xref: /dflybsd-src/sys/kern/lwkt_thread.c (revision d4b8aec4bb44a374c3e91969c1a7df6569da7be3)
1 /*
2  * Copyright (c) 2003-2010 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/kinfo.h>
48 #include <sys/queue.h>
49 #include <sys/sysctl.h>
50 #include <sys/kthread.h>
51 #include <machine/cpu.h>
52 #include <sys/lock.h>
53 #include <sys/caps.h>
54 #include <sys/spinlock.h>
55 #include <sys/ktr.h>
56 
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mplock2.h>
60 
61 #include <sys/dsched.h>
62 
63 #include <vm/vm.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
71 
72 #include <machine/stdarg.h>
73 #include <machine/smp.h>
74 
75 #if !defined(KTR_CTXSW)
76 #define KTR_CTXSW KTR_ALL
77 #endif
78 KTR_INFO_MASTER(ctxsw);
79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 	 sizeof(int) + sizeof(struct thread *));
81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 	 sizeof(int) + sizeof(struct thread *));
83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 	 sizeof (struct thread *) + sizeof(char *));
85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
86 
87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88 
89 #ifdef	INVARIANTS
90 static int panic_on_cscount = 0;
91 #endif
92 static __int64_t switch_count = 0;
93 static __int64_t preempt_hit = 0;
94 static __int64_t preempt_miss = 0;
95 static __int64_t preempt_weird = 0;
96 static __int64_t token_contention_count __debugvar = 0;
97 static int lwkt_use_spin_port;
98 static struct objcache *thread_cache;
99 
100 #ifdef SMP
101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
102 #endif
103 static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
104 
105 extern void cpu_heavy_restore(void);
106 extern void cpu_lwkt_restore(void);
107 extern void cpu_kthread_restore(void);
108 extern void cpu_idle_restore(void);
109 
110 #ifdef __x86_64__
111 
112 static int
113 jg_tos_ok(struct thread *td)
114 {
115 	void *tos;
116 	int tos_ok;
117 
118 	if (td == NULL) {
119 		return 1;
120 	}
121 	KKASSERT(td->td_sp != NULL);
122 	tos = ((void **)td->td_sp)[0];
123 	tos_ok = 0;
124 	if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 	    (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
126 		tos_ok = 1;
127 	}
128 	return tos_ok;
129 }
130 
131 #endif
132 
133 /*
134  * We can make all thread ports use the spin backend instead of the thread
135  * backend.  This should only be set to debug the spin backend.
136  */
137 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138 
139 #ifdef	INVARIANTS
140 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141 #endif
142 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 	    "Successful preemption events");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 	    "Failed preemption events");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
148 #ifdef	INVARIANTS
149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 	&token_contention_count, 0, "spinning due to token contention");
151 #endif
152 static int fairq_enable = 1;
153 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
154 
155 /*
156  * These helper procedures handle the runq, they can only be called from
157  * within a critical section.
158  *
159  * WARNING!  Prior to SMP being brought up it is possible to enqueue and
160  * dequeue threads belonging to other cpus, so be sure to use td->td_gd
161  * instead of 'mycpu' when referencing the globaldata structure.   Once
162  * SMP live enqueuing and dequeueing only occurs on the current cpu.
163  */
164 static __inline
165 void
166 _lwkt_dequeue(thread_t td)
167 {
168     if (td->td_flags & TDF_RUNQ) {
169 	struct globaldata *gd = td->td_gd;
170 
171 	td->td_flags &= ~TDF_RUNQ;
172 	TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
173 	gd->gd_fairq_total_pri -= td->td_pri;
174 	if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
175 		atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
176     }
177 }
178 
179 /*
180  * Priority enqueue.
181  *
182  * NOTE: There are a limited number of lwkt threads runnable since user
183  *	 processes only schedule one at a time per cpu.
184  */
185 static __inline
186 void
187 _lwkt_enqueue(thread_t td)
188 {
189     thread_t xtd;
190 
191     if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
192 	struct globaldata *gd = td->td_gd;
193 
194 	td->td_flags |= TDF_RUNQ;
195 	xtd = TAILQ_FIRST(&gd->gd_tdrunq);
196 	if (xtd == NULL) {
197 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
198 		atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
199 	} else {
200 		while (xtd && xtd->td_pri > td->td_pri)
201 			xtd = TAILQ_NEXT(xtd, td_threadq);
202 		if (xtd)
203 			TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
204 		else
205 			TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 	}
207 	gd->gd_fairq_total_pri += td->td_pri;
208     }
209 }
210 
211 static __boolean_t
212 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
213 {
214 	struct thread *td = (struct thread *)obj;
215 
216 	td->td_kstack = NULL;
217 	td->td_kstack_size = 0;
218 	td->td_flags = TDF_ALLOCATED_THREAD;
219 	return (1);
220 }
221 
222 static void
223 _lwkt_thread_dtor(void *obj, void *privdata)
224 {
225 	struct thread *td = (struct thread *)obj;
226 
227 	KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
228 	    ("_lwkt_thread_dtor: not allocated from objcache"));
229 	KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
230 		td->td_kstack_size > 0,
231 	    ("_lwkt_thread_dtor: corrupted stack"));
232 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
233 }
234 
235 /*
236  * Initialize the lwkt s/system.
237  */
238 void
239 lwkt_init(void)
240 {
241     /* An objcache has 2 magazines per CPU so divide cache size by 2. */
242     thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
243 			NULL, CACHE_NTHREADS/2,
244 			_lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
245 }
246 
247 /*
248  * Schedule a thread to run.  As the current thread we can always safely
249  * schedule ourselves, and a shortcut procedure is provided for that
250  * function.
251  *
252  * (non-blocking, self contained on a per cpu basis)
253  */
254 void
255 lwkt_schedule_self(thread_t td)
256 {
257     crit_enter_quick(td);
258     KASSERT(td != &td->td_gd->gd_idlethread,
259 	    ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
260     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
261     _lwkt_enqueue(td);
262     crit_exit_quick(td);
263 }
264 
265 /*
266  * Deschedule a thread.
267  *
268  * (non-blocking, self contained on a per cpu basis)
269  */
270 void
271 lwkt_deschedule_self(thread_t td)
272 {
273     crit_enter_quick(td);
274     _lwkt_dequeue(td);
275     crit_exit_quick(td);
276 }
277 
278 /*
279  * LWKTs operate on a per-cpu basis
280  *
281  * WARNING!  Called from early boot, 'mycpu' may not work yet.
