xref: /dflybsd-src/sys/kern/lwkt_thread.c (revision a563ca70e68142ccf7f50a6f129665fd8cb66d98)
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 	td->td_mpflags = 0;
295 	return (1);
296 }
297 
298 static void
299 _lwkt_thread_dtor(void *obj, void *privdata)
300 {
301 	struct thread *td = (struct thread *)obj;
302 
303 	KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
304 	    ("_lwkt_thread_dtor: not allocated from objcache"));
305 	KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
306 		td->td_kstack_size > 0,
307 	    ("_lwkt_thread_dtor: corrupted stack"));
308 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
309 }
310 
311 /*
312  * Initialize the lwkt s/system.
313  *
314  * Nominally cache up to 32 thread + kstack structures.
315  */
316 void
317 lwkt_init(void)
318 {
319     TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
320     thread_cache = objcache_create_mbacked(
321 				M_THREAD, sizeof(struct thread),
322 				NULL, lwkt_cache_threads,
323 				_lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
324 }
325 
326 /*
327  * Schedule a thread to run.  As the current thread we can always safely
328  * schedule ourselves, and a shortcut procedure is provided for that
329  * function.
330  *
331  * (non-blocking, self contained on a per cpu basis)
332  */
333 void
334 lwkt_schedule_self(thread_t td)
335 {
336     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
337     crit_enter_quick(td);
338     KASSERT(td != &td->td_gd->gd_idlethread,
339 	    ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
340     KKASSERT(td->td_lwp == NULL ||
341 	     (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
342     _lwkt_enqueue(td);
343     crit_exit_quick(td);
344 }
345 
346 /*
347  * Deschedule a thread.
348  *
349  * (non-blocking, self contained on a per cpu basis)
350  */
351 void
352 lwkt_deschedule_self(thread_t td)
353 {
354     crit_enter_quick(td);
355     _lwkt_dequeue(td);
356     crit_exit_quick(td);
357 }
358 
359 /*
360  * LWKTs operate on a per-cpu basis
361  *
362  * WARNING!  Called from early boot, 'mycpu' may not work yet.
363  */
364 void
365 lwkt_gdinit(struct globaldata *gd)
366 {
367     TAILQ_INIT(&gd->gd_tdrunq);
368     TAILQ_INIT(&gd->gd_tdallq);
369 }
370 
371 /*
372  * Create a new thread.  The thread must be associated with a process context
373  * or LWKT start address before it can be scheduled.  If the target cpu is
374  * -1 the thread will be created on the current cpu.
375  *
376  * If you intend to create a thread without a process context this function
377  * does everything except load the startup and switcher function.
378  */
379 thread_t
380 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
381 {
382     static int cpu_rotator;
383     globaldata_t gd = mycpu;
384     void *stack;
385 
386     /*
387      * If static thread storage is not supplied allocate a thread.  Reuse
388      * a cached free thread if possible.  gd_freetd is used to keep an exiting
389      * thread intact through the exit.
390      */
391     if (td == NULL) {
392 	crit_enter_gd(gd);
393 	if ((td = gd->gd_freetd) != NULL) {
394 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
395 				      TDF_RUNQ)) == 0);
396 	    gd->gd_freetd = NULL;
397 	} else {
398 	    td = objcache_get(thread_cache, M_WAITOK);
399 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
400 				      TDF_RUNQ)) == 0);
401 	}
402 	crit_exit_gd(gd);
403     	KASSERT((td->td_flags &
404 		 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
405 		("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
406     	flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
407     }
408 
409     /*
410      * Try to reuse cached stack.
411      */
412     if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
413 	if (flags & TDF_ALLOCATED_STACK) {
414 	    kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
415 	    stack = NULL;
416 	}
417     }
418     if (stack == NULL) {
419 	stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
420 	flags |= TDF_ALLOCATED_STACK;
421     }
422     if (cpu < 0) {
423 	cpu = ++cpu_rotator;
424 	cpu_ccfence();
425 	cpu %= ncpus;
426     }
427     lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
428     return(td);
429 }
430 
431 /*
432  * Initialize a preexisting thread structure.  This function is used by
433  * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
434  *
435  * All threads start out in a critical section at a priority of
436  * TDPRI_KERN_DAEMON.  Higher level code will modify the priority as
437  * appropriate.  This function may send an IPI message when the
438  * requested cpu is not the current cpu and consequently gd_tdallq may
439  * not be initialized synchronously from the point of view of the originating
440  * cpu.
441  *
442  * NOTE! we have to be careful in regards to creating threads for other cpus
443  * if SMP has not yet been activated.
444  */
445 #ifdef SMP
446 
447 static void
448 lwkt_init_thread_remote(void *arg)
449 {
450     thread_t td = arg;
451 
452     /*
453      * Protected by critical section held by IPI dispatch
454      */
455     TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
456 }
457 
458 #endif
459 
460 /*
461  * lwkt core thread structural initialization.
462  *
463  * NOTE: All threads are initialized as mpsafe threads.
464  */
465 void
466 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
467 		struct globaldata *gd)
468 {
469     globaldata_t mygd = mycpu;
470 
471     bzero(td, sizeof(struct thread));
472     td->td_kstack = stack;
473     td->td_kstack_size = stksize;
474     td->td_flags = flags;
475     td->td_mpflags = 0;
476     td->td_gd = gd;
477     td->td_pri = TDPRI_KERN_DAEMON;
478     td->td_critcount = 1;
479     td->td_toks_have = NULL;
480     td->td_toks_stop = &td->td_toks_base;
481     if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
482 	lwkt_initport_spin(&td->td_msgport);
483     else
484 	lwkt_initport_thread(&td->td_msgport, td);
485     pmap_init_thread(td);
486 #ifdef SMP
487     /*
488      * Normally initializing a thread for a remote cpu requires sending an
489      * IPI.  However, the idlethread is setup before the other cpus are
490      * activated so we have to treat it as a special case.  XXX manipulation
491      * of gd_tdallq requires the BGL.
