442     thread_t std;
1443     globaldata_t gd;
1444 
1445     if (td->td_flags & TDF_VERBOSE)
1446 	kprintf("kthread %p %s has exited\n", td, td->td_comm);
1447     caps_exit(td);
1448 
1449     /*
1450      * Get us into a critical section to interlock gd_freetd and loop
1451      * until we can get it freed.
1452      *
1453      * We have to cache the current td in gd_freetd because objcache_put()ing
1454      * it would rip it out from under us while our thread is still active.
1455      */
1456     gd = mycpu;
1457     crit_enter_quick(td);
1458     while ((std = gd->gd_freetd) != NULL) {
1459 	gd->gd_freetd = NULL;
1460 	objcache_put(thread_cache, std);
1461     }
1462 
1463     /*
1464      * Remove thread resources from kernel lists and deschedule us for
1465      * the last time.
1466      */
1467     if (td->td_flags & TDF_TSLEEPQ)
1468 	tsleep_remove(td);
1469     biosched_done(td);
1470     lwkt_deschedule_self(td);
1471     lwkt_remove_tdallq(td);
1472     if (td->td_flags & TDF_ALLOCATED_THREAD)
1473 	gd->gd_freetd = td;
1474     cpu_thread_exit();
1475 }
1476 
1477 void
1478 lwkt_remove_tdallq(thread_t td)
1479 {
1480     KKASSERT(td->td_gd == mycpu);
1481     TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1482 }
1483 
1484 void
1485 crit_panic(void)
1486 {
1487     thread_t td = curthread;
1488     int lpri = td->td_pri;
1489 
1490     td->td_pri = 0;
1491     panic("td_pri is/would-go negative! %p %d", td, lpri);
1492 }
1493 
1494 #ifdef SMP
1495 
1496 /*
1497  * Called from debugger/panic on cpus which have been stopped.  We must still
1498  * process the IPIQ while stopped, even if we were stopped while in a critical
1499  * section (XXX).
1500  *
1501  * If we are dumping also try to process any pending interrupts.  This may
1502  * or may not work depending on the state of the cpu at the point it was
1503  * stopped.
1504  */
1505 void
1506 lwkt_smp_stopped(void)
1507 {
1508     globaldata_t gd = mycpu;
1509 
1510     crit_enter_gd(gd);
1511     if (dumping) {
1512 	lwkt_process_ipiq();
1513 	splz();
1514     } else {
1515 	lwkt_process_ipiq();
1516     }
1517     crit_exit_gd(gd);
1518 }
1519 
1520 #endif
1521