282  */
283 void
284 lwkt_gdinit(struct globaldata *gd)
285 {
286     TAILQ_INIT(&gd->gd_tdrunq);
287     TAILQ_INIT(&gd->gd_tdallq);
288 }
289 
290 /*
291  * Create a new thread.  The thread must be associated with a process context
292  * or LWKT start address before it can be scheduled.  If the target cpu is
293  * -1 the thread will be created on the current cpu.
294  *
295  * If you intend to create a thread without a process context this function
296  * does everything except load the startup and switcher function.
297  */
298 thread_t
299 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
300 {
301     globaldata_t gd = mycpu;
302     void *stack;
303 
304     /*
305      * If static thread storage is not supplied allocate a thread.  Reuse
306      * a cached free thread if possible.  gd_freetd is used to keep an exiting
307      * thread intact through the exit.
308      */
309     if (td == NULL) {
310 	if ((td = gd->gd_freetd) != NULL)
311 	    gd->gd_freetd = NULL;
312 	else
313 	    td = objcache_get(thread_cache, M_WAITOK);
314     	KASSERT((td->td_flags &
315 		 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
316 		("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
317     	flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
318     }
319 
320     /*
321      * Try to reuse cached stack.
322      */
323     if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
324 	if (flags & TDF_ALLOCATED_STACK) {
325 	    kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
326 	    stack = NULL;
327 	}
328     }
329     if (stack == NULL) {
330 	stack = (void *)kmem_alloc(&kernel_map, stksize);
331 	flags |= TDF_ALLOCATED_STACK;
332     }
333     if (cpu < 0)
334 	lwkt_init_thread(td, stack, stksize, flags, gd);
335     else
336 	lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
337     return(td);
338 }
339 
340 /*
341  * Initialize a preexisting thread structure.  This function is used by
342  * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
343  *
344  * All threads start out in a critical section at a priority of
345  * TDPRI_KERN_DAEMON.  Higher level code will modify the priority as
346  * appropriate.  This function may send an IPI message when the
347  * requested cpu is not the current cpu and consequently gd_tdallq may
348  * not be initialized synchronously from the point of view of the originating
349  * cpu.
350  *
351  * NOTE! we have to be careful in regards to creating threads for other cpus
352  * if SMP has not yet been activated.
353  */
354 #ifdef SMP
355 
356 static void
357 lwkt_init_thread_remote(void *arg)
358 {
359     thread_t td = arg;
360 
361     /*
362      * Protected by critical section held by IPI dispatch
363      */
364     TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
365 }
366 
367 #endif
368 
369 void
370 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
371 		struct globaldata *gd)
372 {
373     globaldata_t mygd = mycpu;
374 
375     bzero(td, sizeof(struct thread));
376     td->td_kstack = stack;
377     td->td_kstack_size = stksize;
378     td->td_flags = flags;
379     td->td_gd = gd;
380     td->td_pri = TDPRI_KERN_DAEMON;
381     td->td_critcount = 1;
382     td->td_toks_stop = &td->td_toks_base;
383 #ifdef SMP
384     if ((flags & TDF_MPSAFE) == 0)
385 	td->td_mpcount = 1;
386 #endif
387     if (lwkt_use_spin_port)
388 	lwkt_initport_spin(&td->td_msgport);
389     else
390 	lwkt_initport_thread(&td->td_msgport, td);
391     pmap_init_thread(td);
392 #ifdef SMP
393     /*
394      * Normally initializing a thread for a remote cpu requires sending an
395      * IPI.  However, the idlethread is setup before the other cpus are
396      * activated so we have to treat it as a special case.  XXX manipulation
397      * of gd_tdallq requires the BGL.
398      */
399     if (gd == mygd || td == &gd->gd_idlethread) {
400 	crit_enter_gd(mygd);
401 	TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
402 	crit_exit_gd(mygd);
403     } else {
404 	lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
405     }
406 #else
407     crit_enter_gd(mygd);
408     TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
409     crit_exit_gd(mygd);
410 #endif
411 
412     dsched_new_thread(td);
413 }
414 
415 void
416 lwkt_set_comm(thread_t td, const char *ctl, ...)
417 {
418     __va_list va;
419 
420     __va_start(va, ctl);
421     kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
422     __va_end(va);
423     KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
424 }
425 
426 void
427 lwkt_hold(thread_t td)
428 {
429     ++td->td_refs;
430 }
431 
432 void
433 lwkt_rele(thread_t td)
434 {
435     KKASSERT(td->td_refs > 0);
436     --td->td_refs;
437 }
438 
439 void
440 lwkt_wait_free(thread_t td)
441 {
442     while (td->td_refs)
443 	tsleep(td, 0, "tdreap", hz);
444 }
445 
446 void
447 lwkt_free_thread(thread_t td)
448 {
449     KASSERT((td->td_flags & TDF_RUNNING) == 0,
450 	("lwkt_free_thread: did not exit! %p", td));
451 
452     if (td->td_flags & TDF_ALLOCATED_THREAD) {
453     	objcache_put(thread_cache, td);
454     } else if (td->td_flags & TDF_ALLOCATED_STACK) {
455 	/* client-allocated struct with internally allocated stack */
456 	KASSERT(td->td_kstack && td->td_kstack_size > 0,
457 	    ("lwkt_free_thread: corrupted stack"));
458 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
459 	td->td_kstack = NULL;
460 	td->td_kstack_size = 0;
461     }
462     KTR_LOG(ctxsw_deadtd, td);
463 }
464 
465 
466 /*
467  * Switch to the next runnable lwkt.  If no LWKTs are runnable then
468  * switch to the idlethread.  Switching must occur within a critical
469  * section to avoid races with the scheduling queue.
470  *
471  * We always have full control over our cpu's run queue.  Other cpus
472  * that wish to manipulate our queue must use the cpu_*msg() calls to
473  * talk to our cpu, so a critical section is all that is needed and
474  * the result is very, very fast thread switching.
475  *
476  * The LWKT scheduler uses a fixed priority model and round-robins at
477  * each priority level.  User process scheduling is a totally
478  * different beast and LWKT priorities should not be confused with
479  * user process priorities.
480  *
481  * The MP lock may be out of sync with the thread's td_mpcount.  lwkt_switch()
482  * cleans it up.  Note that the td_switch() function cannot do anything that
483  * requires the MP lock since the MP lock will have already been setup for
484  * the target thread (not the current thread).  It's nice to have a scheduler
485  * that does not need the MP lock to work because it allows us to do some
486  * really cool high-performance MP lock optimizations.