492      */
493     if (gd == mygd || td == &gd->gd_idlethread) {
494 	crit_enter_gd(mygd);
495 	TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
496 	crit_exit_gd(mygd);
497     } else {
498 	lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
499     }
500 #else
501     crit_enter_gd(mygd);
502     TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
503     crit_exit_gd(mygd);
504 #endif
505 
506     dsched_new_thread(td);
507 }
508 
509 void
510 lwkt_set_comm(thread_t td, const char *ctl, ...)
511 {
512     __va_list va;
513 
514     __va_start(va, ctl);
515     kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
516     __va_end(va);
517     KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
518 }
519 
520 /*
521  * Prevent the thread from getting destroyed.  Note that unlike PHOLD/PRELE
522  * this does not prevent the thread from migrating to another cpu so the
523  * gd_tdallq state is not protected by this.
524  */
525 void
526 lwkt_hold(thread_t td)
527 {
528     atomic_add_int(&td->td_refs, 1);
529 }
530 
531 void
532 lwkt_rele(thread_t td)
533 {
534     KKASSERT(td->td_refs > 0);
535     atomic_add_int(&td->td_refs, -1);
536 }
537 
538 void
539 lwkt_wait_free(thread_t td)
540 {
541     while (td->td_refs)
542 	tsleep(td, 0, "tdreap", hz);
543 }
544 
545 void
546 lwkt_free_thread(thread_t td)
547 {
548     KKASSERT(td->td_refs == 0);
549     KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
550 			      TDF_RUNQ | TDF_TSLEEPQ)) == 0);
551     if (td->td_flags & TDF_ALLOCATED_THREAD) {
552     	objcache_put(thread_cache, td);
553     } else if (td->td_flags & TDF_ALLOCATED_STACK) {
554 	/* client-allocated struct with internally allocated stack */
555 	KASSERT(td->td_kstack && td->td_kstack_size > 0,
556 	    ("lwkt_free_thread: corrupted stack"));
557 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
558 	td->td_kstack = NULL;
559 	td->td_kstack_size = 0;
560     }
561     KTR_LOG(ctxsw_deadtd, td);
562 }
563 
564 
565 /*
566  * Switch to the next runnable lwkt.  If no LWKTs are runnable then
567  * switch to the idlethread.  Switching must occur within a critical
568  * section to avoid races with the scheduling queue.
569  *
570  * We always have full control over our cpu's run queue.  Other cpus
571  * that wish to manipulate our queue must use the cpu_*msg() calls to
572  * talk to our cpu, so a critical section is all that is needed and
573  * the result is very, very fast thread switching.
574  *
575  * The LWKT scheduler uses a fixed priority model and round-robins at
576  * each priority level.  User process scheduling is a totally
577  * different beast and LWKT priorities should not be confused with
578  * user process priorities.
579  *
580  * PREEMPTION NOTE: Preemption occurs via lwkt_preempt().  lwkt_switch()
581  * is not called by the current thread in the preemption case, only when
582  * the preempting thread blocks (in order to return to the original thread).
583  *
584  * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
585  * migration and tsleep deschedule the current lwkt thread and call
586  * lwkt_switch().  In particular, the target cpu of the migration fully
587  * expects the thread to become non-runnable and can deadlock against
588  * cpusync operations if we run any IPIs prior to switching the thread out.
589  *
590  * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
591  * THE CURRENT THREAD HAS BEEN DESCHEDULED!
592  */
593 void
594 lwkt_switch(void)
595 {
596     globaldata_t gd = mycpu;
597     thread_t td = gd->gd_curthread;
598     thread_t ntd;
599     thread_t xtd;
600     int spinning = 0;
601 
602     KKASSERT(gd->gd_processing_ipiq == 0);
603 
604     /*
605      * Switching from within a 'fast' (non thread switched) interrupt or IPI
606      * is illegal.  However, we may have to do it anyway if we hit a fatal
607      * kernel trap or we have paniced.
608      *
609      * If this case occurs save and restore the interrupt nesting level.
610      */
611     if (gd->gd_intr_nesting_level) {
612 	int savegdnest;
613 	int savegdtrap;
614 
615 	if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
616 	    panic("lwkt_switch: Attempt to switch from a "
617 		  "a fast interrupt, ipi, or hard code section, "
618 		  "td %p\n",
619 		  td);
620 	} else {
621 	    savegdnest = gd->gd_intr_nesting_level;
622 	    savegdtrap = gd->gd_trap_nesting_level;
623 	    gd->gd_intr_nesting_level = 0;
624 	    gd->gd_trap_nesting_level = 0;
625 	    if ((td->td_flags & TDF_PANICWARN) == 0) {
626 		td->td_flags |= TDF_PANICWARN;
627 		kprintf("Warning: thread switch from interrupt, IPI, "
628 			"or hard code section.\n"
629 			"thread %p (%s)\n", td, td->td_comm);
630 		print_backtrace(-1);
631 	    }
632 	    lwkt_switch();
633 	    gd->gd_intr_nesting_level = savegdnest;
634 	    gd->gd_trap_nesting_level = savegdtrap;
635 	    return;
636 	}
637     }
638 
639     /*
640      * Release our current user process designation if we are blocking
641      * or if a user reschedule was requested.
642      *
643      * NOTE: This function is NOT called if we are switching into or
644      *	     returning from a preemption.
645      *
646      * NOTE: Releasing our current user process designation may cause
647      *	     it to be assigned to another thread, which in turn will
648      *	     cause us to block in the usched acquire code when we attempt
649      *	     to return to userland.
650      *
651      * NOTE: On SMP systems this can be very nasty when heavy token
652      *	     contention is present so we want to be careful not to
653      *	     release the designation gratuitously.
654      */
655     if (td->td_release &&
656 	(user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
657 	    td->td_release(td);
658     }
659 
660     /*
661      * Release all tokens
662      */
663     crit_enter_gd(gd);
664     if (TD_TOKS_HELD(td))
665 	    lwkt_relalltokens(td);
666 
667     /*
668      * We had better not be holding any spin locks, but don't get into an
669      * endless panic loop.
