xref: /dflybsd-src/sys/kern/lwkt_thread.c (revision dae741e33c840b92a8a53bf9f01157ede145e256)
1 /*
2  * Copyright (c) 2003-2011 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[TDPRI_MAX+1] __debugvar;
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 static void lwkt_setcpu_remote(void *arg);
103 #endif
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 /*
111  * We can make all thread ports use the spin backend instead of the thread
112  * backend.  This should only be set to debug the spin backend.
113  */
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115 
116 #ifdef	INVARIANTS
117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118     "Panic if attempting to switch lwkt's while mastering cpusync");
119 #endif
120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121     "Number of switched threads");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
123     "Successful preemption events");
124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
125     "Failed preemption events");
126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127     "Number of preempted threads.");
128 #ifdef	INVARIANTS
129 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW,
130 	&token_contention_count[0], 0, "spinning due to token contention");
131 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW,
132 	&token_contention_count[1], 0, "spinning due to token contention");
133 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW,
134 	&token_contention_count[2], 0, "spinning due to token contention");
135 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW,
136 	&token_contention_count[3], 0, "spinning due to token contention");
137 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW,
138 	&token_contention_count[4], 0, "spinning due to token contention");
139 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW,
140 	&token_contention_count[5], 0, "spinning due to token contention");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW,
142 	&token_contention_count[6], 0, "spinning due to token contention");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW,
144 	&token_contention_count[7], 0, "spinning due to token contention");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW,
146 	&token_contention_count[8], 0, "spinning due to token contention");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW,
148 	&token_contention_count[9], 0, "spinning due to token contention");
149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW,
150 	&token_contention_count[10], 0, "spinning due to token contention");
151 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW,
152 	&token_contention_count[11], 0, "spinning due to token contention");
153 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW,
154 	&token_contention_count[12], 0, "spinning due to token contention");
155 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW,
156 	&token_contention_count[13], 0, "spinning due to token contention");
157 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW,
158 	&token_contention_count[14], 0, "spinning due to token contention");
159 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW,
160 	&token_contention_count[15], 0, "spinning due to token contention");
161 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW,
162 	&token_contention_count[16], 0, "spinning due to token contention");
163 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW,
164 	&token_contention_count[17], 0, "spinning due to token contention");
165 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW,
166 	&token_contention_count[18], 0, "spinning due to token contention");
167 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW,
168 	&token_contention_count[19], 0, "spinning due to token contention");
169 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW,
170 	&token_contention_count[20], 0, "spinning due to token contention");
171 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW,
172 	&token_contention_count[21], 0, "spinning due to token contention");
173 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW,
174 	&token_contention_count[22], 0, "spinning due to token contention");
175 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW,
176 	&token_contention_count[23], 0, "spinning due to token contention");
177 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW,
178 	&token_contention_count[24], 0, "spinning due to token contention");
179 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW,
180 	&token_contention_count[25], 0, "spinning due to token contention");
181 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW,
182 	&token_contention_count[26], 0, "spinning due to token contention");
183 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW,
184 	&token_contention_count[27], 0, "spinning due to token contention");
185 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW,
186 	&token_contention_count[28], 0, "spinning due to token contention");
187 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW,
188 	&token_contention_count[29], 0, "spinning due to token contention");
189 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW,
190 	&token_contention_count[30], 0, "spinning due to token contention");
191 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW,
192 	&token_contention_count[31], 0, "spinning due to token contention");
193 #endif
194 static int fairq_enable = 0;
195 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
196 	&fairq_enable, 0, "Turn on fairq priority accumulators");
197 static int fairq_bypass = -1;
198 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
199 	&fairq_bypass, 0, "Allow fairq to bypass td on token failure");
200 extern int lwkt_sched_debug;
201 int lwkt_sched_debug = 0;
202 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
203 	&lwkt_sched_debug, 0, "Scheduler debug");
204 static int lwkt_spin_loops = 10;
205 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
206 	&lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
207 static int lwkt_spin_reseq = 0;
208 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
209 	&lwkt_spin_reseq, 0, "Scheduler resequencer enable");
210 static int lwkt_spin_monitor = 0;
211 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
212 	&lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
213 static int lwkt_spin_fatal = 0;	/* disabled */
214 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
215 	&lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
216 static int preempt_enable = 1;
217 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
218 	&preempt_enable, 0, "Enable preemption");
219 static int lwkt_cache_threads = 32;
220 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
221 	&lwkt_cache_threads, 0, "thread+kstack cache");
222 
223 static __cachealign int lwkt_cseq_rindex;
224 static __cachealign int lwkt_cseq_windex;
225 
226 /*
227  * These helper procedures handle the runq, they can only be called from
228  * within a critical section.
229  *
230  * WARNING!  Prior to SMP being brought up it is possible to enqueue and
231  * dequeue threads belonging to other cpus, so be sure to use td->td_gd
232  * instead of 'mycpu' when referencing the globaldata structure.   Once
233  * SMP live enqueuing and dequeueing only occurs on the current cpu.
234  */
235 static __inline
236 void
237 _lwkt_dequeue(thread_t td)
238 {
239     if (td->td_flags & TDF_RUNQ) {
240 	struct globaldata *gd = td->td_gd;
241 
242 	td->td_flags &= ~TDF_RUNQ;
243 	TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
244 	if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
245 		atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
246     }
247 }
248 
249 /*
250  * Priority enqueue.
251  *
252  * NOTE: There are a limited number of lwkt threads runnable since user
253  *	 processes only schedule one at a time per cpu.
254  */
255 static __inline
256 void
257 _lwkt_enqueue(thread_t td)
258 {
259     thread_t xtd;
260 
261     if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
262 	struct globaldata *gd = td->td_gd;
263 
264 	td->td_flags |= TDF_RUNQ;
265 	xtd = TAILQ_FIRST(&gd->gd_tdrunq);
266 	if (xtd == NULL) {
267 	    TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
268 	    atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
269 	} else {
270 	    while (xtd && xtd->td_pri >= td->td_pri)
271 		xtd = TAILQ_NEXT(xtd, td_threadq);
272 	    if (xtd)
273 		TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
274 	    else
275 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
276 	}
277 
278 	/*
279 	 * Request a LWKT reschedule if we are now at the head of the queue.
280 	 */
281 	if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
282 	    need_lwkt_resched();
283     }
284 }
285 
286 static __boolean_t
287 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
288 {
289 	struct thread *td = (struct thread *)obj;
290 
291 	td->td_kstack = NULL;
292 	td->td_kstack_size = 0;
293 	td->td_flags = TDF_ALLOCATED_THREAD;
294 	return (1);
295 }
296 
297 static void
298 _lwkt_thread_dtor(void *obj, void *privdata)
299 {
300 	struct thread *td = (struct thread *)obj;
301 
302 	KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
303 	    ("_lwkt_thread_dtor: not allocated from objcache"));
304 	KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
305 		td->td_kstack_size > 0,
306 	    ("_lwkt_thread_dtor: corrupted stack"));
307 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
308 }
309 
310 /*
311  * Initialize the lwkt s/system.