487  *
488  * PREEMPTION NOTE: Preemption occurs via lwkt_preempt().  lwkt_switch()
489  * is not called by the current thread in the preemption case, only when
490  * the preempting thread blocks (in order to return to the original thread).
491  */
492 void
493 lwkt_switch(void)
494 {
495     globaldata_t gd = mycpu;
496     thread_t td = gd->gd_curthread;
497     thread_t ntd;
498     thread_t xtd;
499     thread_t nlast;
500     int nquserok;
501 #ifdef SMP
502     int mpheld;
503 #endif
504     int didaccumulate;
505     const char *lmsg;	/* diagnostic - 'systat -pv 1' */
506     const void *laddr;
507 
508     /*
509      * Switching from within a 'fast' (non thread switched) interrupt or IPI
510      * is illegal.  However, we may have to do it anyway if we hit a fatal
511      * kernel trap or we have paniced.
512      *
513      * If this case occurs save and restore the interrupt nesting level.
514      */
515     if (gd->gd_intr_nesting_level) {
516 	int savegdnest;
517 	int savegdtrap;
518 
519 	if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
520 	    panic("lwkt_switch: cannot switch from within "
521 		  "a fast interrupt, yet, td %p\n", td);
522 	} else {
523 	    savegdnest = gd->gd_intr_nesting_level;
524 	    savegdtrap = gd->gd_trap_nesting_level;
525 	    gd->gd_intr_nesting_level = 0;
526 	    gd->gd_trap_nesting_level = 0;
527 	    if ((td->td_flags & TDF_PANICWARN) == 0) {
528 		td->td_flags |= TDF_PANICWARN;
529 		kprintf("Warning: thread switch from interrupt or IPI, "
530 			"thread %p (%s)\n", td, td->td_comm);
531 		print_backtrace(-1);
532 	    }
533 	    lwkt_switch();
534 	    gd->gd_intr_nesting_level = savegdnest;
535 	    gd->gd_trap_nesting_level = savegdtrap;
536 	    return;
537 	}
538     }
539 
540     /*
541      * Passive release (used to transition from user to kernel mode
542      * when we block or switch rather then when we enter the kernel).
543      * This function is NOT called if we are switching into a preemption
544      * or returning from a preemption.  Typically this causes us to lose
545      * our current process designation (if we have one) and become a true
546      * LWKT thread, and may also hand the current process designation to
547      * another process and schedule thread.
548      */
549     if (td->td_release)
550 	    td->td_release(td);
551 
552     crit_enter_gd(gd);
553     if (TD_TOKS_HELD(td))
554 	    lwkt_relalltokens(td);
555 
556     /*
557      * We had better not be holding any spin locks, but don't get into an
558      * endless panic loop.
559      */
560     KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
561 	    ("lwkt_switch: still holding a shared spinlock %p!",
562 	     gd->gd_spinlock_rd));
563     KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
564 	    ("lwkt_switch: still holding %d exclusive spinlocks!",
565 	     gd->gd_spinlocks_wr));
566 
567 
568 #ifdef SMP
569     /*
570      * td_mpcount cannot be used to determine if we currently hold the
571      * MP lock because get_mplock() will increment it prior to attempting
572      * to get the lock, and switch out if it can't.  Our ownership of
573      * the actual lock will remain stable while we are in a critical section
574      * (but, of course, another cpu may own or release the lock so the
575      * actual value of mp_lock is not stable).
576      */
577     mpheld = MP_LOCK_HELD();
578 #ifdef	INVARIANTS
579     if (td->td_cscount) {
580 	kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
581 		td);
582 	if (panic_on_cscount)
583 	    panic("switching while mastering cpusync");
584     }
585 #endif
586 #endif
587 
588     /*
589      * If we had preempted another thread on this cpu, resume the preempted
590      * thread.  This occurs transparently, whether the preempted thread
591      * was scheduled or not (it may have been preempted after descheduling
592      * itself).
593      *
594      * We have to setup the MP lock for the original thread after backing
595      * out the adjustment that was made to curthread when the original
596      * was preempted.
597      */
598     if ((ntd = td->td_preempted) != NULL) {
599 	KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
600 #ifdef SMP
601 	if (ntd->td_mpcount && mpheld == 0) {
602 	    panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
603 	       td, ntd, td->td_mpcount, ntd->td_mpcount);
604 	}
605 	if (ntd->td_mpcount) {
606 	    td->td_mpcount -= ntd->td_mpcount;
607 	    KKASSERT(td->td_mpcount >= 0);
608 	}
609 #endif
610 	ntd->td_flags |= TDF_PREEMPT_DONE;
611 
612 	/*
613 	 * The interrupt may have woken a thread up, we need to properly
614 	 * set the reschedule flag if the originally interrupted thread is
615 	 * at a lower priority.
616 	 */
617 	if (TAILQ_FIRST(&gd->gd_tdrunq) &&
618 	    TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
619 	    need_lwkt_resched();
620 	}
621 	/* YYY release mp lock on switchback if original doesn't need it */
622 	goto havethread_preempted;
623     }
624 
625     /*
626      * Implement round-robin fairq with priority insertion.  The priority
627      * insertion is handled by _lwkt_enqueue()
628      *
629      * We have to adjust the MP lock for the target thread.  If we
630      * need the MP lock and cannot obtain it we try to locate a
631      * thread that does not need the MP lock.  If we cannot, we spin
632      * instead of HLT.
633      *
634      * A similar issue exists for the tokens held by the target thread.
635      * If we cannot obtain ownership of the tokens we cannot immediately
636      * schedule the thread.
637      */
638     for (;;) {
639 	clear_lwkt_resched();
640 	didaccumulate = 0;
641 	ntd = TAILQ_FIRST(&gd->gd_tdrunq);
642 
643 	/*
644 	 * Hotpath if we can get all necessary resources.
645 	 *
646 	 * If nothing is runnable switch to the idle thread
647 	 */
648 	if (ntd == NULL) {
649 	    ntd = &gd->gd_idlethread;
650 	    if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
651 		    ntd->td_flags |= TDF_IDLE_NOHLT;
652 #ifdef SMP
653 	    if (ntd->td_mpcount) {
654 		if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
655 		    panic("Idle thread %p was holding the BGL!", ntd);
656 		if (mpheld == 0) {
657 		    cpu_pause();
658 		    continue;
659 		}
660 	    }
661 #endif
662 	    cpu_time.cp_msg[0] = 0;
663 	    cpu_time.cp_stallpc = 0;
664 	    goto haveidle;
665 	}
666 
667 	/*
668 	 * Hotpath schedule
669 	 *
670 	 * NOTE: For UP there is no mplock and lwkt_getalltokens()
671 	 *	     always succeeds.