670      */
671     KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
672 	    ("lwkt_switch: still holding %d exclusive spinlocks!",
673 	     gd->gd_spinlocks_wr));
674 
675 
676 #ifdef SMP
677 #ifdef	INVARIANTS
678     if (td->td_cscount) {
679 	kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
680 		td);
681 	if (panic_on_cscount)
682 	    panic("switching while mastering cpusync");
683     }
684 #endif
685 #endif
686 
687     /*
688      * If we had preempted another thread on this cpu, resume the preempted
689      * thread.  This occurs transparently, whether the preempted thread
690      * was scheduled or not (it may have been preempted after descheduling
691      * itself).
692      *
693      * We have to setup the MP lock for the original thread after backing
694      * out the adjustment that was made to curthread when the original
695      * was preempted.
696      */
697     if ((ntd = td->td_preempted) != NULL) {
698 	KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
699 	ntd->td_flags |= TDF_PREEMPT_DONE;
700 
701 	/*
702 	 * The interrupt may have woken a thread up, we need to properly
703 	 * set the reschedule flag if the originally interrupted thread is
704 	 * at a lower priority.
705 	 *
706 	 * The interrupt may not have descheduled.
707 	 */
708 	if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
709 	    need_lwkt_resched();
710 	goto havethread_preempted;
711     }
712 
713     /*
714      * If we cannot obtain ownership of the tokens we cannot immediately
715      * schedule the target thread.
716      *
717      * Reminder: Again, we cannot afford to run any IPIs in this path if
718      * the current thread has been descheduled.
719      */
720     for (;;) {
721 	clear_lwkt_resched();
722 
723 	/*
724 	 * Hotpath - pull the head of the run queue and attempt to schedule
725 	 * it.
726 	 */
727 	for (;;) {
728 	    ntd = TAILQ_FIRST(&gd->gd_tdrunq);
729 
730 	    if (ntd == NULL) {
731 		/*
732 		 * Runq is empty, switch to idle to allow it to halt.
733 		 */
734 		ntd = &gd->gd_idlethread;
735 #ifdef SMP
736 		if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
737 		    ASSERT_NO_TOKENS_HELD(ntd);
738 #endif
739 		cpu_time.cp_msg[0] = 0;
740 		cpu_time.cp_stallpc = 0;
741 		goto haveidle;
742 	    }
743 	    break;
744 	}
745 
746 	/*
747 	 * Hotpath - schedule ntd.
748 	 *
749 	 * NOTE: For UP there is no mplock and lwkt_getalltokens()
750 	 *	     always succeeds.
751 	 */
752 	if (TD_TOKS_NOT_HELD(ntd) ||
753 	    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
754 	{
755 	    goto havethread;
756 	}
757 
758 	/*
759 	 * Coldpath (SMP only since tokens always succeed on UP)
760 	 *
761 	 * We had some contention on the thread we wanted to schedule.
762 	 * What we do now is try to find a thread that we can schedule
763 	 * in its stead.
764 	 *
765 	 * The coldpath scan does NOT rearrange threads in the run list.
766 	 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
767 	 * the next tick whenever the current head is not the current thread.
768 	 */
769 #ifdef	INVARIANTS
770 	++token_contention_count[ntd->td_pri];
771 	++ntd->td_contended;
772 #endif
773 
774 	if (fairq_bypass > 0)
775 		goto skip;
776 
777 	xtd = NULL;
778 	while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
779 		/*
780 		 * Never schedule threads returning to userland or the
781 		 * user thread scheduler helper thread when higher priority
782 		 * threads are present.
783 		 */
784 		if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
785 			ntd = NULL;
786 			break;
787 		}
788 
789 		/*
790 		 * Try this one.
791 		 */
792 		if (TD_TOKS_NOT_HELD(ntd) ||
793 		    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
794 			goto havethread;
795 		}
796 #ifdef	INVARIANTS
797 		++token_contention_count[ntd->td_pri];
798 		++ntd->td_contended;
799 #endif
800 	}
801 
802 skip:
803 	/*
804 	 * We exhausted the run list, meaning that all runnable threads
805 	 * are contested.
806 	 */
807 	cpu_pause();
808 	ntd = &gd->gd_idlethread;
809 #ifdef SMP
810 	if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
811 	    ASSERT_NO_TOKENS_HELD(ntd);
812 	/* contention case, do not clear contention mask */
813 #endif
814 
815 	/*
816 	 * We are going to have to retry but if the current thread is not
817 	 * on the runq we instead switch through the idle thread to get away
818 	 * from the current thread.  We have to flag for lwkt reschedule
819 	 * to prevent the idle thread from halting.
820 	 *
821 	 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
822 	 *	 instruct it to deal with the potential for deadlocks by
823 	 *	 ordering the tokens by address.
824 	 */
825 	if ((td->td_flags & TDF_RUNQ) == 0) {
826 	    need_lwkt_resched();	/* prevent hlt */
827 	    goto haveidle;
828 	}
829 #if defined(INVARIANTS) && defined(__amd64__)
830 	if ((read_rflags() & PSL_I) == 0) {
831 		cpu_enable_intr();
832 		panic("lwkt_switch() called with interrupts disabled");
833 	}
834 #endif
835 
836 	/*
837 	 * Number iterations so far.  After a certain point we switch to
838 	 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
839 	 */
840 	if (spinning < 0x7FFFFFFF)
841 	    ++spinning;
842 
843 #ifdef SMP
844 	/*
845 	 * lwkt_getalltokens() failed in sorted token mode, we can use
846 	 * monitor/mwait in this case.
847 	 */
848 	if (spinning >= lwkt_spin_loops &&
849 	    (cpu_mi_feature & CPU_MI_MONITOR) &&
850 	    lwkt_spin_monitor)
851 	{
852 	    cpu_mmw_pause_int(&gd->gd_reqflags,
853 			      (gd->gd_reqflags | RQF_SPINNING) &
854 			      ~RQF_IDLECHECK_WK_MASK);
855 	}
856 #endif
857 
858 	/*
859 	 * We already checked that td is still scheduled so this should be
860 	 * safe.