312  *
313  * Nominally cache up to 32 thread + kstack structures.
314  */
315 void
316 lwkt_init(void)
317 {
318     TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
319     thread_cache = objcache_create_mbacked(
320 				M_THREAD, sizeof(struct thread),
321 				NULL, lwkt_cache_threads,
322 				_lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
323 }
324 
325 /*
326  * Schedule a thread to run.  As the current thread we can always safely
327  * schedule ourselves, and a shortcut procedure is provided for that
328  * function.
329  *
330  * (non-blocking, self contained on a per cpu basis)
331  */
332 void
333 lwkt_schedule_self(thread_t td)
334 {
335     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
336     crit_enter_quick(td);
337     KASSERT(td != &td->td_gd->gd_idlethread,
338 	    ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
339     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
340     _lwkt_enqueue(td);
341     crit_exit_quick(td);
342 }
343 
344 /*
345  * Deschedule a thread.
346  *
347  * (non-blocking, self contained on a per cpu basis)
348  */
349 void
350 lwkt_deschedule_self(thread_t td)
351 {
352     crit_enter_quick(td);
353     _lwkt_dequeue(td);
354     crit_exit_quick(td);
355 }
356 
357 /*
358  * LWKTs operate on a per-cpu basis
359  *
360  * WARNING!  Called from early boot, 'mycpu' may not work yet.
361  */
362 void
363 lwkt_gdinit(struct globaldata *gd)
364 {
365     TAILQ_INIT(&gd->gd_tdrunq);
366     TAILQ_INIT(&gd->gd_tdallq);
367 }
368 
369 /*
370  * Create a new thread.  The thread must be associated with a process context
371  * or LWKT start address before it can be scheduled.  If the target cpu is
372  * -1 the thread will be created on the current cpu.
373  *
374  * If you intend to create a thread without a process context this function
375  * does everything except load the startup and switcher function.
376  */
377 thread_t
378 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
379 {
380     static int cpu_rotator;
381     globaldata_t gd = mycpu;
382     void *stack;
383 
384     /*
385      * If static thread storage is not supplied allocate a thread.  Reuse
386      * a cached free thread if possible.  gd_freetd is used to keep an exiting
387      * thread intact through the exit.
388      */
389     if (td == NULL) {
390 	crit_enter_gd(gd);
391 	if ((td = gd->gd_freetd) != NULL) {
392 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
393 				      TDF_RUNQ)) == 0);
394 	    gd->gd_freetd = NULL;
395 	} else {
396 	    td = objcache_get(thread_cache, M_WAITOK);
397 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
398 				      TDF_RUNQ)) == 0);
399 	}
400 	crit_exit_gd(gd);
401     	KASSERT((td->td_flags &
402 		 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
403 		("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
404     	flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
405     }
406 
407     /*
408      * Try to reuse cached stack.
409      */
410     if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
411 	if (flags & TDF_ALLOCATED_STACK) {
412 	    kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
413 	    stack = NULL;
414 	}
415     }
416     if (stack == NULL) {
417 	stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
418 	flags |= TDF_ALLOCATED_STACK;
419     }
420     if (cpu < 0) {
421 	cpu = ++cpu_rotator;
422 	cpu_ccfence();
423 	cpu %= ncpus;
424     }
425     lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
426     return(td);
427 }
428 
429 /*
430  * Initialize a preexisting thread structure.  This function is used by
431  * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
432  *
433  * All threads start out in a critical section at a priority of
434  * TDPRI_KERN_DAEMON.  Higher level code will modify the priority as
435  * appropriate.  This function may send an IPI message when the
436  * requested cpu is not the current cpu and consequently gd_tdallq may
437  * not be initialized synchronously from the point of view of the originating
438  * cpu.
439  *
440  * NOTE! we have to be careful in regards to creating threads for other cpus
441  * if SMP has not yet been activated.
442  */
443 #ifdef SMP
444 
445 static void
446 lwkt_init_thread_remote(void *arg)
447 {
448     thread_t td = arg;
449 
450     /*
451      * Protected by critical section held by IPI dispatch
452      */
453     TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
454 }
455 
456 #endif
457 
458 /*
459  * lwkt core thread structural initialization.
460  *
461  * NOTE: All threads are initialized as mpsafe threads.
462  */
463 void
464 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
465 		struct globaldata *gd)
466 {
467     globaldata_t mygd = mycpu;
468 
469     bzero(td, sizeof(struct thread));
470     td->td_kstack = stack;
471     td->td_kstack_size = stksize;
472     td->td_flags = flags;
473     td->td_gd = gd;
474     td->td_pri = TDPRI_KERN_DAEMON;
475     td->td_critcount = 1;
476     td->td_toks_stop = &td->td_toks_base;
477     if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
478 	lwkt_initport_spin(&td->td_msgport);
479     else
480 	lwkt_initport_thread(&td->td_msgport, td);
481     pmap_init_thread(td);
482 #ifdef SMP
483     /*
484      * Normally initializing a thread for a remote cpu requires sending an
485      * IPI.  However, the idlethread is setup before the other cpus are
486      * activated so we have to treat it as a special case.  XXX manipulation
487      * of gd_tdallq requires the BGL.
488      */
489     if (gd == mygd || td == &gd->gd_idlethread) {
490 	crit_enter_gd(mygd);
491 	TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
492 	crit_exit_gd(mygd);
493     } else {
494 	lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
495     }
496 #else
497     crit_enter_gd(mygd);
498     TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
499     crit_exit_gd(mygd);
500 #endif
501 
502     dsched_new_thread(td);
503 }
504 
505 void
506 lwkt_set_comm(thread_t td, const char *ctl, ...)
507 {
508     __va_list va;
509 
510     __va_start(va, ctl);
511     kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
512     __va_end(va);
513     KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
514 }
515 
516 void
517 lwkt_hold(thread_t td)
518 {
519     atomic_add_int(&td->td_refs, 1);
520 }
521 
522 void
523 lwkt_rele(thread_t td)
524 {
525     KKASSERT(td->td_refs > 0);
526     atomic_add_int(&td->td_refs, -1);
527 }
528 
529 void
530 lwkt_wait_free(thread_t td)
531 {
532     while (td->td_refs)
533 	tsleep(td, 0, "tdreap", hz);
534 }
535 
536 void
537 lwkt_free_thread(thread_t td)
538 {
539     KKASSERT(td->td_refs == 0);
540     KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
541     if (td->td_flags & TDF_ALLOCATED_THREAD) {
542     	objcache_put(thread_cache, td);
543     } else if (td->td_flags & TDF_ALLOCATED_STACK) {
544 	/* client-allocated struct with internally allocated stack */
545 	KASSERT(td->td_kstack && td->td_kstack_size > 0,
546 	    ("lwkt_free_thread: corrupted stack"));
547 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
548 	td->td_kstack = NULL;
549 	td->td_kstack_size = 0;
550     }
551     KTR_LOG(ctxsw_deadtd, td);
552 }
553 
554 
555 /*
556  * Switch to the next runnable lwkt.  If no LWKTs are runnable then
557  * switch to the idlethread.  Switching must occur within a critical
558  * section to avoid races with the scheduling queue.
