xref: /dflybsd-src/sys/kern/lwkt_thread.c (revision 28e81a2874cc70f7840d5527a55b62f8caed09b8)
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 = 0;
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 	td->td_kstack = NULL;
310 	td->td_flags = 0;
311 }
312 
313 /*
314  * Initialize the lwkt s/system.
315  *
316  * Nominally cache up to 32 thread + kstack structures.  Cache more on
317  * systems with a lot of cpu cores.
318  */
319 void
320 lwkt_init(void)
321 {
322     TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
323     if (lwkt_cache_threads == 0) {
324 	lwkt_cache_threads = ncpus * 4;
325 	if (lwkt_cache_threads < 32)
326 	    lwkt_cache_threads = 32;
327     }
328     thread_cache = objcache_create_mbacked(
329 				M_THREAD, sizeof(struct thread),
330 				NULL, lwkt_cache_threads,
331 				_lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
332 }
333 
334 /*
335  * Schedule a thread to run.  As the current thread we can always safely
336  * schedule ourselves, and a shortcut procedure is provided for that
337  * function.
338  *
339  * (non-blocking, self contained on a per cpu basis)
340  */
341 void
342 lwkt_schedule_self(thread_t td)
343 {
344     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
345     crit_enter_quick(td);
346     KASSERT(td != &td->td_gd->gd_idlethread,
347 	    ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
348     KKASSERT(td->td_lwp == NULL ||
349 	     (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
350     _lwkt_enqueue(td);
351     crit_exit_quick(td);
352 }
353 
354 /*
355  * Deschedule a thread.
356  *
357  * (non-blocking, self contained on a per cpu basis)
358  */
359 void
360 lwkt_deschedule_self(thread_t td)
361 {
362     crit_enter_quick(td);
363     _lwkt_dequeue(td);
364     crit_exit_quick(td);
365 }
366 
367 /*
368  * LWKTs operate on a per-cpu basis
369  *
370  * WARNING!  Called from early boot, 'mycpu' may not work yet.
371  */
372 void
373 lwkt_gdinit(struct globaldata *gd)
374 {
375     TAILQ_INIT(&gd->gd_tdrunq);
376     TAILQ_INIT(&gd->gd_tdallq);
377 }
378 
379 /*
380  * Create a new thread.  The thread must be associated with a process context
381  * or LWKT start address before it can be scheduled.  If the target cpu is
382  * -1 the thread will be created on the current cpu.
383  *
384  * If you intend to create a thread without a process context this function
385  * does everything except load the startup and switcher function.
386  */
387 thread_t
388 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
389 {
390     static int cpu_rotator;
391     globaldata_t gd = mycpu;
392     void *stack;
393 
394     /*
395      * If static thread storage is not supplied allocate a thread.  Reuse
396      * a cached free thread if possible.  gd_freetd is used to keep an exiting
397      * thread intact through the exit.
398      */
399     if (td == NULL) {
400 	crit_enter_gd(gd);
401 	if ((td = gd->gd_freetd) != NULL) {
402 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
403 				      TDF_RUNQ)) == 0);
404 	    gd->gd_freetd = NULL;
405 	} else {
406 	    td = objcache_get(thread_cache, M_WAITOK);
407 	    KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
408 				      TDF_RUNQ)) == 0);
409 	}
410 	crit_exit_gd(gd);
411     	KASSERT((td->td_flags &
412 		 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
413 		 TDF_ALLOCATED_THREAD,
414 		("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
415     	flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
416     }
417 
418     /*
419      * Try to reuse cached stack.
420      */
421     if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
422 	if (flags & TDF_ALLOCATED_STACK) {
423 	    kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
424 	    stack = NULL;
425 	}
426     }
427     if (stack == NULL) {
428 	stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
429 	flags |= TDF_ALLOCATED_STACK;
430     }
431     if (cpu < 0) {
432 	cpu = ++cpu_rotator;
433 	cpu_ccfence();
434 	cpu %= ncpus;
435     }
436     lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
437     return(td);
438 }
439 
440 /*
441  * Initialize a preexisting thread structure.  This function is used by
442  * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
443  *
444  * All threads start out in a critical section at a priority of
445  * TDPRI_KERN_DAEMON.  Higher level code will modify the priority as
446  * appropriate.  This function may send an IPI message when the
447  * requested cpu is not the current cpu and consequently gd_tdallq may
448  * not be initialized synchronously from the point of view of the originating
449  * cpu.
450  *
451  * NOTE! we have to be careful in regards to creating threads for other cpus
452  * if SMP has not yet been activated.
453  */
454 #ifdef SMP
455 
456 static void
457 lwkt_init_thread_remote(void *arg)
458 {
459     thread_t td = arg;
460 
461     /*
462      * Protected by critical section held by IPI dispatch
463      */
464     TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
465 }
466 
467 #endif
468 
469 /*
470  * lwkt core thread structural initialization.
471  *
472  * NOTE: All threads are initialized as mpsafe threads.