861 	 */
862 	splz_check();
863 
864 	/*
865 	 * This experimental resequencer is used as a fall-back to reduce
866 	 * hw cache line contention by placing each core's scheduler into a
867 	 * time-domain-multplexed slot.
868 	 *
869 	 * The resequencer is disabled by default.  It's functionality has
870 	 * largely been superceeded by the token algorithm which limits races
871 	 * to a subset of cores.
872 	 *
873 	 * The resequencer algorithm tends to break down when more than
874 	 * 20 cores are contending.  What appears to happen is that new
875 	 * tokens can be obtained out of address-sorted order by new cores
876 	 * while existing cores languish in long delays between retries and
877 	 * wind up being starved-out of the token acquisition.
878 	 */
879 	if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
880 	    int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
881 	    int oseq;
882 
883 	    while ((oseq = lwkt_cseq_rindex) != cseq) {
884 		cpu_ccfence();
885 #if 1
886 		if (cpu_mi_feature & CPU_MI_MONITOR) {
887 		    cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
888 		} else {
889 #endif
890 		    cpu_pause();
891 		    cpu_lfence();
892 #if 1
893 		}
894 #endif
895 	    }
896 	    DELAY(1);
897 	    atomic_add_int(&lwkt_cseq_rindex, 1);
898 	}
899 	/* highest level for(;;) loop */
900     }
901 
902 havethread:
903     /*
904      * Clear gd_idle_repeat when doing a normal switch to a non-idle
905      * thread.
906      */
907     ntd->td_wmesg = NULL;
908     ++gd->gd_cnt.v_swtch;
909     gd->gd_idle_repeat = 0;
910 
911 havethread_preempted:
912     /*
913      * If the new target does not need the MP lock and we are holding it,
914      * release the MP lock.  If the new target requires the MP lock we have
915      * already acquired it for the target.
916      */
917     ;
918 haveidle:
919     KASSERT(ntd->td_critcount,
920 	    ("priority problem in lwkt_switch %d %d",
921 	    td->td_critcount, ntd->td_critcount));
922 
923     if (td != ntd) {
924 	/*
925 	 * Execute the actual thread switch operation.  This function
926 	 * returns to the current thread and returns the previous thread
927 	 * (which may be different from the thread we switched to).
928 	 *
929 	 * We are responsible for marking ntd as TDF_RUNNING.
930 	 */
931 	++switch_count;
932 	KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
933 	ntd->td_flags |= TDF_RUNNING;
934 	lwkt_switch_return(td->td_switch(ntd));
935 	/* ntd invalid, td_switch() can return a different thread_t */
936     }
937 
938     /*
939      * catch-all.  XXX is this strictly needed?
940      */
941     splz_check();
942 
943     /* NOTE: current cpu may have changed after switch */
944     crit_exit_quick(td);
945 }
946 
947 /*
948  * Called by assembly in the td_switch (thread restore path) for thread
949  * bootstrap cases which do not 'return' to lwkt_switch().
950  */
951 void
952 lwkt_switch_return(thread_t otd)
953 {
954 #ifdef SMP
955 	globaldata_t rgd;
956 
957 	/*
958 	 * Check if otd was migrating.  Now that we are on ntd we can finish
959 	 * up the migration.  This is a bit messy but it is the only place
960 	 * where td is known to be fully descheduled.
961 	 *
962 	 * We can only activate the migration if otd was migrating but not
963 	 * held on the cpu due to a preemption chain.  We still have to
964 	 * clear TDF_RUNNING on the old thread either way.
965 	 *
966 	 * We are responsible for clearing the previously running thread's
967 	 * TDF_RUNNING.
968 	 */
969 	if ((rgd = otd->td_migrate_gd) != NULL &&
970 	    (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
971 		KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
972 			 (TDF_MIGRATING | TDF_RUNNING));
973 		otd->td_migrate_gd = NULL;
974 		otd->td_flags &= ~TDF_RUNNING;
975 		lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
976 	} else {
977 		otd->td_flags &= ~TDF_RUNNING;
978 	}
979 #else
980 	otd->td_flags &= ~TDF_RUNNING;
981 #endif
982 }
983 
984 /*
985  * Request that the target thread preempt the current thread.  Preemption
986  * can only occur if our only critical section is the one that we were called
987  * with, the relative priority of the target thread is higher, and the target
988  * thread holds no tokens.  This also only works if we are not holding any
989  * spinlocks (obviously).
990  *
991  * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION.  Typically
992  * this is called via lwkt_schedule() through the td_preemptable callback.
993  * critcount is the managed critical priority that we should ignore in order
994  * to determine whether preemption is possible (aka usually just the crit
995  * priority of lwkt_schedule() itself).
996  *
997  * Preemption is typically limited to interrupt threads.
998  *
999  * Operation works in a fairly straight-forward manner.  The normal
1000  * scheduling code is bypassed and we switch directly to the target
1001  * thread.  When the target thread attempts to block or switch away
1002  * code at the base of lwkt_switch() will switch directly back to our
1003  * thread.  Our thread is able to retain whatever tokens it holds and
1004  * if the target needs one of them the target will switch back to us
1005  * and reschedule itself normally.
1006  */
1007 void
1008 lwkt_preempt(thread_t ntd, int critcount)
1009 {
1010     struct globaldata *gd = mycpu;
1011     thread_t xtd;
1012     thread_t td;
1013     int save_gd_intr_nesting_level;
1014 
1015     /*
1016      * The caller has put us in a critical section.  We can only preempt
1017      * if the caller of the caller was not in a critical section (basically
1018      * a local interrupt), as determined by the 'critcount' parameter.  We
1019      * also can't preempt if the caller is holding any spinlocks (even if
1020      * he isn't in a critical section).  This also handles the tokens test.
1021      *
1022      * YYY The target thread must be in a critical section (else it must
1023      * inherit our critical section?  I dunno yet).