559  *
560  * We always have full control over our cpu's run queue.  Other cpus
561  * that wish to manipulate our queue must use the cpu_*msg() calls to
562  * talk to our cpu, so a critical section is all that is needed and
563  * the result is very, very fast thread switching.
564  *
565  * The LWKT scheduler uses a fixed priority model and round-robins at
566  * each priority level.  User process scheduling is a totally
567  * different beast and LWKT priorities should not be confused with
568  * user process priorities.
569  *
570  * PREEMPTION NOTE: Preemption occurs via lwkt_preempt().  lwkt_switch()
571  * is not called by the current thread in the preemption case, only when
572  * the preempting thread blocks (in order to return to the original thread).
573  *
574  * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
575  * migration and tsleep deschedule the current lwkt thread and call
576  * lwkt_switch().  In particular, the target cpu of the migration fully
577  * expects the thread to become non-runnable and can deadlock against
578  * cpusync operations if we run any IPIs prior to switching the thread out.
579  *
580  * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
581  * THE CURRENT THREAD HAS BEEN DESCHEDULED!
582  */
583 void
584 lwkt_switch(void)
585 {
586     globaldata_t gd = mycpu;
587     thread_t td = gd->gd_curthread;
588     thread_t ntd;
589     thread_t xtd;
590     int spinning = 0;
591 
592     KKASSERT(gd->gd_processing_ipiq == 0);
593 
594     /*
595      * Switching from within a 'fast' (non thread switched) interrupt or IPI
596      * is illegal.  However, we may have to do it anyway if we hit a fatal
597      * kernel trap or we have paniced.
598      *
599      * If this case occurs save and restore the interrupt nesting level.
600      */
601     if (gd->gd_intr_nesting_level) {
602 	int savegdnest;
603 	int savegdtrap;
604 
605 	if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
606 	    panic("lwkt_switch: Attempt to switch from a "
607 		  "a fast interrupt, ipi, or hard code section, "
608 		  "td %p\n",
609 		  td);
610 	} else {
611 	    savegdnest = gd->gd_intr_nesting_level;
612 	    savegdtrap = gd->gd_trap_nesting_level;
613 	    gd->gd_intr_nesting_level = 0;
614 	    gd->gd_trap_nesting_level = 0;
615 	    if ((td->td_flags & TDF_PANICWARN) == 0) {
616 		td->td_flags |= TDF_PANICWARN;
617 		kprintf("Warning: thread switch from interrupt, IPI, "
618 			"or hard code section.\n"
619 			"thread %p (%s)\n", td, td->td_comm);
620 		print_backtrace(-1);
621 	    }
622 	    lwkt_switch();
623 	    gd->gd_intr_nesting_level = savegdnest;
624 	    gd->gd_trap_nesting_level = savegdtrap;
625 	    return;
626 	}
627     }
628 
629     /*
630      * Release our current user process designation if we are blocking
631      * or if a user reschedule was requested.
632      *
633      * NOTE: This function is NOT called if we are switching into or
634      *	     returning from a preemption.
635      *
636      * NOTE: Releasing our current user process designation may cause
637      *	     it to be assigned to another thread, which in turn will
638      *	     cause us to block in the usched acquire code when we attempt
639      *	     to return to userland.
640      *
641      * NOTE: On SMP systems this can be very nasty when heavy token
642      *	     contention is present so we want to be careful not to
643      *	     release the designation gratuitously.
644      */
645     if (td->td_release &&
646 	(user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
647 	    td->td_release(td);
648     }
649 
650     /*
651      * Release all tokens
652      */
653     crit_enter_gd(gd);
654     if (TD_TOKS_HELD(td))
655 	    lwkt_relalltokens(td);
656 
657     /*
658      * We had better not be holding any spin locks, but don't get into an
659      * endless panic loop.
660      */
661     KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
662 	    ("lwkt_switch: still holding %d exclusive spinlocks!",
663 	     gd->gd_spinlocks_wr));
664 
665 
666 #ifdef SMP
667 #ifdef	INVARIANTS
668     if (td->td_cscount) {
669 	kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
670 		td);
671 	if (panic_on_cscount)
672 	    panic("switching while mastering cpusync");
673     }
674 #endif
675 #endif
676 
677     /*
678      * If we had preempted another thread on this cpu, resume the preempted
679      * thread.  This occurs transparently, whether the preempted thread
680      * was scheduled or not (it may have been preempted after descheduling
681      * itself).
682      *
683      * We have to setup the MP lock for the original thread after backing
684      * out the adjustment that was made to curthread when the original
685      * was preempted.
686      */
687     if ((ntd = td->td_preempted) != NULL) {
688 	KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
689 	ntd->td_flags |= TDF_PREEMPT_DONE;
690 
691 	/*
692 	 * The interrupt may have woken a thread up, we need to properly
693 	 * set the reschedule flag if the originally interrupted thread is
694 	 * at a lower priority.
695 	 *
696 	 * The interrupt may not have descheduled.
697 	 */
698 	if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
699 	    need_lwkt_resched();
700 	goto havethread_preempted;
701     }
702 
703     /*
704      * If we cannot obtain ownership of the tokens we cannot immediately
705      * schedule the target thread.
706      *
707      * Reminder: Again, we cannot afford to run any IPIs in this path if
708      * the current thread has been descheduled.
709      */
710     for (;;) {
711 	clear_lwkt_resched();
712 
713 	/*
714 	 * Hotpath - pull the head of the run queue and attempt to schedule
715 	 * it.
716 	 */
717 	for (;;) {
718 	    ntd = TAILQ_FIRST(&gd->gd_tdrunq);
719 
720 	    if (ntd == NULL) {
721 		/*
722 		 * Runq is empty, switch to idle to allow it to halt.
723 		 */
724 		ntd = &gd->gd_idlethread;
725 #ifdef SMP
726 		if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
727 		    ASSERT_NO_TOKENS_HELD(ntd);
728 #endif
729 		cpu_time.cp_msg[0] = 0;
730 		cpu_time.cp_stallpc = 0;
731 		goto haveidle;
732 	    }
733 	    break;
734 	}
735 
736 	/*
737 	 * Hotpath - schedule ntd.
738 	 *
739 	 * NOTE: For UP there is no mplock and lwkt_getalltokens()
740 	 *	     always succeeds.