473  */
474 void
475 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
476 		struct globaldata *gd)
477 {
478     globaldata_t mygd = mycpu;
479 
480     bzero(td, sizeof(struct thread));
481     td->td_kstack = stack;
482     td->td_kstack_size = stksize;
483     td->td_flags = flags;
484     td->td_mpflags = 0;
485     td->td_gd = gd;
486     td->td_pri = TDPRI_KERN_DAEMON;
487     td->td_critcount = 1;
488     td->td_toks_have = NULL;
489     td->td_toks_stop = &td->td_toks_base;
490     if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
491 	lwkt_initport_spin(&td->td_msgport);
492     else
493 	lwkt_initport_thread(&td->td_msgport, td);
494     pmap_init_thread(td);
495 #ifdef SMP
496     /*
497      * Normally initializing a thread for a remote cpu requires sending an
498      * IPI.  However, the idlethread is setup before the other cpus are
499      * activated so we have to treat it as a special case.  XXX manipulation
500      * of gd_tdallq requires the BGL.
501      */
502     if (gd == mygd || td == &gd->gd_idlethread) {
503 	crit_enter_gd(mygd);
504 	TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
505 	crit_exit_gd(mygd);
506     } else {
507 	lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
508     }
509 #else
510     crit_enter_gd(mygd);
511     TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
512     crit_exit_gd(mygd);
513 #endif
514 
515     dsched_new_thread(td);
516 }
517 
518 void
519 lwkt_set_comm(thread_t td, const char *ctl, ...)
520 {
521     __va_list va;
522 
523     __va_start(va, ctl);
524     kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
525     __va_end(va);
526     KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
527 }
528 
529 /*
530  * Prevent the thread from getting destroyed.  Note that unlike PHOLD/PRELE
531  * this does not prevent the thread from migrating to another cpu so the
532  * gd_tdallq state is not protected by this.
533  */
534 void
535 lwkt_hold(thread_t td)
536 {
537     atomic_add_int(&td->td_refs, 1);
538 }
539 
540 void
541 lwkt_rele(thread_t td)
542 {
543     KKASSERT(td->td_refs > 0);
544     atomic_add_int(&td->td_refs, -1);
545 }
546 
547 void
548 lwkt_free_thread(thread_t td)
549 {
550     KKASSERT(td->td_refs == 0);
551     KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
552 			      TDF_RUNQ | TDF_TSLEEPQ)) == 0);
553     if (td->td_flags & TDF_ALLOCATED_THREAD) {
554     	objcache_put(thread_cache, td);
555     } else if (td->td_flags & TDF_ALLOCATED_STACK) {
556 	/* client-allocated struct with internally allocated stack */
557 	KASSERT(td->td_kstack && td->td_kstack_size > 0,
558 	    ("lwkt_free_thread: corrupted stack"));
559 	kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
560 	td->td_kstack = NULL;
561 	td->td_kstack_size = 0;
562     }
563     KTR_LOG(ctxsw_deadtd, td);
564 }
565 
566 
567 /*
568  * Switch to the next runnable lwkt.  If no LWKTs are runnable then
569  * switch to the idlethread.  Switching must occur within a critical
570  * section to avoid races with the scheduling queue.
571  *
572  * We always have full control over our cpu's run queue.  Other cpus
573  * that wish to manipulate our queue must use the cpu_*msg() calls to
574  * talk to our cpu, so a critical section is all that is needed and
575  * the result is very, very fast thread switching.
576  *
577  * The LWKT scheduler uses a fixed priority model and round-robins at
578  * each priority level.  User process scheduling is a totally
579  * different beast and LWKT priorities should not be confused with
580  * user process priorities.
581  *
582  * PREEMPTION NOTE: Preemption occurs via lwkt_preempt().  lwkt_switch()
583  * is not called by the current thread in the preemption case, only when
584  * the preempting thread blocks (in order to return to the original thread).
585  *
586  * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
587  * migration and tsleep deschedule the current lwkt thread and call
588  * lwkt_switch().  In particular, the target cpu of the migration fully
589  * expects the thread to become non-runnable and can deadlock against
590  * cpusync operations if we run any IPIs prior to switching the thread out.
591  *
592  * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
593  * THE CURRENT THREAD HAS BEEN DESCHEDULED!
594  */
595 void
596 lwkt_switch(void)
597 {
598     globaldata_t gd = mycpu;
599     thread_t td = gd->gd_curthread;
600     thread_t ntd;
601     thread_t xtd;
602     int spinning = 0;
603 
604     KKASSERT(gd->gd_processing_ipiq == 0);
605     KKASSERT(td->td_flags & TDF_RUNNING);
606 
607     /*
608      * Switching from within a 'fast' (non thread switched) interrupt or IPI
609      * is illegal.  However, we may have to do it anyway if we hit a fatal
610      * kernel trap or we have paniced.
611      *
612      * If this case occurs save and restore the interrupt nesting level.