1024      */
1025     KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1026 
1027     td = gd->gd_curthread;
1028     if (preempt_enable == 0) {
1029 	++preempt_miss;
1030 	return;
1031     }
1032     if (ntd->td_pri <= td->td_pri) {
1033 	++preempt_miss;
1034 	return;
1035     }
1036     if (td->td_critcount > critcount) {
1037 	++preempt_miss;
1038 	return;
1039     }
1040 #ifdef SMP
1041     if (ntd->td_gd != gd) {
1042 	++preempt_miss;
1043 	return;
1044     }
1045 #endif
1046     /*
1047      * We don't have to check spinlocks here as they will also bump
1048      * td_critcount.
1049      *
1050      * Do not try to preempt if the target thread is holding any tokens.
1051      * We could try to acquire the tokens but this case is so rare there
1052      * is no need to support it.
1053      */
1054     KKASSERT(gd->gd_spinlocks_wr == 0);
1055 
1056     if (TD_TOKS_HELD(ntd)) {
1057 	++preempt_miss;
1058 	return;
1059     }
1060     if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1061 	++preempt_weird;
1062 	return;
1063     }
1064     if (ntd->td_preempted) {
1065 	++preempt_hit;
1066 	return;
1067     }
1068     KKASSERT(gd->gd_processing_ipiq == 0);
1069 
1070     /*
1071      * Since we are able to preempt the current thread, there is no need to
1072      * call need_lwkt_resched().
1073      *
1074      * We must temporarily clear gd_intr_nesting_level around the switch
1075      * since switchouts from the target thread are allowed (they will just
1076      * return to our thread), and since the target thread has its own stack.
1077      *
1078      * A preemption must switch back to the original thread, assert the
1079      * case.
1080      */
1081     ++preempt_hit;
1082     ntd->td_preempted = td;
1083     td->td_flags |= TDF_PREEMPT_LOCK;
1084     KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1085     save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1086     gd->gd_intr_nesting_level = 0;
1087     ntd->td_flags |= TDF_RUNNING;
1088     xtd = td->td_switch(ntd);
1089     KKASSERT(xtd == ntd);
1090     lwkt_switch_return(xtd);
1091     gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1092 
1093     KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1094     ntd->td_preempted = NULL;
1095     td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1096 }
1097 
1098 /*
1099  * Conditionally call splz() if gd_reqflags indicates work is pending.
1100  * This will work inside a critical section but not inside a hard code
1101  * section.
1102  *
1103  * (self contained on a per cpu basis)
1104  */
1105 void
1106 splz_check(void)
1107 {
1108     globaldata_t gd = mycpu;
1109     thread_t td = gd->gd_curthread;
1110 
1111     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1112 	gd->gd_intr_nesting_level == 0 &&
1113 	td->td_nest_count < 2)
1114     {
1115 	splz();
1116     }
1117 }
1118 
1119 /*
1120  * This version is integrated into crit_exit, reqflags has already
1121  * been tested but td_critcount has not.
1122  *
1123  * We only want to execute the splz() on the 1->0 transition of
1124  * critcount and not in a hard code section or if too deeply nested.
1125  */
1126 void
1127 lwkt_maybe_splz(thread_t td)
1128 {
1129     globaldata_t gd = td->td_gd;
1130 
1131     if (td->td_critcount == 0 &&
1132 	gd->gd_intr_nesting_level == 0 &&
1133 	td->td_nest_count < 2)
1134     {
1135 	splz();
1136     }
1137 }
1138 
1139 /*
1140  * Drivers which set up processing co-threads can call this function to
1141  * run the co-thread at a higher priority and to allow it to preempt
1142  * normal threads.
1143  */
1144 void
1145 lwkt_set_interrupt_support_thread(void)
1146 {
1147 	thread_t td = curthread;
1148 
1149         lwkt_setpri_self(TDPRI_INT_SUPPORT);
1150 	td->td_flags |= TDF_INTTHREAD;
1151 	td->td_preemptable = lwkt_preempt;
1152 }
1153 
1154 
1155 /*
1156  * This function is used to negotiate a passive release of the current
1157  * process/lwp designation with the user scheduler, allowing the user
1158  * scheduler to schedule another user thread.  The related kernel thread
1159  * (curthread) continues running in the released state.
1160  */
1161 void
1162 lwkt_passive_release(struct thread *td)
1163 {
1164     struct lwp *lp = td->td_lwp;
1165 
1166     td->td_release = NULL;
1167     lwkt_setpri_self(TDPRI_KERN_USER);
1168     lp->lwp_proc->p_usched->release_curproc(lp);
1169 }
1170 
1171 
1172 /*
1173  * This implements a LWKT yield, allowing a kernel thread to yield to other
1174  * kernel threads at the same or higher priority.  This function can be
1175  * called in a tight loop and will typically only yield once per tick.
1176  *
1177  * Most kernel threads run at the same priority in order to allow equal
1178  * sharing.
1179  *
1180  * (self contained on a per cpu basis)
1181  */
1182 void
1183 lwkt_yield(void)
1184 {
1185     globaldata_t gd = mycpu;
1186     thread_t td = gd->gd_curthread;
1187 
1188     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1189 	splz();
1190     if (lwkt_resched_wanted()) {
1191 	lwkt_schedule_self(curthread);
1192 	lwkt_switch();
1193     }
1194 }
1195 
1196 /*
1197  * This yield is designed for kernel threads with a user context.
1198  *
1199  * The kernel acting on behalf of the user is potentially cpu-bound,
1200  * this function will efficiently allow other threads to run and also
1201  * switch to other processes by releasing.
1202  *
1203  * The lwkt_user_yield() function is designed to have very low overhead
1204  * if no yield is determined to be needed.
1205  */
1206 void
1207 lwkt_user_yield(void)
1208 {
1209     globaldata_t gd = mycpu;
1210     thread_t td = gd->gd_curthread;
1211 
1212     /*
1213      * Always run any pending interrupts in case we are in a critical
1214      * section.
1215      */
1216     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1217 	splz();
1218 
1219     /*
1220      * Switch (which forces a release) if another kernel thread needs
1221      * the cpu, if userland wants us to resched, or if our kernel
1222      * quantum has run out.
1223      */
1224     if (lwkt_resched_wanted() ||
1225 	user_resched_wanted())
1226     {
1227 	lwkt_switch();
1228     }
1229 
1230 #if 0
1231     /*
1232      * Reacquire the current process if we are released.