672 	 */
673 	if (ntd->td_fairq_accum >= 0 &&
674 #ifdef SMP
675 	    (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
676 #endif
677 	    (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
678 	) {
679 #ifdef SMP
680 	    clr_mplock_contention_mask(gd);
681 #endif
682 	    goto havethread;
683 	}
684 
685 	lmsg = NULL;
686 	laddr = NULL;
687 
688 #ifdef SMP
689 	/* Reload mpheld (it become stale after mplock/token ops) */
690 	mpheld = MP_LOCK_HELD();
691 	if (ntd->td_mpcount && mpheld == 0) {
692 	    lmsg = "mplock";
693 	    laddr = ntd->td_mplock_stallpc;
694 	}
695 #endif
696 
697 	/*
698 	 * Coldpath - unable to schedule ntd, continue looking for threads
699 	 * to schedule.  This is only allowed of the (presumably) kernel
700 	 * thread exhausted its fair share.  A kernel thread stuck on
701 	 * resources does not currently allow a user thread to get in
702 	 * front of it.
703 	 */
704 #ifdef SMP
705 	nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
706 		    (ntd->td_fairq_accum < 0));
707 #else
708 	nquserok = 1;
709 #endif
710 	nlast = NULL;
711 
712 	for (;;) {
713 	    /*
714 	     * If the fair-share scheduler ran out ntd gets moved to the
715 	     * end and its accumulator will be bumped, if it didn't we
716 	     * maintain the same queue position.
717 	     *
718 	     * nlast keeps track of the last element prior to any moves.
719 	     */
720 	    if (ntd->td_fairq_accum < 0) {
721 		xtd = TAILQ_NEXT(ntd, td_threadq);
722 		lwkt_fairq_accumulate(gd, ntd);
723 		didaccumulate = 1;
724 		TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
725 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
726 		if (nlast == NULL) {
727 		    nlast = ntd;
728 		    if (xtd == NULL)
729 			xtd = ntd;
730 		}
731 		ntd = xtd;
732 	    } else {
733 		ntd = TAILQ_NEXT(ntd, td_threadq);
734 	    }
735 
736 	    /*
737 	     * If we exhausted the run list switch to the idle thread.
738 	     * Since one or more threads had resource acquisition issues
739 	     * we do not allow the idle thread to halt.
740 	     *
741 	     * NOTE: nlast can be NULL.
742 	     */
743 	    if (ntd == nlast) {
744 		cpu_pause();
745 		ntd = &gd->gd_idlethread;
746 		ntd->td_flags |= TDF_IDLE_NOHLT;
747 #ifdef SMP
748 		set_mplock_contention_mask(gd);
749 		cpu_mplock_contested();
750 		if (ntd->td_mpcount) {
751 		    mpheld = MP_LOCK_HELD();
752 		    if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
753 			panic("Idle thread %p was holding the BGL!", ntd);
754 		    if (mpheld == 0) {
755 			cpu_pause();
756 			break;		/* try again from the top, almost */
757 		    }
758 		}
759 #endif
760 
761 		/*
762 		 * If fairq accumulations occured we do not schedule the
763 		 * idle thread.  This will cause us to try again from
764 		 * the (almost) top.
765 		 */
766 		if (didaccumulate)
767 			break;		/* try again from the top, almost */
768 		if (lmsg)
769 		    strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
770 		cpu_time.cp_stallpc = (uintptr_t)laddr;
771 		goto haveidle;
772 	    }
773 
774 	    /*
775 	     * Try to switch to this thread.
776 	     *
777 	     * NOTE: For UP there is no mplock and lwkt_getalltokens()
778 	     *	     always succeeds.
779 	     */
780 	    if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok) &&
781 		ntd->td_fairq_accum >= 0 &&
782 #ifdef SMP
783 		(ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
784 #endif
785 		(!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
786 	    ) {
787 #ifdef SMP
788 		    clr_mplock_contention_mask(gd);
789 #endif
790 		    goto havethread;
791 	    }
792 #ifdef SMP
793 	    /* Reload mpheld (it become stale after mplock/token ops) */
794 	    mpheld = MP_LOCK_HELD();
795 	    if (ntd->td_mpcount && mpheld == 0) {
796 		lmsg = "mplock";
797 		laddr = ntd->td_mplock_stallpc;
798 	    }
799 
800 	    if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
801 		nquserok = 0;
802 #endif
803 	}
804     }
805 
806     /*
807      * Do the actual switch.  WARNING: mpheld is stale here.
808      *
809      * We must always decrement td_fairq_accum on non-idle threads just
810      * in case a thread never gets a tick due to being in a continuous
811      * critical section.  The page-zeroing code does that.
812      *
813      * If the thread we came up with is a higher or equal priority verses
814      * the thread at the head of the queue we move our thread to the
815      * front.  This way we can always check the front of the queue.
816      */
817 havethread:
818     ++gd->gd_cnt.v_swtch;
819     --ntd->td_fairq_accum;
820     xtd = TAILQ_FIRST(&gd->gd_tdrunq);
821     if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
822 	TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
823 	TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
824     }
825 havethread_preempted:
826     ;
827     /*
828      * If the new target does not need the MP lock and we are holding it,
829      * release the MP lock.  If the new target requires the MP lock we have
830      * already acquired it for the target.
831      *
832      * WARNING: mpheld is stale here.
833      */
834 haveidle:
835     KASSERT(ntd->td_critcount,
836 	    ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
837 #ifdef SMP
838     if (ntd->td_mpcount == 0 ) {
839 	if (MP_LOCK_HELD())
840 	    cpu_rel_mplock();
841     } else {
842 	ASSERT_MP_LOCK_HELD(ntd);
843     }
844 #endif
845     if (td != ntd) {
846 	++switch_count;
847 #ifdef __x86_64__
848 	{
849 	    int tos_ok __debugvar = jg_tos_ok(ntd);
850 	    KKASSERT(tos_ok);
851 	}
852 #endif
853 	KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
854 	td->td_switch(ntd);
855     }
856     /* NOTE: current cpu may have changed after switch */
857     crit_exit_quick(td);
858 }
859 
860 /*
861  * Request that the target thread preempt the current thread.  Preemption
862  * only works under a specific set of conditions:
863  *
864  *	- We are not preempting ourselves
865  *	- The target thread is owned by the current cpu
866  *	- We are not currently being preempted
867  *	- The target is not currently being preempted
868  *	- We are not holding any spin locks
869  *	- The target thread is not holding any tokens
870  *	- We are able to satisfy the target's MP lock requirements (if any).