613      */
614     if (gd->gd_intr_nesting_level) {
615 	int savegdnest;
616 	int savegdtrap;
617 
618 	if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
619 	    panic("lwkt_switch: Attempt to switch from a "
620 		  "a fast interrupt, ipi, or hard code section, "
621 		  "td %p\n",
622 		  td);
623 	} else {
624 	    savegdnest = gd->gd_intr_nesting_level;
625 	    savegdtrap = gd->gd_trap_nesting_level;
626 	    gd->gd_intr_nesting_level = 0;
627 	    gd->gd_trap_nesting_level = 0;
628 	    if ((td->td_flags & TDF_PANICWARN) == 0) {
629 		td->td_flags |= TDF_PANICWARN;
630 		kprintf("Warning: thread switch from interrupt, IPI, "
631 			"or hard code section.\n"
632 			"thread %p (%s)\n", td, td->td_comm);
633 		print_backtrace(-1);
634 	    }
635 	    lwkt_switch();
636 	    gd->gd_intr_nesting_level = savegdnest;
637 	    gd->gd_trap_nesting_level = savegdtrap;
638 	    return;
639 	}
640     }
641 
642     /*
643      * Release our current user process designation if we are blocking
644      * or if a user reschedule was requested.
645      *
646      * NOTE: This function is NOT called if we are switching into or
647      *	     returning from a preemption.
648      *
649      * NOTE: Releasing our current user process designation may cause
650      *	     it to be assigned to another thread, which in turn will
651      *	     cause us to block in the usched acquire code when we attempt
652      *	     to return to userland.
653      *
654      * NOTE: On SMP systems this can be very nasty when heavy token
655      *	     contention is present so we want to be careful not to
656      *	     release the designation gratuitously.
657      */
658     if (td->td_release &&
659 	(user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
660 	    td->td_release(td);
661     }
662 
663     /*
664      * Release all tokens
665      */
666     crit_enter_gd(gd);
667     if (TD_TOKS_HELD(td))
668 	    lwkt_relalltokens(td);
669 
670     /*
671      * We had better not be holding any spin locks, but don't get into an
672      * endless panic loop.
673      */
674     KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
675 	    ("lwkt_switch: still holding %d exclusive spinlocks!",
676 	     gd->gd_spinlocks_wr));
677 
678 
679 #ifdef SMP
680 #ifdef	INVARIANTS
681     if (td->td_cscount) {
682 	kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
683 		td);
684 	if (panic_on_cscount)
685 	    panic("switching while mastering cpusync");
686     }
687 #endif
688 #endif
689 
690     /*
691      * If we had preempted another thread on this cpu, resume the preempted
692      * thread.  This occurs transparently, whether the preempted thread
693      * was scheduled or not (it may have been preempted after descheduling
694      * itself).
695      *
696      * We have to setup the MP lock for the original thread after backing
697      * out the adjustment that was made to curthread when the original
698      * was preempted.
699      */
700     if ((ntd = td->td_preempted) != NULL) {
701 	KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
702 	ntd->td_flags |= TDF_PREEMPT_DONE;
703 
704 	/*
705 	 * The interrupt may have woken a thread up, we need to properly
706 	 * set the reschedule flag if the originally interrupted thread is
707 	 * at a lower priority.
708 	 *
709 	 * The interrupt may not have descheduled.
710 	 */
711 	if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
712 	    need_lwkt_resched();
713 	goto havethread_preempted;
714     }
715 
716     /*
717      * If we cannot obtain ownership of the tokens we cannot immediately
718      * schedule the target thread.
719      *
720      * Reminder: Again, we cannot afford to run any IPIs in this path if
721      * the current thread has been descheduled.
722      */
723     for (;;) {
724 	clear_lwkt_resched();
725 
726 	/*
727 	 * Hotpath - pull the head of the run queue and attempt to schedule
728 	 * it.
729 	 */
730 	for (;;) {
731 	    ntd = TAILQ_FIRST(&gd->gd_tdrunq);
732 
733 	    if (ntd == NULL) {
734 		/*
735 		 * Runq is empty, switch to idle to allow it to halt.
736 		 */
737 		ntd = &gd->gd_idlethread;
738 #ifdef SMP
739 		if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
740 		    ASSERT_NO_TOKENS_HELD(ntd);
741 #endif
742 		cpu_time.cp_msg[0] = 0;
743 		cpu_time.cp_stallpc = 0;
744 		goto haveidle;
745 	    }
746 	    break;
747 	}
748 
749 	/*
750 	 * Hotpath - schedule ntd.
751 	 *
752 	 * NOTE: For UP there is no mplock and lwkt_getalltokens()
753 	 *	     always succeeds.
754 	 */
755 	if (TD_TOKS_NOT_HELD(ntd) ||
756 	    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
757 	{
758 	    goto havethread;
759 	}
760 
761 	/*
762 	 * Coldpath (SMP only since tokens always succeed on UP)
763 	 *
764 	 * We had some contention on the thread we wanted to schedule.