741 	 */
742 	if (TD_TOKS_NOT_HELD(ntd) ||
743 	    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
744 	{
745 	    goto havethread;
746 	}
747 
748 	/*
749 	 * Coldpath (SMP only since tokens always succeed on UP)
750 	 *
751 	 * We had some contention on the thread we wanted to schedule.
752 	 * What we do now is try to find a thread that we can schedule
753 	 * in its stead.
754 	 *
755 	 * The coldpath scan does NOT rearrange threads in the run list.
756 	 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
757 	 * the next tick whenever the current head is not the current thread.
758 	 */
759 #ifdef	INVARIANTS
760 	++token_contention_count[ntd->td_pri];
761 	++ntd->td_contended;
762 #endif
763 
764 	if (fairq_bypass > 0)
765 		goto skip;
766 
767 	xtd = NULL;
768 	while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
769 		/*
770 		 * Never schedule threads returning to userland or the
771 		 * user thread scheduler helper thread when higher priority
772 		 * threads are present.
773 		 */
774 		if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
775 			ntd = NULL;
776 			break;
777 		}
778 
779 		/*
780 		 * Try this one.
781 		 */
782 		if (TD_TOKS_NOT_HELD(ntd) ||
783 		    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
784 			goto havethread;
785 		}
786 #ifdef	INVARIANTS
787 		++token_contention_count[ntd->td_pri];
788 		++ntd->td_contended;
789 #endif
790 	}
791 
792 skip:
793 	/*
794 	 * We exhausted the run list, meaning that all runnable threads
795 	 * are contested.
796 	 */
797 	cpu_pause();
798 	ntd = &gd->gd_idlethread;
799 #ifdef SMP
800 	if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
801 	    ASSERT_NO_TOKENS_HELD(ntd);
802 	/* contention case, do not clear contention mask */
803 #endif
804 
805 	/*
806 	 * We are going to have to retry but if the current thread is not
807 	 * on the runq we instead switch through the idle thread to get away
808 	 * from the current thread.  We have to flag for lwkt reschedule
809 	 * to prevent the idle thread from halting.
810 	 *
811 	 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
812 	 *	 instruct it to deal with the potential for deadlocks by
813 	 *	 ordering the tokens by address.
814 	 */
815 	if ((td->td_flags & TDF_RUNQ) == 0) {
816 	    need_lwkt_resched();	/* prevent hlt */
817 	    goto haveidle;
818 	}
819 #if defined(INVARIANTS) && defined(__amd64__)
820 	if ((read_rflags() & PSL_I) == 0) {
821 		cpu_enable_intr();
822 		panic("lwkt_switch() called with interrupts disabled");
823 	}
824 #endif
825 
826 	/*
827 	 * Number iterations so far.  After a certain point we switch to
828 	 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
829 	 */
830 	if (spinning < 0x7FFFFFFF)
831 	    ++spinning;
832 
833 #ifdef SMP
834 	/*
835 	 * lwkt_getalltokens() failed in sorted token mode, we can use
836 	 * monitor/mwait in this case.
837 	 */
838 	if (spinning >= lwkt_spin_loops &&
839 	    (cpu_mi_feature & CPU_MI_MONITOR) &&
840 	    lwkt_spin_monitor)
841 	{
842 	    cpu_mmw_pause_int(&gd->gd_reqflags,
843 			      (gd->gd_reqflags | RQF_SPINNING) &
844 			      ~RQF_IDLECHECK_WK_MASK);
845 	}
846 #endif
847 
848 	/*
849 	 * We already checked that td is still scheduled so this should be
850 	 * safe.
851 	 */
852 	splz_check();
853 
854 	/*
855 	 * This experimental resequencer is used as a fall-back to reduce
856 	 * hw cache line contention by placing each core's scheduler into a
857 	 * time-domain-multplexed slot.
858 	 *
859 	 * The resequencer is disabled by default.  It's functionality has
860 	 * largely been superceeded by the token algorithm which limits races
861 	 * to a subset of cores.
862 	 *
863 	 * The resequencer algorithm tends to break down when more than
864 	 * 20 cores are contending.  What appears to happen is that new
865 	 * tokens can be obtained out of address-sorted order by new cores
866 	 * while existing cores languish in long delays between retries and
867 	 * wind up being starved-out of the token acquisition.
868 	 */
869 	if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
870 	    int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
871 	    int oseq;
872 
873 	    while ((oseq = lwkt_cseq_rindex) != cseq) {
874 		cpu_ccfence();
875 #if 1
876 		if (cpu_mi_feature & CPU_MI_MONITOR) {
877 		    cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
878 		} else {
879 #endif
880 		    cpu_pause();
881 		    cpu_lfence();
882 #if 1
883 		}
884 #endif
885 	    }
886 	    DELAY(1);
887 	    atomic_add_int(&lwkt_cseq_rindex, 1);
888 	}
889 	/* highest level for(;;) loop */
890     }
891 
892 havethread:
893     /*
894      * If the thread we came up with is a higher or equal priority verses
895      * the thread at the head of the queue we move our thread to the
896      * front.  This way we can always check the front of the queue.
897      *
898      * Clear gd_idle_repeat when doing a normal switch to a non-idle
899      * thread.
900      */
901     ntd->td_wmesg = NULL;
902     ++gd->gd_cnt.v_swtch;
903 #if 0
904     xtd = TAILQ_FIRST(&gd->gd_tdrunq);
905     if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
906 	TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
907 	TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
908     }
909 #endif
910     gd->gd_idle_repeat = 0;
911 
912 havethread_preempted:
913     /*
914      * If the new target does not need the MP lock and we are holding it,
915      * release the MP lock.  If the new target requires the MP lock we have
916      * already acquired it for the target.
917      */
918     ;
919 haveidle:
920     KASSERT(ntd->td_critcount,
921 	    ("priority problem in lwkt_switch %d %d",
922 	    td->td_critcount, ntd->td_critcount));
923 
924     if (td != ntd) {
925 	/*
926 	 * Execute the actual thread switch operation.  This function
927 	 * returns to the current thread and returns the previous thread
928 	 * (which may be different from the thread we switched to).
929 	 *
930 	 * We are responsible for marking ntd as TDF_RUNNING.
931 	 */
932 	++switch_count;
933 	KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
934 	ntd->td_flags |= TDF_RUNNING;
935 	lwkt_switch_return(td->td_switch(ntd));
936 	/* ntd invalid, td_switch() can return a different thread_t */
937     }
938 
939 #if 1
940     /*
941      * catch-all
942      */
943     splz_check();
944 #endif
945     /* NOTE: current cpu may have changed after switch */
946     crit_exit_quick(td);
947 }
948 
949 /*
950  * Called by assembly in the td_switch (thread restore path) for thread
951  * bootstrap cases which do not 'return' to lwkt_switch().
952  */
953 void
954 lwkt_switch_return(thread_t otd)
955 {
956 #ifdef SMP
957 	globaldata_t rgd;
958 
959 	/*
960 	 * Check if otd was migrating.  Now that we are on ntd we can finish
961 	 * up the migration.  This is a bit messy but it is the only place
962 	 * where td is known to be fully descheduled.