1233      *
1234      * XXX not implemented atm.  The kernel may be holding locks and such,
1235      *     so we want the thread to continue to receive cpu.
1236      */
1237     if (td->td_release == NULL && lp) {
1238 	lp->lwp_proc->p_usched->acquire_curproc(lp);
1239 	td->td_release = lwkt_passive_release;
1240 	lwkt_setpri_self(TDPRI_USER_NORM);
1241     }
1242 #endif
1243 }
1244 
1245 /*
1246  * Generic schedule.  Possibly schedule threads belonging to other cpus and
1247  * deal with threads that might be blocked on a wait queue.
1248  *
1249  * We have a little helper inline function which does additional work after
1250  * the thread has been enqueued, including dealing with preemption and
1251  * setting need_lwkt_resched() (which prevents the kernel from returning
1252  * to userland until it has processed higher priority threads).
1253  *
1254  * It is possible for this routine to be called after a failed _enqueue
1255  * (due to the target thread migrating, sleeping, or otherwise blocked).
1256  * We have to check that the thread is actually on the run queue!
1257  */
1258 static __inline
1259 void
1260 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1261 {
1262     if (ntd->td_flags & TDF_RUNQ) {
1263 	if (ntd->td_preemptable) {
1264 	    ntd->td_preemptable(ntd, ccount);	/* YYY +token */
1265 	}
1266     }
1267 }
1268 
1269 static __inline
1270 void
1271 _lwkt_schedule(thread_t td)
1272 {
1273     globaldata_t mygd = mycpu;
1274 
1275     KASSERT(td != &td->td_gd->gd_idlethread,
1276 	    ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1277     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1278     crit_enter_gd(mygd);
1279     KKASSERT(td->td_lwp == NULL ||
1280 	     (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1281 
1282     if (td == mygd->gd_curthread) {
1283 	_lwkt_enqueue(td);
1284     } else {
1285 	/*
1286 	 * If we own the thread, there is no race (since we are in a
1287 	 * critical section).  If we do not own the thread there might
1288 	 * be a race but the target cpu will deal with it.
1289 	 */
1290 #ifdef SMP
1291 	if (td->td_gd == mygd) {
1292 	    _lwkt_enqueue(td);
1293 	    _lwkt_schedule_post(mygd, td, 1);
1294 	} else {
1295 	    lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1296 	}
1297 #else
1298 	_lwkt_enqueue(td);
1299 	_lwkt_schedule_post(mygd, td, 1);
1300 #endif
1301     }
1302     crit_exit_gd(mygd);
1303 }
1304 
1305 void
1306 lwkt_schedule(thread_t td)
1307 {
1308     _lwkt_schedule(td);
1309 }
1310 
1311 void
1312 lwkt_schedule_noresched(thread_t td)	/* XXX not impl */
1313 {
1314     _lwkt_schedule(td);
1315 }
1316 
1317 #ifdef SMP
1318 
1319 /*
1320  * When scheduled remotely if frame != NULL the IPIQ is being
1321  * run via doreti or an interrupt then preemption can be allowed.
1322  *
1323  * To allow preemption we have to drop the critical section so only
1324  * one is present in _lwkt_schedule_post.
1325  */
1326 static void
1327 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1328 {
1329     thread_t td = curthread;
1330     thread_t ntd = arg;
1331 
1332     if (frame && ntd->td_preemptable) {
1333 	crit_exit_noyield(td);
1334 	_lwkt_schedule(ntd);
1335 	crit_enter_quick(td);
1336     } else {
1337 	_lwkt_schedule(ntd);
1338     }
1339 }
1340 
1341 /*
1342  * Thread migration using a 'Pull' method.  The thread may or may not be
1343  * the current thread.  It MUST be descheduled and in a stable state.
1344  * lwkt_giveaway() must be called on the cpu owning the thread.
1345  *
1346  * At any point after lwkt_giveaway() is called, the target cpu may
1347  * 'pull' the thread by calling lwkt_acquire().
1348  *
1349  * We have to make sure the thread is not sitting on a per-cpu tsleep
1350  * queue or it will blow up when it moves to another cpu.
1351  *
1352  * MPSAFE - must be called under very specific conditions.
1353  */
1354 void
1355 lwkt_giveaway(thread_t td)
1356 {
1357     globaldata_t gd = mycpu;
1358 
1359     crit_enter_gd(gd);
1360     if (td->td_flags & TDF_TSLEEPQ)
1361 	tsleep_remove(td);
1362     KKASSERT(td->td_gd == gd);
1363     TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1364     td->td_flags |= TDF_MIGRATING;
1365     crit_exit_gd(gd);
1366 }
1367 
1368 void
1369 lwkt_acquire(thread_t td)
1370 {
1371     globaldata_t gd;
1372     globaldata_t mygd;
1373     int retry = 10000000;
1374 
1375     KKASSERT(td->td_flags & TDF_MIGRATING);
1376     gd = td->td_gd;
1377     mygd = mycpu;
1378     if (gd != mycpu) {
1379 	cpu_lfence();
1380 	KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1381 	crit_enter_gd(mygd);
1382 	DEBUG_PUSH_INFO("lwkt_acquire");
1383 	while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1384 #ifdef SMP
1385 	    lwkt_process_ipiq();
1386 #endif
1387 	    cpu_lfence();
1388 	    if (--retry == 0) {
1389 		kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1390 			td, td->td_flags);
1391 		retry = 10000000;
1392 	    }
1393 	}
1394 	DEBUG_POP_INFO();
1395 	cpu_mfence();
1396 	td->td_gd = mygd;
1397 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1398 	td->td_flags &= ~TDF_MIGRATING;
1399 	crit_exit_gd(mygd);
1400     } else {
1401 	crit_enter_gd(mygd);
1402 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1403 	td->td_flags &= ~TDF_MIGRATING;
1404 	crit_exit_gd(mygd);
1405     }
1406 }
1407 
1408 #endif
1409 
1410 /*
1411  * Generic deschedule.  Descheduling threads other then your own should be
1412  * done only in carefully controlled circumstances.  Descheduling is
1413  * asynchronous.