765 	 * What we do now is try to find a thread that we can schedule
766 	 * in its stead.
767 	 *
768 	 * The coldpath scan does NOT rearrange threads in the run list.
769 	 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
770 	 * the next tick whenever the current head is not the current thread.
771 	 */
772 #ifdef	INVARIANTS
773 	++token_contention_count[ntd->td_pri];
774 	++ntd->td_contended;
775 #endif
776 
777 	if (fairq_bypass > 0)
778 		goto skip;
779 
780 	xtd = NULL;
781 	while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
782 		/*
783 		 * Never schedule threads returning to userland or the
784 		 * user thread scheduler helper thread when higher priority
785 		 * threads are present.
786 		 */
787 		if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
788 			ntd = NULL;
789 			break;
790 		}
791 
792 		/*
793 		 * Try this one.
794 		 */
795 		if (TD_TOKS_NOT_HELD(ntd) ||
796 		    lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
797 			goto havethread;
798 		}
799 #ifdef	INVARIANTS
800 		++token_contention_count[ntd->td_pri];
801 		++ntd->td_contended;
802 #endif
803 	}
804 
805 skip:
806 	/*
807 	 * We exhausted the run list, meaning that all runnable threads
808 	 * are contested.
809 	 */
810 	cpu_pause();
811 	ntd = &gd->gd_idlethread;
812 #ifdef SMP
813 	if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
814 	    ASSERT_NO_TOKENS_HELD(ntd);
815 	/* contention case, do not clear contention mask */
816 #endif
817 
818 	/*
819 	 * We are going to have to retry but if the current thread is not
820 	 * on the runq we instead switch through the idle thread to get away
821 	 * from the current thread.  We have to flag for lwkt reschedule
822 	 * to prevent the idle thread from halting.
823 	 *
824 	 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
825 	 *	 instruct it to deal with the potential for deadlocks by
826 	 *	 ordering the tokens by address.
827 	 */
828 	if ((td->td_flags & TDF_RUNQ) == 0) {
829 	    need_lwkt_resched();	/* prevent hlt */
830 	    goto haveidle;
831 	}
832 #if defined(INVARIANTS) && defined(__amd64__)
833 	if ((read_rflags() & PSL_I) == 0) {
834 		cpu_enable_intr();
835 		panic("lwkt_switch() called with interrupts disabled");
836 	}
837 #endif
838 
839 	/*
840 	 * Number iterations so far.  After a certain point we switch to
841 	 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
842 	 */
843 	if (spinning < 0x7FFFFFFF)
844 	    ++spinning;
845 
846 #ifdef SMP
847 	/*
848 	 * lwkt_getalltokens() failed in sorted token mode, we can use
849 	 * monitor/mwait in this case.
850 	 */
851 	if (spinning >= lwkt_spin_loops &&
852 	    (cpu_mi_feature & CPU_MI_MONITOR) &&
853 	    lwkt_spin_monitor)
854 	{
855 	    cpu_mmw_pause_int(&gd->gd_reqflags,
856 			      (gd->gd_reqflags | RQF_SPINNING) &
857 			      ~RQF_IDLECHECK_WK_MASK);
858 	}
859 #endif
860 
861 	/*
862 	 * We already checked that td is still scheduled so this should be
863 	 * safe.
864 	 */
865 	splz_check();
866 
867 	/*
868 	 * This experimental resequencer is used as a fall-back to reduce
869 	 * hw cache line contention by placing each core's scheduler into a
870 	 * time-domain-multplexed slot.
871 	 *
872 	 * The resequencer is disabled by default.  It's functionality has
873 	 * largely been superceeded by the token algorithm which limits races
874 	 * to a subset of cores.
875 	 *
876 	 * The resequencer algorithm tends to break down when more than
877 	 * 20 cores are contending.  What appears to happen is that new
878 	 * tokens can be obtained out of address-sorted order by new cores
879 	 * while existing cores languish in long delays between retries and
880 	 * wind up being starved-out of the token acquisition.
881 	 */
882 	if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
883 	    int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
884 	    int oseq;
885 
886 	    while ((oseq = lwkt_cseq_rindex) != cseq) {
887 		cpu_ccfence();
888 #if 1
889 		if (cpu_mi_feature & CPU_MI_MONITOR) {
890 		    cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
891 		} else {
892 #endif
893 		    cpu_pause();
894 		    cpu_lfence();
895 #if 1
896 		}
897 #endif
898 	    }
899 	    DELAY(1);
900 	    atomic_add_int(&lwkt_cseq_rindex, 1);
901 	}
902 	/* highest level for(;;) loop */
903     }
904 
905 havethread:
906     /*
907      * Clear gd_idle_repeat when doing a normal switch to a non-idle
908      * thread.
909      */
910     ntd->td_wmesg = NULL;
911     ++gd->gd_cnt.v_swtch;
912     gd->gd_idle_repeat = 0;
913 
914 havethread_preempted:
915     /*
916      * If the new target does not need the MP lock and we are holding it,
917      * release the MP lock.  If the new target requires the MP lock we have
918      * already acquired it for the target.