963 	 *
964 	 * We can only activate the migration if otd was migrating but not
965 	 * held on the cpu due to a preemption chain.  We still have to
966 	 * clear TDF_RUNNING on the old thread either way.
967 	 *
968 	 * We are responsible for clearing the previously running thread's
969 	 * TDF_RUNNING.
970 	 */
971 	if ((rgd = otd->td_migrate_gd) != NULL &&
972 	    (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
973 		KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
974 			 (TDF_MIGRATING | TDF_RUNNING));
975 		otd->td_migrate_gd = NULL;
976 		otd->td_flags &= ~TDF_RUNNING;
977 		lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
978 	} else {
979 		otd->td_flags &= ~TDF_RUNNING;
980 	}
981 #else
982 	otd->td_flags &= ~TDF_RUNNING;
983 #endif
984 }
985 
986 /*
987  * Request that the target thread preempt the current thread.  Preemption
988  * only works under a specific set of conditions:
989  *
990  *	- We are not preempting ourselves
991  *	- The target thread is owned by the current cpu
992  *	- We are not currently being preempted
993  *	- The target is not currently being preempted
994  *	- We are not holding any spin locks
995  *	- The target thread is not holding any tokens
996  *	- We are able to satisfy the target's MP lock requirements (if any).
997  *
998  * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION.  Typically
999  * this is called via lwkt_schedule() through the td_preemptable callback.
1000  * critcount is the managed critical priority that we should ignore in order
1001  * to determine whether preemption is possible (aka usually just the crit
1002  * priority of lwkt_schedule() itself).
1003  *
1004  * XXX at the moment we run the target thread in a critical section during
1005  * the preemption in order to prevent the target from taking interrupts
1006  * that *WE* can't.  Preemption is strictly limited to interrupt threads
1007  * and interrupt-like threads, outside of a critical section, and the
1008  * preempted source thread will be resumed the instant the target blocks
1009  * whether or not the source is scheduled (i.e. preemption is supposed to
1010  * be as transparent as possible).
1011  */
1012 void
1013 lwkt_preempt(thread_t ntd, int critcount)
1014 {
1015     struct globaldata *gd = mycpu;
1016     thread_t xtd;
1017     thread_t td;
1018     int save_gd_intr_nesting_level;
1019 
1020     /*
1021      * The caller has put us in a critical section.  We can only preempt
1022      * if the caller of the caller was not in a critical section (basically
1023      * a local interrupt), as determined by the 'critcount' parameter.  We
1024      * also can't preempt if the caller is holding any spinlocks (even if
1025      * he isn't in a critical section).  This also handles the tokens test.
1026      *
1027      * YYY The target thread must be in a critical section (else it must
1028      * inherit our critical section?  I dunno yet).
1029      */
1030     KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1031 
1032     td = gd->gd_curthread;
1033     if (preempt_enable == 0) {
1034 #if 0
1035 	if (ntd->td_pri > td->td_pri)
1036 	    need_lwkt_resched();
1037 #endif
1038 	++preempt_miss;
1039 	return;
1040     }
1041     if (ntd->td_pri <= td->td_pri) {
1042 	++preempt_miss;
1043 	return;
1044     }
1045     if (td->td_critcount > critcount) {
1046 	++preempt_miss;
1047 #if 0
1048 	need_lwkt_resched();
1049 #endif
1050 	return;
1051     }
1052 #ifdef SMP
1053     if (ntd->td_gd != gd) {
1054 	++preempt_miss;
1055 #if 0
1056 	need_lwkt_resched();
1057 #endif
1058 	return;
1059     }
1060 #endif
1061     /*
1062      * We don't have to check spinlocks here as they will also bump
1063      * td_critcount.
1064      *
1065      * Do not try to preempt if the target thread is holding any tokens.
1066      * We could try to acquire the tokens but this case is so rare there
1067      * is no need to support it.
1068      */
1069     KKASSERT(gd->gd_spinlocks_wr == 0);
1070 
1071     if (TD_TOKS_HELD(ntd)) {
1072 	++preempt_miss;
1073 #if 0
1074 	need_lwkt_resched();
1075 #endif
1076 	return;
1077     }
1078     if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1079 	++preempt_weird;
1080 #if 0
1081 	need_lwkt_resched();
1082 #endif
1083 	return;
1084     }
1085     if (ntd->td_preempted) {
1086 	++preempt_hit;
1087 #if 0
1088 	need_lwkt_resched();
1089 #endif
1090 	return;
1091     }
1092     KKASSERT(gd->gd_processing_ipiq == 0);
1093 
1094     /*
1095      * Since we are able to preempt the current thread, there is no need to
1096      * call need_lwkt_resched().
1097      *
1098      * We must temporarily clear gd_intr_nesting_level around the switch
1099      * since switchouts from the target thread are allowed (they will just
1100      * return to our thread), and since the target thread has its own stack.
1101      *
1102      * A preemption must switch back to the original thread, assert the
1103      * case.
1104      */
1105     ++preempt_hit;
1106     ntd->td_preempted = td;
1107     td->td_flags |= TDF_PREEMPT_LOCK;
1108     KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1109     save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1110     gd->gd_intr_nesting_level = 0;
1111     ntd->td_flags |= TDF_RUNNING;
1112     xtd = td->td_switch(ntd);
1113     KKASSERT(xtd == ntd);
1114     lwkt_switch_return(xtd);
1115     gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1116 
1117     KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1118     ntd->td_preempted = NULL;
1119     td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1120 }
1121 
1122 /*
1123  * Conditionally call splz() if gd_reqflags indicates work is pending.
1124  * This will work inside a critical section but not inside a hard code
1125  * section.
1126  *
1127  * (self contained on a per cpu basis)
1128  */
1129 void
1130 splz_check(void)
1131 {
1132     globaldata_t gd = mycpu;
1133     thread_t td = gd->gd_curthread;
1134 
1135     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1136 	gd->gd_intr_nesting_level == 0 &&
1137 	td->td_nest_count < 2)
1138     {
1139 	splz();
1140     }
1141 }
1142 
1143 /*
1144  * This version is integrated into crit_exit, reqflags has already
1145  * been tested but td_critcount has not.
1146  *
1147  * We only want to execute the splz() on the 1->0 transition of
1148  * critcount and not in a hard code section or if too deeply nested.
1149  */
1150 void
1151 lwkt_maybe_splz(thread_t td)
1152 {
1153     globaldata_t gd = td->td_gd;
1154 
1155     if (td->td_critcount == 0 &&
1156 	gd->gd_intr_nesting_level == 0 &&
1157 	td->td_nest_count < 2)
1158     {
1159 	splz();
1160     }
1161 }
1162 
1163 /*
1164  * Drivers which set up processing co-threads can call this function to
1165  * run the co-thread at a higher priority and to allow it to preempt
1166  * normal threads.