871  *
872  * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION.  Typically
873  * this is called via lwkt_schedule() through the td_preemptable callback.
874  * critcount is the managed critical priority that we should ignore in order
875  * to determine whether preemption is possible (aka usually just the crit
876  * priority of lwkt_schedule() itself).
877  *
878  * XXX at the moment we run the target thread in a critical section during
879  * the preemption in order to prevent the target from taking interrupts
880  * that *WE* can't.  Preemption is strictly limited to interrupt threads
881  * and interrupt-like threads, outside of a critical section, and the
882  * preempted source thread will be resumed the instant the target blocks
883  * whether or not the source is scheduled (i.e. preemption is supposed to
884  * be as transparent as possible).
885  *
886  * The target thread inherits our MP count (added to its own) for the
887  * duration of the preemption in order to preserve the atomicy of the
888  * MP lock during the preemption.  Therefore, any preempting targets must be
889  * careful in regards to MP assertions.  Note that the MP count may be
890  * out of sync with the physical mp_lock, but we do not have to preserve
891  * the original ownership of the lock if it was out of synch (that is, we
892  * can leave it synchronized on return).
893  */
894 void
895 lwkt_preempt(thread_t ntd, int critcount)
896 {
897     struct globaldata *gd = mycpu;
898     thread_t td;
899 #ifdef SMP
900     int mpheld;
901     int savecnt;
902 #endif
903 
904     /*
905      * The caller has put us in a critical section.  We can only preempt
906      * if the caller of the caller was not in a critical section (basically
907      * a local interrupt), as determined by the 'critcount' parameter.  We
908      * also can't preempt if the caller is holding any spinlocks (even if
909      * he isn't in a critical section).  This also handles the tokens test.
910      *
911      * YYY The target thread must be in a critical section (else it must
912      * inherit our critical section?  I dunno yet).
913      *
914      * Set need_lwkt_resched() unconditionally for now YYY.
915      */
916     KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
917 
918     td = gd->gd_curthread;
919     if (ntd->td_pri <= td->td_pri) {
920 	++preempt_miss;
921 	return;
922     }
923     if (td->td_critcount > critcount) {
924 	++preempt_miss;
925 	need_lwkt_resched();
926 	return;
927     }
928 #ifdef SMP
929     if (ntd->td_gd != gd) {
930 	++preempt_miss;
931 	need_lwkt_resched();
932 	return;
933     }
934 #endif
935     /*
936      * Take the easy way out and do not preempt if we are holding
937      * any spinlocks.  We could test whether the thread(s) being
938      * preempted interlock against the target thread's tokens and whether
939      * we can get all the target thread's tokens, but this situation
940      * should not occur very often so its easier to simply not preempt.
941      * Also, plain spinlocks are impossible to figure out at this point so
942      * just don't preempt.
943      *
944      * Do not try to preempt if the target thread is holding any tokens.
945      * We could try to acquire the tokens but this case is so rare there
946      * is no need to support it.
947      */
948     if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
949 	++preempt_miss;
950 	need_lwkt_resched();
951 	return;
952     }
953     if (TD_TOKS_HELD(ntd)) {
954 	++preempt_miss;
955 	need_lwkt_resched();
956 	return;
957     }
958     if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
959 	++preempt_weird;
960 	need_lwkt_resched();
961 	return;
962     }
963     if (ntd->td_preempted) {
964 	++preempt_hit;
965 	need_lwkt_resched();
966 	return;
967     }
968 #ifdef SMP
969     /*
970      * note: an interrupt might have occured just as we were transitioning
971      * to or from the MP lock.  In this case td_mpcount will be pre-disposed
972      * (non-zero) but not actually synchronized with the actual state of the
973      * lock.  We can use it to imply an MP lock requirement for the
974      * preemption but we cannot use it to test whether we hold the MP lock
975      * or not.
976      */
977     savecnt = td->td_mpcount;
978     mpheld = MP_LOCK_HELD();
979     ntd->td_mpcount += td->td_mpcount;
980     if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
981 	ntd->td_mpcount -= td->td_mpcount;
982 	++preempt_miss;
983 	need_lwkt_resched();
984 	return;
985     }
986 #endif
987 
988     /*
989      * Since we are able to preempt the current thread, there is no need to
990      * call need_lwkt_resched().
991      */
992     ++preempt_hit;
993     ntd->td_preempted = td;
994     td->td_flags |= TDF_PREEMPT_LOCK;
995     KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
996     td->td_switch(ntd);
997 
998     KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
999 #ifdef SMP
1000     KKASSERT(savecnt == td->td_mpcount);
1001     mpheld = MP_LOCK_HELD();
1002     if (mpheld && td->td_mpcount == 0)
1003 	cpu_rel_mplock();
1004     else if (mpheld == 0 && td->td_mpcount)
1005 	panic("lwkt_preempt(): MP lock was not held through");
1006 #endif
1007     ntd->td_preempted = NULL;
1008     td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1009 }
1010 
1011 /*
1012  * Conditionally call splz() if gd_reqflags indicates work is pending.
1013  *
1014  * td_nest_count prevents deep nesting via splz() or doreti() which
1015  * might otherwise blow out the kernel stack.  Note that except for
1016  * this special case, we MUST call splz() here to handle any
1017  * pending ints, particularly after we switch, or we might accidently
1018  * halt the cpu with interrupts pending.
1019  *
1020  * (self contained on a per cpu basis)
1021  */
1022 void
1023 splz_check(void)
1024 {
1025     globaldata_t gd = mycpu;
1026     thread_t td = gd->gd_curthread;
1027 
1028     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1029 	splz();
1030 }
1031 
1032 /*
1033  * This function is used to negotiate a passive release of the current
1034  * process/lwp designation with the user scheduler, allowing the user
1035  * scheduler to schedule another user thread.  The related kernel thread
1036  * (curthread) continues running in the released state.
1037  */
1038 void
1039 lwkt_passive_release(struct thread *td)
1040 {
1041     struct lwp *lp = td->td_lwp;
1042 
1043     td->td_release = NULL;
1044     lwkt_setpri_self(TDPRI_KERN_USER);
1045     lp->lwp_proc->p_usched->release_curproc(lp);
1046 }
1047 
1048 
1049 /*
1050  * This implements a normal yield.  This routine is virtually a nop if
1051  * there is nothing to yield to but it will always run any pending interrupts
1052  * if called from a critical section.