919      */
920     ;
921 haveidle:
922     KASSERT(ntd->td_critcount,
923 	    ("priority problem in lwkt_switch %d %d",
924 	    td->td_critcount, ntd->td_critcount));
925 
926     if (td != ntd) {
927 	/*
928 	 * Execute the actual thread switch operation.  This function
929 	 * returns to the current thread and returns the previous thread
930 	 * (which may be different from the thread we switched to).
931 	 *
932 	 * We are responsible for marking ntd as TDF_RUNNING.
933 	 */
934 	KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
935 	++switch_count;
936 	KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
937 	ntd->td_flags |= TDF_RUNNING;
938 	lwkt_switch_return(td->td_switch(ntd));
939 	/* ntd invalid, td_switch() can return a different thread_t */
940     }
941 
942     /*
943      * catch-all.  XXX is this strictly needed?
944      */
945     splz_check();
946 
947     /* NOTE: current cpu may have changed after switch */
948     crit_exit_quick(td);
949 }
950 
951 /*
952  * Called by assembly in the td_switch (thread restore path) for thread
953  * bootstrap cases which do not 'return' to lwkt_switch().
954  */
955 void
956 lwkt_switch_return(thread_t otd)
957 {
958 #ifdef SMP
959 	globaldata_t rgd;
960 
961 	/*
962 	 * Check if otd was migrating.  Now that we are on ntd we can finish
963 	 * up the migration.  This is a bit messy but it is the only place
964 	 * where td is known to be fully descheduled.
965 	 *
966 	 * We can only activate the migration if otd was migrating but not
967 	 * held on the cpu due to a preemption chain.  We still have to
968 	 * clear TDF_RUNNING on the old thread either way.
969 	 *
970 	 * We are responsible for clearing the previously running thread's
971 	 * TDF_RUNNING.
972 	 */
973 	if ((rgd = otd->td_migrate_gd) != NULL &&
974 	    (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
975 		KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
976 			 (TDF_MIGRATING | TDF_RUNNING));
977 		otd->td_migrate_gd = NULL;
978 		otd->td_flags &= ~TDF_RUNNING;
979 		lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
980 	} else {
981 		otd->td_flags &= ~TDF_RUNNING;
982 	}
983 #else
984 	otd->td_flags &= ~TDF_RUNNING;
985 #endif
986 }
987 
988 /*
989  * Request that the target thread preempt the current thread.  Preemption
990  * can only occur if our only critical section is the one that we were called
991  * with, the relative priority of the target thread is higher, and the target
992  * thread holds no tokens.  This also only works if we are not holding any
993  * spinlocks (obviously).
994  *
995  * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION.  Typically
996  * this is called via lwkt_schedule() through the td_preemptable callback.
997  * critcount is the managed critical priority that we should ignore in order
998  * to determine whether preemption is possible (aka usually just the crit
999  * priority of lwkt_schedule() itself).
1000  *
1001  * Preemption is typically limited to interrupt threads.
1002  *
1003  * Operation works in a fairly straight-forward manner.  The normal
1004  * scheduling code is bypassed and we switch directly to the target
1005  * thread.  When the target thread attempts to block or switch away
1006  * code at the base of lwkt_switch() will switch directly back to our
1007  * thread.  Our thread is able to retain whatever tokens it holds and
1008  * if the target needs one of them the target will switch back to us
1009  * and reschedule itself normally.
1010  */
1011 void
1012 lwkt_preempt(thread_t ntd, int critcount)
1013 {
1014     struct globaldata *gd = mycpu;
1015     thread_t xtd;
1016     thread_t td;
1017     int save_gd_intr_nesting_level;
1018 
1019     /*
1020      * The caller has put us in a critical section.  We can only preempt
1021      * if the caller of the caller was not in a critical section (basically
1022      * a local interrupt), as determined by the 'critcount' parameter.  We
1023      * also can't preempt if the caller is holding any spinlocks (even if
1024      * he isn't in a critical section).  This also handles the tokens test.
1025      *
1026      * YYY The target thread must be in a critical section (else it must
1027      * inherit our critical section?  I dunno yet).
1028      */
1029     KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1030 
1031     td = gd->gd_curthread;
1032     if (preempt_enable == 0) {
1033 	++preempt_miss;
1034 	return;
1035     }
1036     if (ntd->td_pri <= td->td_pri) {
1037 	++preempt_miss;
1038 	return;
1039     }
1040     if (td->td_critcount > critcount) {
1041 	++preempt_miss;
1042 	return;
1043     }
1044 #ifdef SMP
1045     if (td->td_cscount) {
1046 	++preempt_miss;
1047 	return;
1048     }
1049     if (ntd->td_gd != gd) {
1050 	++preempt_miss;
1051 	return;
1052     }
1053 #endif
1054     /*
1055      * We don't have to check spinlocks here as they will also bump
1056      * td_critcount.