1414  *
1415  * This function may block if the cpu has run out of messages.
1416  */
1417 void
1418 lwkt_deschedule(thread_t td)
1419 {
1420     crit_enter();
1421 #ifdef SMP
1422     if (td == curthread) {
1423 	_lwkt_dequeue(td);
1424     } else {
1425 	if (td->td_gd == mycpu) {
1426 	    _lwkt_dequeue(td);
1427 	} else {
1428 	    lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1429 	}
1430     }
1431 #else
1432     _lwkt_dequeue(td);
1433 #endif
1434     crit_exit();
1435 }
1436 
1437 /*
1438  * Set the target thread's priority.  This routine does not automatically
1439  * switch to a higher priority thread, LWKT threads are not designed for
1440  * continuous priority changes.  Yield if you want to switch.
1441  */
1442 void
1443 lwkt_setpri(thread_t td, int pri)
1444 {
1445     if (td->td_pri != pri) {
1446 	KKASSERT(pri >= 0);
1447 	crit_enter();
1448 	if (td->td_flags & TDF_RUNQ) {
1449 	    KKASSERT(td->td_gd == mycpu);
1450 	    _lwkt_dequeue(td);
1451 	    td->td_pri = pri;
1452 	    _lwkt_enqueue(td);
1453 	} else {
1454 	    td->td_pri = pri;
1455 	}
1456 	crit_exit();
1457     }
1458 }
1459 
1460 /*
1461  * Set the initial priority for a thread prior to it being scheduled for
1462  * the first time.  The thread MUST NOT be scheduled before or during
1463  * this call.  The thread may be assigned to a cpu other then the current
1464  * cpu.
1465  *
1466  * Typically used after a thread has been created with TDF_STOPPREQ,
1467  * and before the thread is initially scheduled.
1468  */
1469 void
1470 lwkt_setpri_initial(thread_t td, int pri)
1471 {
1472     KKASSERT(pri >= 0);
1473     KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1474     td->td_pri = pri;
1475 }
1476 
1477 void
1478 lwkt_setpri_self(int pri)
1479 {
1480     thread_t td = curthread;
1481 
1482     KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1483     crit_enter();
1484     if (td->td_flags & TDF_RUNQ) {
1485 	_lwkt_dequeue(td);
1486 	td->td_pri = pri;
1487 	_lwkt_enqueue(td);
1488     } else {
1489 	td->td_pri = pri;
1490     }
1491     crit_exit();
1492 }
1493 
1494 /*
1495  * hz tick scheduler clock for LWKT threads
1496  */
1497 void
1498 lwkt_schedulerclock(thread_t td)
1499 {
1500     globaldata_t gd = td->td_gd;
1501     thread_t xtd;
1502 
1503     if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1504 	/*
1505 	 * If the current thread is at the head of the runq shift it to the
1506 	 * end of any equal-priority threads and request a LWKT reschedule
1507 	 * if it moved.
1508 	 */
1509 	xtd = TAILQ_NEXT(td, td_threadq);
1510 	if (xtd && xtd->td_pri == td->td_pri) {
1511 	    TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1512 	    while (xtd && xtd->td_pri == td->td_pri)
1513 		xtd = TAILQ_NEXT(xtd, td_threadq);
1514 	    if (xtd)
1515 		TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1516 	    else
1517 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1518 	    need_lwkt_resched();
1519 	}
1520     } else {
1521 	/*
1522 	 * If we scheduled a thread other than the one at the head of the
1523 	 * queue always request a reschedule every tick.
1524 	 */
1525 	need_lwkt_resched();
1526     }
1527 }
1528 
1529 /*
1530  * Migrate the current thread to the specified cpu.
1531  *
1532  * This is accomplished by descheduling ourselves from the current cpu
1533  * and setting td_migrate_gd.  The lwkt_switch() code will detect that the
1534  * 'old' thread wants to migrate after it has been completely switched out
1535  * and will complete the migration.
1536  *
1537  * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1538  *
1539  * We must be sure to release our current process designation (if a user
1540  * process) before clearing out any tsleepq we are on because the release
1541  * code may re-add us.
1542  *
1543  * We must be sure to remove ourselves from the current cpu's tsleepq
1544  * before potentially moving to another queue.  The thread can be on
1545  * a tsleepq due to a left-over tsleep_interlock().
1546  */
1547 
1548 void
1549 lwkt_setcpu_self(globaldata_t rgd)
1550 {
1551 #ifdef SMP
1552     thread_t td = curthread;
1553 
1554     if (td->td_gd != rgd) {
1555 	crit_enter_quick(td);
1556 
1557 	if (td->td_release)
1558 	    td->td_release(td);
1559 	if (td->td_flags & TDF_TSLEEPQ)
1560 	    tsleep_remove(td);
1561 
1562 	/*
1563 	 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1564 	 * trying to deschedule ourselves and switch away, then deschedule
1565 	 * ourself, remove us from tdallq, and set td_migrate_gd.  Finally,
1566 	 * call lwkt_switch() to complete the operation.
1567 	 */
1568 	td->td_flags |= TDF_MIGRATING;
1569 	lwkt_deschedule_self(td);
1570 	TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1571 	td->td_migrate_gd = rgd;
1572 	lwkt_switch();
1573 
1574 	/*
1575 	 * We are now on the target cpu
1576 	 */
1577 	KKASSERT(rgd == mycpu);
1578 	TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1579 	crit_exit_quick(td);
1580     }
1581 #endif
1582 }
1583 
1584 void
1585 lwkt_migratecpu(int cpuid)
1586 {
1587 #ifdef SMP
1588 	globaldata_t rgd;
1589 
1590 	rgd = globaldata_find(cpuid);
1591 	lwkt_setcpu_self(rgd);
1592 #endif
1593 }
1594 
1595 #ifdef SMP
1596 /*
1597  * Remote IPI for cpu migration (called while in a critical section so we
1598  * do not have to enter another one).
1599  *
1600  * The thread (td) has already been completely descheduled from the
1601  * originating cpu and we can simply assert the case.  The thread is
1602  * assigned to the new cpu and enqueued.