1053  *
1054  * This yield is designed for kernel threads without a user context.
1055  *
1056  * (self contained on a per cpu basis)
1057  */
1058 void
1059 lwkt_yield(void)
1060 {
1061     globaldata_t gd = mycpu;
1062     thread_t td = gd->gd_curthread;
1063     thread_t xtd;
1064 
1065     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1066 	splz();
1067     if (td->td_fairq_accum < 0) {
1068 	lwkt_schedule_self(curthread);
1069 	lwkt_switch();
1070     } else {
1071 	xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1072 	if (xtd && xtd->td_pri > td->td_pri) {
1073 	    lwkt_schedule_self(curthread);
1074 	    lwkt_switch();
1075 	}
1076     }
1077 }
1078 
1079 /*
1080  * This yield is designed for kernel threads with a user context.
1081  *
1082  * The kernel acting on behalf of the user is potentially cpu-bound,
1083  * this function will efficiently allow other threads to run and also
1084  * switch to other processes by releasing.
1085  *
1086  * The lwkt_user_yield() function is designed to have very low overhead
1087  * if no yield is determined to be needed.
1088  */
1089 void
1090 lwkt_user_yield(void)
1091 {
1092     globaldata_t gd = mycpu;
1093     thread_t td = gd->gd_curthread;
1094 
1095     /*
1096      * Always run any pending interrupts in case we are in a critical
1097      * section.
1098      */
1099     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1100 	splz();
1101 
1102 #ifdef SMP
1103     /*
1104      * XXX SEVERE TEMPORARY HACK.  A cpu-bound operation running in the
1105      * kernel can prevent other cpus from servicing interrupt threads
1106      * which still require the MP lock (which is a lot of them).  This
1107      * has a chaining effect since if the interrupt is blocked, so is
1108      * the event, so normal scheduling will not pick up on the problem.
1109      */
1110     if (mp_lock_contention_mask && td->td_mpcount) {
1111 	yield_mplock(td);
1112     }
1113 #endif
1114 
1115     /*
1116      * Switch (which forces a release) if another kernel thread needs
1117      * the cpu, if userland wants us to resched, or if our kernel
1118      * quantum has run out.
1119      */
1120     if (lwkt_resched_wanted() ||
1121 	user_resched_wanted() ||
1122 	td->td_fairq_accum < 0)
1123     {
1124 	lwkt_switch();
1125     }
1126 
1127 #if 0
1128     /*
1129      * Reacquire the current process if we are released.
1130      *
1131      * XXX not implemented atm.  The kernel may be holding locks and such,
1132      *     so we want the thread to continue to receive cpu.
1133      */
1134     if (td->td_release == NULL && lp) {
1135 	lp->lwp_proc->p_usched->acquire_curproc(lp);
1136 	td->td_release = lwkt_passive_release;
1137 	lwkt_setpri_self(TDPRI_USER_NORM);
1138     }
1139 #endif
1140 }
1141 
1142 /*
1143  * Generic schedule.  Possibly schedule threads belonging to other cpus and
1144  * deal with threads that might be blocked on a wait queue.
1145  *
1146  * We have a little helper inline function which does additional work after
1147  * the thread has been enqueued, including dealing with preemption and
1148  * setting need_lwkt_resched() (which prevents the kernel from returning
1149  * to userland until it has processed higher priority threads).
1150  *
1151  * It is possible for this routine to be called after a failed _enqueue
1152  * (due to the target thread migrating, sleeping, or otherwise blocked).
1153  * We have to check that the thread is actually on the run queue!
1154  *
1155  * reschedok is an optimized constant propagated from lwkt_schedule() or
1156  * lwkt_schedule_noresched().  By default it is non-zero, causing a
1157  * reschedule to be requested if the target thread has a higher priority.
1158  * The port messaging code will set MSG_NORESCHED and cause reschedok to
1159  * be 0, prevented undesired reschedules.
1160  */
1161 static __inline
1162 void
1163 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1164 {
1165     thread_t otd;
1166 
1167     if (ntd->td_flags & TDF_RUNQ) {
1168 	if (ntd->td_preemptable && reschedok) {
1169 	    ntd->td_preemptable(ntd, ccount);	/* YYY +token */
1170 	} else if (reschedok) {
1171 	    otd = curthread;
1172 	    if (ntd->td_pri > otd->td_pri)
1173 		need_lwkt_resched();
1174 	}
1175 
1176 	/*
1177 	 * Give the thread a little fair share scheduler bump if it
1178 	 * has been asleep for a while.  This is primarily to avoid
1179 	 * a degenerate case for interrupt threads where accumulator
1180 	 * crosses into negative territory unnecessarily.
1181 	 */
1182 	if (ntd->td_fairq_lticks != ticks) {
1183 	    ntd->td_fairq_lticks = ticks;
1184 	    ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1185 	    if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1186 		    ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1187 	}
1188     }
1189 }
1190 
1191 static __inline
1192 void
1193 _lwkt_schedule(thread_t td, int reschedok)
1194 {
1195     globaldata_t mygd = mycpu;
1196 
1197     KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1198     crit_enter_gd(mygd);
1199     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1200     if (td == mygd->gd_curthread) {
1201 	_lwkt_enqueue(td);
1202     } else {
1203 	/*
1204 	 * If we own the thread, there is no race (since we are in a
1205 	 * critical section).  If we do not own the thread there might
1206 	 * be a race but the target cpu will deal with it.
1207 	 */
1208 #ifdef SMP
1209 	if (td->td_gd == mygd) {
1210 	    _lwkt_enqueue(td);
1211 	    _lwkt_schedule_post(mygd, td, 1, reschedok);
1212 	} else {
1213 	    lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1214 	}
1215 #else
1216 	_lwkt_enqueue(td);
1217 	_lwkt_schedule_post(mygd, td, 1, reschedok);
1218 #endif
1219     }
1220     crit_exit_gd(mygd);
1221 }
1222 
1223 void
1224 lwkt_schedule(thread_t td)
1225 {
1226     _lwkt_schedule(td, 1);
1227 }
1228 
1229 void
1230 lwkt_schedule_noresched(thread_t td)
1231 {
1232     _lwkt_schedule(td, 0);
1233 }
1234 
1235 #ifdef SMP
1236 
1237 /*
1238  * When scheduled remotely if frame != NULL the IPIQ is being
1239  * run via doreti or an interrupt then preemption can be allowed.