1167  */
1168 void
1169 lwkt_set_interrupt_support_thread(void)
1170 {
1171 	thread_t td = curthread;
1172 
1173         lwkt_setpri_self(TDPRI_INT_SUPPORT);
1174 	td->td_flags |= TDF_INTTHREAD;
1175 	td->td_preemptable = lwkt_preempt;
1176 }
1177 
1178 
1179 /*
1180  * This function is used to negotiate a passive release of the current
1181  * process/lwp designation with the user scheduler, allowing the user
1182  * scheduler to schedule another user thread.  The related kernel thread
1183  * (curthread) continues running in the released state.
1184  */
1185 void
1186 lwkt_passive_release(struct thread *td)
1187 {
1188     struct lwp *lp = td->td_lwp;
1189 
1190     td->td_release = NULL;
1191     lwkt_setpri_self(TDPRI_KERN_USER);
1192     lp->lwp_proc->p_usched->release_curproc(lp);
1193 }
1194 
1195 
1196 /*
1197  * This implements a LWKT yield, allowing a kernel thread to yield to other
1198  * kernel threads at the same or higher priority.  This function can be
1199  * called in a tight loop and will typically only yield once per tick.
1200  *
1201  * Most kernel threads run at the same priority in order to allow equal
1202  * sharing.
1203  *
1204  * (self contained on a per cpu basis)
1205  */
1206 void
1207 lwkt_yield(void)
1208 {
1209     globaldata_t gd = mycpu;
1210     thread_t td = gd->gd_curthread;
1211 
1212     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1213 	splz();
1214     if (lwkt_resched_wanted()) {
1215 	lwkt_schedule_self(curthread);
1216 	lwkt_switch();
1217     }
1218 }
1219 
1220 /*
1221  * This yield is designed for kernel threads with a user context.
1222  *
1223  * The kernel acting on behalf of the user is potentially cpu-bound,
1224  * this function will efficiently allow other threads to run and also
1225  * switch to other processes by releasing.
1226  *
1227  * The lwkt_user_yield() function is designed to have very low overhead
1228  * if no yield is determined to be needed.
1229  */
1230 void
1231 lwkt_user_yield(void)
1232 {
1233     globaldata_t gd = mycpu;
1234     thread_t td = gd->gd_curthread;
1235 
1236     /*
1237      * Always run any pending interrupts in case we are in a critical
1238      * section.
1239      */
1240     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1241 	splz();
1242 
1243     /*
1244      * Switch (which forces a release) if another kernel thread needs
1245      * the cpu, if userland wants us to resched, or if our kernel
1246      * quantum has run out.
1247      */
1248     if (lwkt_resched_wanted() ||
1249 	user_resched_wanted())
1250     {
1251 	lwkt_switch();
1252     }
1253 
1254 #if 0
1255     /*
1256      * Reacquire the current process if we are released.
1257      *
1258      * XXX not implemented atm.  The kernel may be holding locks and such,
1259      *     so we want the thread to continue to receive cpu.
1260      */
1261     if (td->td_release == NULL && lp) {
1262 	lp->lwp_proc->p_usched->acquire_curproc(lp);
1263 	td->td_release = lwkt_passive_release;
1264 	lwkt_setpri_self(TDPRI_USER_NORM);
1265     }
1266 #endif
1267 }
1268 
1269 /*
1270  * Generic schedule.  Possibly schedule threads belonging to other cpus and
1271  * deal with threads that might be blocked on a wait queue.
1272  *
1273  * We have a little helper inline function which does additional work after
1274  * the thread has been enqueued, including dealing with preemption and
1275  * setting need_lwkt_resched() (which prevents the kernel from returning
1276  * to userland until it has processed higher priority threads).
1277  *
1278  * It is possible for this routine to be called after a failed _enqueue
1279  * (due to the target thread migrating, sleeping, or otherwise blocked).
1280  * We have to check that the thread is actually on the run queue!
1281  */
1282 static __inline
1283 void
1284 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1285 {
1286     if (ntd->td_flags & TDF_RUNQ) {
1287 	if (ntd->td_preemptable) {
1288 	    ntd->td_preemptable(ntd, ccount);	/* YYY +token */
1289 	}
1290     }
1291 }
1292 
1293 static __inline
1294 void
1295 _lwkt_schedule(thread_t td)
1296 {
1297     globaldata_t mygd = mycpu;
1298 
1299     KASSERT(td != &td->td_gd->gd_idlethread,
1300 	    ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1301     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1302     crit_enter_gd(mygd);
1303     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1304     if (td == mygd->gd_curthread) {
1305 	_lwkt_enqueue(td);
1306     } else {
1307 	/*
1308 	 * If we own the thread, there is no race (since we are in a
1309 	 * critical section).  If we do not own the thread there might
1310 	 * be a race but the target cpu will deal with it.
1311 	 */
1312 #ifdef SMP
1313 	if (td->td_gd == mygd) {
1314 	    _lwkt_enqueue(td);
1315 	    _lwkt_schedule_post(mygd, td, 1);
1316 	} else {
1317 	    lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1318 	}
1319 #else
1320 	_lwkt_enqueue(td);
1321 	_lwkt_schedule_post(mygd, td, 1);
1322 #endif
1323     }
1324     crit_exit_gd(mygd);
1325 }
1326 
1327 void
1328 lwkt_schedule(thread_t td)
1329 {
1330     _lwkt_schedule(td);
1331 }
1332 
1333 void
1334 lwkt_schedule_noresched(thread_t td)	/* XXX not impl */
1335 {
1336     _lwkt_schedule(td);
1337 }
1338 
1339 #ifdef SMP
1340 
1341 /*
1342  * When scheduled remotely if frame != NULL the IPIQ is being
1343  * run via doreti or an interrupt then preemption can be allowed.
1344  *
1345  * To allow preemption we have to drop the critical section so only
1346  * one is present in _lwkt_schedule_post.
1347  */
1348 static void
1349 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1350 {
1351     thread_t td = curthread;
1352     thread_t ntd = arg;
1353 
1354     if (frame && ntd->td_preemptable) {
1355 	crit_exit_noyield(td);
1356 	_lwkt_schedule(ntd);
1357 	crit_enter_quick(td);
1358     } else {
1359 	_lwkt_schedule(ntd);
1360     }
1361 }
1362 
1363 /*
1364  * Thread migration using a 'Pull' method.  The thread may or may not be
1365  * the current thread.  It MUST be descheduled and in a stable state.
1366  * lwkt_giveaway() must be called on the cpu owning the thread.
1367  *
1368  * At any point after lwkt_giveaway() is called, the target cpu may
1369  * 'pull' the thread by calling lwkt_acquire().
1370  *
1371  * We have to make sure the thread is not sitting on a per-cpu tsleep
1372  * queue or it will blow up when it moves to another cpu.