1057      *
1058      * Do not try to preempt if the target thread is holding any tokens.
1059      * We could try to acquire the tokens but this case is so rare there
1060      * is no need to support it.
1061      */
1062     KKASSERT(gd->gd_spinlocks_wr == 0);
1063 
1064     if (TD_TOKS_HELD(ntd)) {
1065 	++preempt_miss;
1066 	return;
1067     }
1068     if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1069 	++preempt_weird;
1070 	return;
1071     }
1072     if (ntd->td_preempted) {
1073 	++preempt_hit;
1074 	return;
1075     }
1076     KKASSERT(gd->gd_processing_ipiq == 0);
1077 
1078     /*
1079      * Since we are able to preempt the current thread, there is no need to
1080      * call need_lwkt_resched().
1081      *
1082      * We must temporarily clear gd_intr_nesting_level around the switch
1083      * since switchouts from the target thread are allowed (they will just
1084      * return to our thread), and since the target thread has its own stack.
1085      *
1086      * A preemption must switch back to the original thread, assert the
1087      * case.
1088      */
1089     ++preempt_hit;
1090     ntd->td_preempted = td;
1091     td->td_flags |= TDF_PREEMPT_LOCK;
1092     KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1093     save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1094     gd->gd_intr_nesting_level = 0;
1095 
1096     KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1097     ntd->td_flags |= TDF_RUNNING;
1098     xtd = td->td_switch(ntd);
1099     KKASSERT(xtd == ntd);
1100     lwkt_switch_return(xtd);
1101     gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1102 
1103     KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1104     ntd->td_preempted = NULL;
1105     td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1106 }
1107 
1108 /*
1109  * Conditionally call splz() if gd_reqflags indicates work is pending.
1110  * This will work inside a critical section but not inside a hard code
1111  * section.
1112  *
1113  * (self contained on a per cpu basis)
1114  */
1115 void
1116 splz_check(void)
1117 {
1118     globaldata_t gd = mycpu;
1119     thread_t td = gd->gd_curthread;
1120 
1121     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1122 	gd->gd_intr_nesting_level == 0 &&
1123 	td->td_nest_count < 2)
1124     {
1125 	splz();
1126     }
1127 }
1128 
1129 /*
1130  * This version is integrated into crit_exit, reqflags has already
1131  * been tested but td_critcount has not.
1132  *
1133  * We only want to execute the splz() on the 1->0 transition of
1134  * critcount and not in a hard code section or if too deeply nested.
1135  */
1136 void
1137 lwkt_maybe_splz(thread_t td)
1138 {
1139     globaldata_t gd = td->td_gd;
1140 
1141     if (td->td_critcount == 0 &&
1142 	gd->gd_intr_nesting_level == 0 &&
1143 	td->td_nest_count < 2)
1144     {
1145 	splz();
1146     }
1147 }
1148 
1149 /*
1150  * Drivers which set up processing co-threads can call this function to
1151  * run the co-thread at a higher priority and to allow it to preempt
1152  * normal threads.
1153  */
1154 void
1155 lwkt_set_interrupt_support_thread(void)
1156 {
1157 	thread_t td = curthread;
1158 
1159         lwkt_setpri_self(TDPRI_INT_SUPPORT);
1160 	td->td_flags |= TDF_INTTHREAD;
1161 	td->td_preemptable = lwkt_preempt;
1162 }
1163 
1164 
1165 /*
1166  * This function is used to negotiate a passive release of the current
1167  * process/lwp designation with the user scheduler, allowing the user
1168  * scheduler to schedule another user thread.  The related kernel thread
1169  * (curthread) continues running in the released state.
1170  */
1171 void
1172 lwkt_passive_release(struct thread *td)
1173 {
1174     struct lwp *lp = td->td_lwp;
1175 
1176     td->td_release = NULL;
1177     lwkt_setpri_self(TDPRI_KERN_USER);
1178     lp->lwp_proc->p_usched->release_curproc(lp);
1179 }
1180 
1181 
1182 /*
1183  * This implements a LWKT yield, allowing a kernel thread to yield to other
1184  * kernel threads at the same or higher priority.  This function can be
1185  * called in a tight loop and will typically only yield once per tick.
1186  *
1187  * Most kernel threads run at the same priority in order to allow equal
1188  * sharing.
1189  *
1190  * (self contained on a per cpu basis)
1191  */
1192 void
1193 lwkt_yield(void)
1194 {
1195     globaldata_t gd = mycpu;
1196     thread_t td = gd->gd_curthread;
1197 
1198     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1199 	splz();
1200     if (lwkt_resched_wanted()) {
1201 	lwkt_schedule_self(curthread);
1202 	lwkt_switch();
1203     }
1204 }
1205 
1206 /*
1207  * This yield is designed for kernel threads with a user context.
1208  *
1209  * The kernel acting on behalf of the user is potentially cpu-bound,
1210  * this function will efficiently allow other threads to run and also
1211  * switch to other processes by releasing.