1240  *
1241  * To allow preemption we have to drop the critical section so only
1242  * one is present in _lwkt_schedule_post.
1243  */
1244 static void
1245 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1246 {
1247     thread_t td = curthread;
1248     thread_t ntd = arg;
1249 
1250     if (frame && ntd->td_preemptable) {
1251 	crit_exit_noyield(td);
1252 	_lwkt_schedule(ntd, 1);
1253 	crit_enter_quick(td);
1254     } else {
1255 	_lwkt_schedule(ntd, 1);
1256     }
1257 }
1258 
1259 /*
1260  * Thread migration using a 'Pull' method.  The thread may or may not be
1261  * the current thread.  It MUST be descheduled and in a stable state.
1262  * lwkt_giveaway() must be called on the cpu owning the thread.
1263  *
1264  * At any point after lwkt_giveaway() is called, the target cpu may
1265  * 'pull' the thread by calling lwkt_acquire().
1266  *
1267  * We have to make sure the thread is not sitting on a per-cpu tsleep
1268  * queue or it will blow up when it moves to another cpu.
1269  *
1270  * MPSAFE - must be called under very specific conditions.
1271  */
1272 void
1273 lwkt_giveaway(thread_t td)
1274 {
1275     globaldata_t gd = mycpu;
1276 
1277     crit_enter_gd(gd);
1278     if (td->td_flags & TDF_TSLEEPQ)
1279 	tsleep_remove(td);
1280     KKASSERT(td->td_gd == gd);
1281     TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1282     td->td_flags |= TDF_MIGRATING;
1283     crit_exit_gd(gd);
1284 }
1285 
1286 void
1287 lwkt_acquire(thread_t td)
1288 {
1289     globaldata_t gd;
1290     globaldata_t mygd;
1291 
1292     KKASSERT(td->td_flags & TDF_MIGRATING);
1293     gd = td->td_gd;
1294     mygd = mycpu;
1295     if (gd != mycpu) {
1296 	cpu_lfence();
1297 	KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1298 	crit_enter_gd(mygd);
1299 	while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1300 #ifdef SMP
1301 	    lwkt_process_ipiq();
1302 #endif
1303 	    cpu_lfence();
1304 	}
1305 	td->td_gd = mygd;
1306 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1307 	td->td_flags &= ~TDF_MIGRATING;
1308 	crit_exit_gd(mygd);
1309     } else {
1310 	crit_enter_gd(mygd);
1311 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1312 	td->td_flags &= ~TDF_MIGRATING;
1313 	crit_exit_gd(mygd);
1314     }
1315 }
1316 
1317 #endif
1318 
1319 /*
1320  * Generic deschedule.  Descheduling threads other then your own should be
1321  * done only in carefully controlled circumstances.  Descheduling is
1322  * asynchronous.
1323  *
1324  * This function may block if the cpu has run out of messages.
1325  */
1326 void
1327 lwkt_deschedule(thread_t td)
1328 {
1329     crit_enter();
1330 #ifdef SMP
1331     if (td == curthread) {
1332 	_lwkt_dequeue(td);
1333     } else {
1334 	if (td->td_gd == mycpu) {
1335 	    _lwkt_dequeue(td);
1336 	} else {
1337 	    lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1338 	}
1339     }
1340 #else
1341     _lwkt_dequeue(td);
1342 #endif
1343     crit_exit();
1344 }
1345 
1346 /*
1347  * Set the target thread's priority.  This routine does not automatically
1348  * switch to a higher priority thread, LWKT threads are not designed for
1349  * continuous priority changes.  Yield if you want to switch.
1350  */
1351 void
1352 lwkt_setpri(thread_t td, int pri)
1353 {
1354     KKASSERT(td->td_gd == mycpu);
1355     if (td->td_pri != pri) {
1356 	KKASSERT(pri >= 0);
1357 	crit_enter();
1358 	if (td->td_flags & TDF_RUNQ) {
1359 	    _lwkt_dequeue(td);
1360 	    td->td_pri = pri;
1361 	    _lwkt_enqueue(td);
1362 	} else {
1363 	    td->td_pri = pri;
1364 	}
1365 	crit_exit();
1366     }
1367 }
1368 
1369 /*
1370  * Set the initial priority for a thread prior to it being scheduled for
1371  * the first time.  The thread MUST NOT be scheduled before or during
1372  * this call.  The thread may be assigned to a cpu other then the current
1373  * cpu.
1374  *
1375  * Typically used after a thread has been created with TDF_STOPPREQ,
1376  * and before the thread is initially scheduled.
1377  */
1378 void
1379 lwkt_setpri_initial(thread_t td, int pri)
1380 {
1381     KKASSERT(pri >= 0);
1382     KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1383     td->td_pri = pri;
1384 }
1385 
1386 void
1387 lwkt_setpri_self(int pri)
1388 {
1389     thread_t td = curthread;
1390 
1391     KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1392     crit_enter();
1393     if (td->td_flags & TDF_RUNQ) {
1394 	_lwkt_dequeue(td);
1395 	td->td_pri = pri;
1396 	_lwkt_enqueue(td);
1397     } else {
1398 	td->td_pri = pri;
1399     }
1400     crit_exit();
1401 }
1402 
1403 /*
1404  * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1405  *
1406  * Example: two competing threads, same priority N.  decrement by (2*N)
1407  * increment by N*8, each thread will get 4 ticks.
1408  */
1409 void
1410 lwkt_fairq_schedulerclock(thread_t td)
1411 {
1412     if (fairq_enable) {
1413 	while (td) {
1414 	    if (td != &td->td_gd->gd_idlethread) {
1415 		td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1416 		if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1417 			td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1418 		if (td->td_fairq_accum < 0)
1419 			need_lwkt_resched();
1420 		td->td_fairq_lticks = ticks;
1421 	    }
1422 	    td = td->td_preempted;
1423 	}
1424     }
1425 }
1426 
1427 static void
1428 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1429 {
1430 	td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1431 	if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1432 		td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1433 }
1434 
1435 /*
1436  * Migrate the current thread to the specified cpu.
1437  *
1438  * This is accomplished by descheduling ourselves from the current cpu,
1439  * moving our thread to the tdallq of the target cpu, IPI messaging the
1440  * target cpu, and switching out.  TDF_MIGRATING prevents scheduling
1441  * races while the thread is being migrated.
1442  *
1443  * We must be sure to remove ourselves from the current cpu's tsleepq
1444  * before potentially moving to another queue.  The thread can be on
1445  * a tsleepq due to a left-over tsleep_interlock().