1373  *
1374  * MPSAFE - must be called under very specific conditions.
1375  */
1376 void
1377 lwkt_giveaway(thread_t td)
1378 {
1379     globaldata_t gd = mycpu;
1380 
1381     crit_enter_gd(gd);
1382     if (td->td_flags & TDF_TSLEEPQ)
1383 	tsleep_remove(td);
1384     KKASSERT(td->td_gd == gd);
1385     TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1386     td->td_flags |= TDF_MIGRATING;
1387     crit_exit_gd(gd);
1388 }
1389 
1390 void
1391 lwkt_acquire(thread_t td)
1392 {
1393     globaldata_t gd;
1394     globaldata_t mygd;
1395     int retry = 10000000;
1396 
1397     KKASSERT(td->td_flags & TDF_MIGRATING);
1398     gd = td->td_gd;
1399     mygd = mycpu;
1400     if (gd != mycpu) {
1401 	cpu_lfence();
1402 	KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1403 	crit_enter_gd(mygd);
1404 	DEBUG_PUSH_INFO("lwkt_acquire");
1405 	while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1406 #ifdef SMP
1407 	    lwkt_process_ipiq();
1408 #endif
1409 	    cpu_lfence();
1410 	    if (--retry == 0) {
1411 		kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1412 			td, td->td_flags);
1413 		retry = 10000000;
1414 	    }
1415 	}
1416 	DEBUG_POP_INFO();
1417 	cpu_mfence();
1418 	td->td_gd = mygd;
1419 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1420 	td->td_flags &= ~TDF_MIGRATING;
1421 	crit_exit_gd(mygd);
1422     } else {
1423 	crit_enter_gd(mygd);
1424 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1425 	td->td_flags &= ~TDF_MIGRATING;
1426 	crit_exit_gd(mygd);
1427     }
1428 }
1429 
1430 #endif
1431 
1432 /*
1433  * Generic deschedule.  Descheduling threads other then your own should be
1434  * done only in carefully controlled circumstances.  Descheduling is
1435  * asynchronous.
1436  *
1437  * This function may block if the cpu has run out of messages.
1438  */
1439 void
1440 lwkt_deschedule(thread_t td)
1441 {
1442     crit_enter();
1443 #ifdef SMP
1444     if (td == curthread) {
1445 	_lwkt_dequeue(td);
1446     } else {
1447 	if (td->td_gd == mycpu) {
1448 	    _lwkt_dequeue(td);
1449 	} else {
1450 	    lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1451 	}
1452     }
1453 #else
1454     _lwkt_dequeue(td);
1455 #endif
1456     crit_exit();
1457 }
1458 
1459 /*
1460  * Set the target thread's priority.  This routine does not automatically
1461  * switch to a higher priority thread, LWKT threads are not designed for
1462  * continuous priority changes.  Yield if you want to switch.
1463  */
1464 void
1465 lwkt_setpri(thread_t td, int pri)
1466 {
1467     if (td->td_pri != pri) {
1468 	KKASSERT(pri >= 0);
1469 	crit_enter();
1470 	if (td->td_flags & TDF_RUNQ) {
1471 	    KKASSERT(td->td_gd == mycpu);
1472 	    _lwkt_dequeue(td);
1473 	    td->td_pri = pri;
1474 	    _lwkt_enqueue(td);
1475 	} else {
1476 	    td->td_pri = pri;
1477 	}
1478 	crit_exit();
1479     }
1480 }
1481 
1482 /*
1483  * Set the initial priority for a thread prior to it being scheduled for
1484  * the first time.  The thread MUST NOT be scheduled before or during
1485  * this call.  The thread may be assigned to a cpu other then the current
1486  * cpu.
1487  *
1488  * Typically used after a thread has been created with TDF_STOPPREQ,
1489  * and before the thread is initially scheduled.
1490  */
1491 void
1492 lwkt_setpri_initial(thread_t td, int pri)
1493 {
1494     KKASSERT(pri >= 0);
1495     KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1496     td->td_pri = pri;
1497 }
1498 
1499 void
1500 lwkt_setpri_self(int pri)
1501 {
1502     thread_t td = curthread;
1503 
1504     KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1505     crit_enter();
1506     if (td->td_flags & TDF_RUNQ) {
1507 	_lwkt_dequeue(td);
1508 	td->td_pri = pri;
1509 	_lwkt_enqueue(td);
1510     } else {
1511 	td->td_pri = pri;
1512     }
1513     crit_exit();
1514 }
1515 
1516 /*
1517  * hz tick scheduler clock for LWKT threads
1518  */
1519 void
1520 lwkt_schedulerclock(thread_t td)
1521 {
1522     globaldata_t gd = td->td_gd;
1523     thread_t xtd;
1524 
1525     if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1526 	/*
1527 	 * If the current thread is at the head of the runq shift it to the
1528 	 * end of any equal-priority threads and request a LWKT reschedule
1529 	 * if it moved.
1530 	 */
1531 	xtd = TAILQ_NEXT(td, td_threadq);
1532 	if (xtd && xtd->td_pri == td->td_pri) {
1533 	    TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1534 	    while (xtd && xtd->td_pri == td->td_pri)
1535 		xtd = TAILQ_NEXT(xtd, td_threadq);
1536 	    if (xtd)
1537 		TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1538 	    else
1539 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1540 	    need_lwkt_resched();
1541 	}
1542     } else {
1543 	/*
1544 	 * If we scheduled a thread other than the one at the head of the
1545 	 * queue always request a reschedule every tick.
1546 	 */
1547 	need_lwkt_resched();
1548     }
1549 }
1550 
1551 /*
1552  * Migrate the current thread to the specified cpu.
1553  *
1554  * This is accomplished by descheduling ourselves from the current cpu
1555  * and setting td_migrate_gd.  The lwkt_switch() code will detect that the
1556  * 'old' thread wants to migrate after it has been completely switched out
1557  * and will complete the migration.
1558  *
1559  * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1560  *
1561  * We must be sure to release our current process designation (if a user
1562  * process) before clearing out any tsleepq we are on because the release
1563  * code may re-add us.
1564  *
1565  * We must be sure to remove ourselves from the current cpu's tsleepq
1566  * before potentially moving to another queue.  The thread can be on
1567  * a tsleepq due to a left-over tsleep_interlock().
1568  */
1569 
1570 void
1571 lwkt_setcpu_self(globaldata_t rgd)
1572 {
1573 #ifdef SMP
1574     thread_t td = curthread;
1575 
1576     if (td->td_gd != rgd) {
1577 	crit_enter_quick(td);
1578 
1579 	if (td->td_release)
1580 	    td->td_release(td);
1581 	if (td->td_flags & TDF_TSLEEPQ)
1582 	    tsleep_remove(td);
1583 
1584 	/*
1585 	 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1586 	 * trying to deschedule ourselves and switch away, then deschedule
1587 	 * ourself, remove us from tdallq, and set td_migrate_gd.  Finally,
1588 	 * call lwkt_switch() to complete the operation.