1603  *
1604  * The thread will re-add itself to tdallq when it resumes execution.
1605  */
1606 static void
1607 lwkt_setcpu_remote(void *arg)
1608 {
1609     thread_t td = arg;
1610     globaldata_t gd = mycpu;
1611 
1612     KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1613     td->td_gd = gd;
1614     cpu_mfence();
1615     td->td_flags &= ~TDF_MIGRATING;
1616     KKASSERT(td->td_migrate_gd == NULL);
1617     KKASSERT(td->td_lwp == NULL ||
1618 	    (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1619     _lwkt_enqueue(td);
1620 }
1621 #endif
1622 
1623 struct lwp *
1624 lwkt_preempted_proc(void)
1625 {
1626     thread_t td = curthread;
1627     while (td->td_preempted)
1628 	td = td->td_preempted;
1629     return(td->td_lwp);
1630 }
1631 
1632 /*
1633  * Create a kernel process/thread/whatever.  It shares it's address space
1634  * with proc0 - ie: kernel only.
1635  *
1636  * If the cpu is not specified one will be selected.  In the future
1637  * specifying a cpu of -1 will enable kernel thread migration between
1638  * cpus.
1639  */
1640 int
1641 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1642 	    thread_t template, int tdflags, int cpu, const char *fmt, ...)
1643 {
1644     thread_t td;
1645     __va_list ap;
1646 
1647     td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1648 			   tdflags);
1649     if (tdp)
1650 	*tdp = td;
1651     cpu_set_thread_handler(td, lwkt_exit, func, arg);
1652 
1653     /*
1654      * Set up arg0 for 'ps' etc
1655      */
1656     __va_start(ap, fmt);
1657     kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1658     __va_end(ap);
1659 
1660     /*
1661      * Schedule the thread to run
1662      */
1663     if (td->td_flags & TDF_NOSTART)
1664 	td->td_flags &= ~TDF_NOSTART;
1665     else
1666 	lwkt_schedule(td);
1667     return 0;
1668 }
1669 
1670 /*
1671  * Destroy an LWKT thread.   Warning!  This function is not called when
1672  * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1673  * uses a different reaping mechanism.
1674  */
1675 void
1676 lwkt_exit(void)
1677 {
1678     thread_t td = curthread;
1679     thread_t std;
1680     globaldata_t gd;
1681 
1682     /*
1683      * Do any cleanup that might block here
1684      */
1685     if (td->td_flags & TDF_VERBOSE)
1686 	kprintf("kthread %p %s has exited\n", td, td->td_comm);
1687     caps_exit(td);
1688     biosched_done(td);
1689     dsched_exit_thread(td);
1690 
1691     /*
1692      * Get us into a critical section to interlock gd_freetd and loop
1693      * until we can get it freed.
1694      *
1695      * We have to cache the current td in gd_freetd because objcache_put()ing
1696      * it would rip it out from under us while our thread is still active.
1697      */
1698     gd = mycpu;
1699     crit_enter_quick(td);
1700     lwkt_wait_free(td);
1701     while ((std = gd->gd_freetd) != NULL) {
1702 	KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1703 	gd->gd_freetd = NULL;
1704 	objcache_put(thread_cache, std);
1705 	lwkt_wait_free(td);
1706     }
1707 
1708     /*
1709      * Remove thread resources from kernel lists and deschedule us for
1710      * the last time.  We cannot block after this point or we may end
1711      * up with a stale td on the tsleepq.
1712      *
1713      * None of this may block, the critical section is the only thing
1714      * protecting tdallq and the only thing preventing new lwkt_hold()
1715      * thread refs now.
1716      */
1717     if (td->td_flags & TDF_TSLEEPQ)
1718 	tsleep_remove(td);
1719     lwkt_deschedule_self(td);
1720     lwkt_remove_tdallq(td);
1721     KKASSERT(td->td_refs == 0);
1722 
1723     /*
1724      * Final cleanup
1725      */
1726     KKASSERT(gd->gd_freetd == NULL);
1727     if (td->td_flags & TDF_ALLOCATED_THREAD)
1728 	gd->gd_freetd = td;
1729     cpu_thread_exit();
1730 }
1731 
1732 void
1733 lwkt_remove_tdallq(thread_t td)
1734 {
1735     KKASSERT(td->td_gd == mycpu);
1736     TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1737 }
1738 
1739 /*
1740  * Code reduction and branch prediction improvements.  Call/return
1741  * overhead on modern cpus often degenerates into 0 cycles due to
1742  * the cpu's branch prediction hardware and return pc cache.  We
1743  * can take advantage of this by not inlining medium-complexity
1744  * functions and we can also reduce the branch prediction impact
1745  * by collapsing perfectly predictable branches into a single
1746  * procedure instead of duplicating it.
1747  *
1748  * Is any of this noticeable?  Probably not, so I'll take the
1749  * smaller code size.
1750  */
1751 void
1752 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1753 {
1754     _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1755 }
1756 
1757 void
1758 crit_panic(void)
1759 {
1760     thread_t td = curthread;
1761     int lcrit = td->td_critcount;
1762 
1763     td->td_critcount = 0;
1764     panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1765     /* NOT REACHED */
1766 }
1767 
1768 #ifdef SMP
1769 
1770 /*
1771  * Called from debugger/panic on cpus which have been stopped.  We must still
1772  * process the IPIQ while stopped, even if we were stopped while in a critical
1773  * section (XXX).
1774  *
1775  * If we are dumping also try to process any pending interrupts.  This may
1776  * or may not work depending on the state of the cpu at the point it was
1777  * stopped.
1778  */
1779 void
1780 lwkt_smp_stopped(void)
1781 {
1782     globaldata_t gd = mycpu;
1783 
1784     crit_enter_gd(gd);
1785     if (dumping) {
1786 	lwkt_process_ipiq();
1787 	splz();
1788     } else {
1789 	lwkt_process_ipiq();
1790     }
1791     crit_exit_gd(gd);
1792 }
1793 
1794 #endif
1795