1446  */
1447 #ifdef SMP
1448 static void lwkt_setcpu_remote(void *arg);
1449 #endif
1450 
1451 void
1452 lwkt_setcpu_self(globaldata_t rgd)
1453 {
1454 #ifdef SMP
1455     thread_t td = curthread;
1456 
1457     if (td->td_gd != rgd) {
1458 	crit_enter_quick(td);
1459 	if (td->td_flags & TDF_TSLEEPQ)
1460 	    tsleep_remove(td);
1461 	td->td_flags |= TDF_MIGRATING;
1462 	lwkt_deschedule_self(td);
1463 	TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1464 	lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1465 	lwkt_switch();
1466 	/* we are now on the target cpu */
1467 	TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1468 	crit_exit_quick(td);
1469     }
1470 #endif
1471 }
1472 
1473 void
1474 lwkt_migratecpu(int cpuid)
1475 {
1476 #ifdef SMP
1477 	globaldata_t rgd;
1478 
1479 	rgd = globaldata_find(cpuid);
1480 	lwkt_setcpu_self(rgd);
1481 #endif
1482 }
1483 
1484 /*
1485  * Remote IPI for cpu migration (called while in a critical section so we
1486  * do not have to enter another one).  The thread has already been moved to
1487  * our cpu's allq, but we must wait for the thread to be completely switched
1488  * out on the originating cpu before we schedule it on ours or the stack
1489  * state may be corrupt.  We clear TDF_MIGRATING after flushing the GD
1490  * change to main memory.
1491  *
1492  * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1493  * against wakeups.  It is best if this interface is used only when there
1494  * are no pending events that might try to schedule the thread.
1495  */
1496 #ifdef SMP
1497 static void
1498 lwkt_setcpu_remote(void *arg)
1499 {
1500     thread_t td = arg;
1501     globaldata_t gd = mycpu;
1502 
1503     while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1504 #ifdef SMP
1505 	lwkt_process_ipiq();
1506 #endif
1507 	cpu_lfence();
1508     }
1509     td->td_gd = gd;
1510     cpu_sfence();
1511     td->td_flags &= ~TDF_MIGRATING;
1512     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1513     _lwkt_enqueue(td);
1514 }
1515 #endif
1516 
1517 struct lwp *
1518 lwkt_preempted_proc(void)
1519 {
1520     thread_t td = curthread;
1521     while (td->td_preempted)
1522 	td = td->td_preempted;
1523     return(td->td_lwp);
1524 }
1525 
1526 /*
1527  * Create a kernel process/thread/whatever.  It shares it's address space
1528  * with proc0 - ie: kernel only.
1529  *
1530  * NOTE!  By default new threads are created with the MP lock held.  A
1531  * thread which does not require the MP lock should release it by calling
1532  * rel_mplock() at the start of the new thread.
1533  */
1534 int
1535 lwkt_create(void (*func)(void *), void *arg,
1536     struct thread **tdp, thread_t template, int tdflags, int cpu,
1537     const char *fmt, ...)
1538 {
1539     thread_t td;
1540     __va_list ap;
1541 
1542     td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1543 			   tdflags);
1544     if (tdp)
1545 	*tdp = td;
1546     cpu_set_thread_handler(td, lwkt_exit, func, arg);
1547 
1548     /*
1549      * Set up arg0 for 'ps' etc
1550      */
1551     __va_start(ap, fmt);
1552     kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1553     __va_end(ap);
1554 
1555     /*
1556      * Schedule the thread to run
1557      */
1558     if ((td->td_flags & TDF_STOPREQ) == 0)
1559 	lwkt_schedule(td);
1560     else
1561 	td->td_flags &= ~TDF_STOPREQ;
1562     return 0;
1563 }
1564 
1565 /*
1566  * Destroy an LWKT thread.   Warning!  This function is not called when
1567  * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1568  * uses a different reaping mechanism.
1569  */
1570 void
1571 lwkt_exit(void)
1572 {
1573     thread_t td = curthread;
1574     thread_t std;
1575     globaldata_t gd;
1576 
1577     if (td->td_flags & TDF_VERBOSE)
1578 	kprintf("kthread %p %s has exited\n", td, td->td_comm);
1579     caps_exit(td);
1580 
1581     /*
1582      * Get us into a critical section to interlock gd_freetd and loop
1583      * until we can get it freed.
1584      *
1585      * We have to cache the current td in gd_freetd because objcache_put()ing
1586      * it would rip it out from under us while our thread is still active.
1587      */
1588     gd = mycpu;
1589     crit_enter_quick(td);
1590     while ((std = gd->gd_freetd) != NULL) {
1591 	gd->gd_freetd = NULL;
1592 	objcache_put(thread_cache, std);
1593     }
1594 
1595     /*
1596      * Remove thread resources from kernel lists and deschedule us for
1597      * the last time.
1598      */
1599     if (td->td_flags & TDF_TSLEEPQ)
1600 	tsleep_remove(td);
1601     biosched_done(td);
1602     dsched_exit_thread(td);
1603     lwkt_deschedule_self(td);
1604     lwkt_remove_tdallq(td);
1605     if (td->td_flags & TDF_ALLOCATED_THREAD)
1606 	gd->gd_freetd = td;
1607     cpu_thread_exit();
1608 }
1609 
1610 void
1611 lwkt_remove_tdallq(thread_t td)
1612 {
1613     KKASSERT(td->td_gd == mycpu);
1614     TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1615 }
1616 
1617 void
1618 crit_panic(void)
1619 {
1620     thread_t td = curthread;
1621     int lcrit = td->td_critcount;
1622 
1623     td->td_critcount = 0;
1624     panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1625 }
1626 
1627 #ifdef SMP
1628 
1629 /*
1630  * Called from debugger/panic on cpus which have been stopped.  We must still
1631  * process the IPIQ while stopped, even if we were stopped while in a critical
1632  * section (XXX).
1633  *
1634  * If we are dumping also try to process any pending interrupts.  This may
1635  * or may not work depending on the state of the cpu at the point it was
1636  * stopped.
1637  */
1638 void
1639 lwkt_smp_stopped(void)
1640 {
1641     globaldata_t gd = mycpu;
1642 
1643     crit_enter_gd(gd);
1644     if (dumping) {
1645 	lwkt_process_ipiq();
1646 	splz();
1647     } else {
1648 	lwkt_process_ipiq();
1649     }
1650     crit_exit_gd(gd);
1651 }
1652 
1653 #endif
1654