1589 	 */
1590 	td->td_flags |= TDF_MIGRATING;
1591 	lwkt_deschedule_self(td);
1592 	TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1593 	td->td_migrate_gd = rgd;
1594 	lwkt_switch();
1595 
1596 	/*
1597 	 * We are now on the target cpu
1598 	 */
1599 	KKASSERT(rgd == mycpu);
1600 	TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1601 	crit_exit_quick(td);
1602     }
1603 #endif
1604 }
1605 
1606 void
1607 lwkt_migratecpu(int cpuid)
1608 {
1609 #ifdef SMP
1610 	globaldata_t rgd;
1611 
1612 	rgd = globaldata_find(cpuid);
1613 	lwkt_setcpu_self(rgd);
1614 #endif
1615 }
1616 
1617 #ifdef SMP
1618 /*
1619  * Remote IPI for cpu migration (called while in a critical section so we
1620  * do not have to enter another one).
1621  *
1622  * The thread (td) has already been completely descheduled from the
1623  * originating cpu and we can simply assert the case.  The thread is
1624  * assigned to the new cpu and enqueued.
1625  *
1626  * The thread will re-add itself to tdallq when it resumes execution.
1627  */
1628 static void
1629 lwkt_setcpu_remote(void *arg)
1630 {
1631     thread_t td = arg;
1632     globaldata_t gd = mycpu;
1633 
1634     KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1635     td->td_gd = gd;
1636     cpu_mfence();
1637     td->td_flags &= ~TDF_MIGRATING;
1638     KKASSERT(td->td_migrate_gd == NULL);
1639     KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1640     _lwkt_enqueue(td);
1641 }
1642 #endif
1643 
1644 struct lwp *
1645 lwkt_preempted_proc(void)
1646 {
1647     thread_t td = curthread;
1648     while (td->td_preempted)
1649 	td = td->td_preempted;
1650     return(td->td_lwp);
1651 }
1652 
1653 /*
1654  * Create a kernel process/thread/whatever.  It shares it's address space
1655  * with proc0 - ie: kernel only.
1656  *
1657  * If the cpu is not specified one will be selected.  In the future
1658  * specifying a cpu of -1 will enable kernel thread migration between
1659  * cpus.
1660  */
1661 int
1662 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1663 	    thread_t template, int tdflags, int cpu, const char *fmt, ...)
1664 {
1665     thread_t td;
1666     __va_list ap;
1667 
1668     td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1669 			   tdflags);
1670     if (tdp)
1671 	*tdp = td;
1672     cpu_set_thread_handler(td, lwkt_exit, func, arg);
1673 
1674     /*
1675      * Set up arg0 for 'ps' etc
1676      */
1677     __va_start(ap, fmt);
1678     kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1679     __va_end(ap);
1680 
1681     /*
1682      * Schedule the thread to run
1683      */
1684     if ((td->td_flags & TDF_STOPREQ) == 0)
1685 	lwkt_schedule(td);
1686     else
1687 	td->td_flags &= ~TDF_STOPREQ;
1688     return 0;
1689 }
1690 
1691 /*
1692  * Destroy an LWKT thread.   Warning!  This function is not called when
1693  * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1694  * uses a different reaping mechanism.
1695  */
1696 void
1697 lwkt_exit(void)
1698 {
1699     thread_t td = curthread;
1700     thread_t std;
1701     globaldata_t gd;
1702 
1703     /*
1704      * Do any cleanup that might block here
1705      */
1706     if (td->td_flags & TDF_VERBOSE)
1707 	kprintf("kthread %p %s has exited\n", td, td->td_comm);
1708     caps_exit(td);
1709     biosched_done(td);
1710     dsched_exit_thread(td);
1711 
1712     /*
1713      * Get us into a critical section to interlock gd_freetd and loop
1714      * until we can get it freed.
1715      *
1716      * We have to cache the current td in gd_freetd because objcache_put()ing
1717      * it would rip it out from under us while our thread is still active.
1718      */
1719     gd = mycpu;
1720     crit_enter_quick(td);
1721     while ((std = gd->gd_freetd) != NULL) {
1722 	KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1723 	gd->gd_freetd = NULL;
1724 	objcache_put(thread_cache, std);
1725     }
1726 
1727     /*
1728      * Remove thread resources from kernel lists and deschedule us for
1729      * the last time.  We cannot block after this point or we may end
1730      * up with a stale td on the tsleepq.
1731      */
1732     if (td->td_flags & TDF_TSLEEPQ)
1733 	tsleep_remove(td);
1734     lwkt_deschedule_self(td);
1735     lwkt_remove_tdallq(td);
1736     KKASSERT(td->td_refs == 0);
1737 
1738     /*
1739      * Final cleanup
1740      */
1741     KKASSERT(gd->gd_freetd == NULL);
1742     if (td->td_flags & TDF_ALLOCATED_THREAD)
1743 	gd->gd_freetd = td;
1744     cpu_thread_exit();
1745 }
1746 
1747 void
1748 lwkt_remove_tdallq(thread_t td)
1749 {
1750     KKASSERT(td->td_gd == mycpu);
1751     TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1752 }
1753 
1754 /*
1755  * Code reduction and branch prediction improvements.  Call/return
1756  * overhead on modern cpus often degenerates into 0 cycles due to
1757  * the cpu's branch prediction hardware and return pc cache.  We
1758  * can take advantage of this by not inlining medium-complexity
1759  * functions and we can also reduce the branch prediction impact
1760  * by collapsing perfectly predictable branches into a single
1761  * procedure instead of duplicating it.
1762  *
1763  * Is any of this noticeable?  Probably not, so I'll take the
1764  * smaller code size.
1765  */
1766 void
1767 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1768 {
1769     _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1770 }
1771 
1772 void
1773 crit_panic(void)
1774 {
1775     thread_t td = curthread;
1776     int lcrit = td->td_critcount;
1777 
1778     td->td_critcount = 0;
1779     panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1780     /* NOT REACHED */
1781 }
1782 
1783 #ifdef SMP
1784 
1785 /*
1786  * Called from debugger/panic on cpus which have been stopped.  We must still
1787  * process the IPIQ while stopped, even if we were stopped while in a critical
1788  * section (XXX).
1789  *
1790  * If we are dumping also try to process any pending interrupts.  This may
1791  * or may not work depending on the state of the cpu at the point it was
1792  * stopped.
1793  */
1794 void
1795 lwkt_smp_stopped(void)
1796 {
1797     globaldata_t gd = mycpu;
1798 
1799     crit_enter_gd(gd);
1800     if (dumping) {
1801 	lwkt_process_ipiq();
1802 	splz();
1803     } else {
1804 	lwkt_process_ipiq();
1805     }
1806     crit_exit_gd(gd);
1807 }
1808 
1809 #endif
1810