1212  *
1213  * The lwkt_user_yield() function is designed to have very low overhead
1214  * if no yield is determined to be needed.
1215  */
1216 void
1217 lwkt_user_yield(void)
1218 {
1219     globaldata_t gd = mycpu;
1220     thread_t td = gd->gd_curthread;
1221 
1222     /*
1223      * Always run any pending interrupts in case we are in a critical
1224      * section.
1225      */
1226     if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1227 	splz();
1228 
1229     /*
1230      * Switch (which forces a release) if another kernel thread needs
1231      * the cpu, if userland wants us to resched, or if our kernel
1232      * quantum has run out.
1233      */
1234     if (lwkt_resched_wanted() ||
1235 	user_resched_wanted())
1236     {
1237 	lwkt_switch();
1238     }
1239 
1240 #if 0
1241     /*
1242      * Reacquire the current process if we are released.
1243      *
1244      * XXX not implemented atm.  The kernel may be holding locks and such,
1245      *     so we want the thread to continue to receive cpu.
1246      */
1247     if (td->td_release == NULL && lp) {
1248 	lp->lwp_proc->p_usched->acquire_curproc(lp);
1249 	td->td_release = lwkt_passive_release;
1250 	lwkt_setpri_self(TDPRI_USER_NORM);
1251     }
1252 #endif
1253 }
1254 
1255 /*
1256  * Generic schedule.  Possibly schedule threads belonging to other cpus and
1257  * deal with threads that might be blocked on a wait queue.
1258  *
1259  * We have a little helper inline function which does additional work after
1260  * the thread has been enqueued, including dealing with preemption and
1261  * setting need_lwkt_resched() (which prevents the kernel from returning
1262  * to userland until it has processed higher priority threads).
1263  *
1264  * It is possible for this routine to be called after a failed _enqueue
1265  * (due to the target thread migrating, sleeping, or otherwise blocked).
1266  * We have to check that the thread is actually on the run queue!
1267  */
1268 static __inline
1269 void
1270 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1271 {
1272     if (ntd->td_flags & TDF_RUNQ) {
1273 	if (ntd->td_preemptable) {
1274 	    ntd->td_preemptable(ntd, ccount);	/* YYY +token */
1275 	}
1276     }
1277 }
1278 
1279 static __inline
1280 void
1281 _lwkt_schedule(thread_t td)
1282 {
1283     globaldata_t mygd = mycpu;
1284 
1285     KASSERT(td != &td->td_gd->gd_idlethread,
1286 	    ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1287     KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1288     crit_enter_gd(mygd);
1289     KKASSERT(td->td_lwp == NULL ||
1290 	     (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1291 
1292     if (td == mygd->gd_curthread) {
1293 	_lwkt_enqueue(td);
1294     } else {
1295 	/*
1296 	 * If we own the thread, there is no race (since we are in a
1297 	 * critical section).  If we do not own the thread there might
1298 	 * be a race but the target cpu will deal with it.
1299 	 */
1300 #ifdef SMP
1301 	if (td->td_gd == mygd) {
1302 	    _lwkt_enqueue(td);
1303 	    _lwkt_schedule_post(mygd, td, 1);
1304 	} else {
1305 	    lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1306 	}
1307 #else
1308 	_lwkt_enqueue(td);
1309 	_lwkt_schedule_post(mygd, td, 1);
1310 #endif
1311     }
1312     crit_exit_gd(mygd);
1313 }
1314 
1315 void
1316 lwkt_schedule(thread_t td)
1317 {
1318     _lwkt_schedule(td);
1319 }
1320 
1321 void
1322 lwkt_schedule_noresched(thread_t td)	/* XXX not impl */
1323 {
1324     _lwkt_schedule(td);
1325 }
1326 
1327 #ifdef SMP
1328 
1329 /*
1330  * When scheduled remotely if frame != NULL the IPIQ is being
1331  * run via doreti or an interrupt then preemption can be allowed.
1332  *
1333  * To allow preemption we have to drop the critical section so only
1334  * one is present in _lwkt_schedule_post.
1335  */
1336 static void
1337 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1338 {
1339     thread_t td = curthread;
1340     thread_t ntd = arg;
1341 
1342     if (frame && ntd->td_preemptable) {
1343 	crit_exit_noyield(td);
1344 	_lwkt_schedule(ntd);
1345 	crit_enter_quick(td);
1346     } else {
1347 	_lwkt_schedule(ntd);
1348     }
1349 }
1350 
1351 /*
1352  * Thread migration using a 'Pull' method.  The thread may or may not be
1353  * the current thread.  It MUST be descheduled and in a stable state.
1354  * lwkt_giveaway() must be called on the cpu owning the thread.
1355  *
1356  * At any point after lwkt_giveaway() is called, the target cpu may
1357  * 'pull' the thread by calling lwkt_acquire().
1358  *
1359  * We have to make sure the thread is not sitting on a per-cpu tsleep
1360  * queue or it will blow up when it moves to another cpu.
1361  *
1362  * MPSAFE - must be called under very specific conditions.
1363  */
1364 void
1365 lwkt_giveaway(thread_t td)
1366 {
1367     globaldata_t gd = mycpu;
1368 
1369     crit_enter_gd(gd);
1370     if (td->td_flags & TDF_TSLEEPQ)
1371 	tsleep_remove(td);
1372     KKASSERT(td->td_gd == gd);
1373     TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1374     td->td_flags |= TDF_MIGRATING;
1375     crit_exit_gd(gd);
1376 }
1377 
1378 void
1379 lwkt_acquire(thread_t td)
1380 {
1381     globaldata_t gd;
1382     globaldata_t mygd;
1383     int retry = 10000000;
1384 
1385     KKASSERT(td->td_flags & TDF_MIGRATING);
1386     gd = td->td_gd;
1387     mygd = mycpu;
1388     if (gd != mycpu) {
1389 	cpu_lfence();
1390 	KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1391 	crit_enter_gd(mygd);
1392 	DEBUG_PUSH_INFO("lwkt_acquire");
1393 	while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1394 #ifdef SMP
1395 	    lwkt_process_ipiq();
1396 #endif
1397 	    cpu_lfence();
1398 	    if (--retry == 0) {
1399 		kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1400 			td, td->td_flags);
1401 		retry = 10000000;
1402 	    }
1403 	}
1404 	DEBUG_POP_INFO();
1405 	cpu_mfence();
1406 	td->td_gd = mygd;
1407 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1408 	td->td_flags &= ~TDF_MIGRATING;
1409 	crit_exit_gd(mygd);
1410     } else {
1411 	crit_enter_gd(mygd);
1412 	TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1413 	td->td_flags &= ~TDF_MIGRATING;
1414 	crit_exit_gd(mygd);
1415     }
1416 }
1417 
1418 #endif
1419 
1420 /*
1421  * Generic deschedule.  Descheduling threads other then your own should be
1422  * done only in carefully controlled circumstances.  Descheduling is
1423  * asynchronous.
1424  *
1425  * This function may block if the cpu has run out of messages.
1426  */
1427 void
1428 lwkt_deschedule(thread_t td)
1429 {
1430     crit_enter();
1431 #ifdef SMP
1432     if (td == curthread) {
1433 	_lwkt_dequeue(td);
1434     } else {
1435 	if (td->td_gd == mycpu) {
1436 	    _lwkt_dequeue(td);
1437 	} else {
1438 	    lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1439 	}
1440     }
1441 #else
1442     _lwkt_dequeue(td);
1443 #endif
1444     crit_exit();
1445 }
1446 
1447 /*
1448  * Set the target thread's priority.  This routine does not automatically
1449  * switch to a higher priority thread, LWKT threads are not designed for
1450  * continuous priority changes.  Yield if you want to switch.
1451  */
1452 void
1453 lwkt_setpri(thread_t td, int pri)
1454 {
1455     if (td->td_pri != pri) {
1456 	KKASSERT(pri >= 0);
1457 	crit_enter();
1458 	if (td->td_flags & TDF_RUNQ) {
1459 	    KKASSERT(td->td_gd == mycpu);
1460 	    _lwkt_dequeue(td);
1461 	    td->td_pri = pri;
1462 	    _lwkt_enqueue(td);
1463 	} else {
1464 	    td->td_pri = pri;
1465 	}
1466 	crit_exit();
1467     }
1468 }
1469 
1470 /*
1471  * Set the initial priority for a thread prior to it being scheduled for
1472  * the first time.  The thread MUST NOT be scheduled before or during
1473  * this call.  The thread may be assigned to a cpu other then the current
1474  * cpu.
1475  *
1476  * Typically used after a thread has been created with TDF_STOPPREQ,
1477  * and before the thread is initially scheduled.
1478  */
1479 void
1480 lwkt_setpri_initial(thread_t td, int pri)
1481 {
1482     KKASSERT(pri >= 0);
1483     KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1484     td->td_pri = pri;
1485 }
1486 
1487 void
1488 lwkt_setpri_self(int pri)
1489 {
1490     thread_t td = curthread;
1491 
1492     KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1493     crit_enter();
1494     if (td->td_flags & TDF_RUNQ) {
1495 	_lwkt_dequeue(td);
1496 	td->td_pri = pri;
1497 	_lwkt_enqueue(td);
1498     } else {
1499 	td->td_pri = pri;
1500     }
1501     crit_exit();
1502 }
1503 
1504 /*
1505  * hz tick scheduler clock for LWKT threads
1506  */
1507 void
1508 lwkt_schedulerclock(thread_t td)
1509 {
1510     globaldata_t gd = td->td_gd;
1511     thread_t xtd;
1512 
1513     if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1514 	/*
1515 	 * If the current thread is at the head of the runq shift it to the
1516 	 * end of any equal-priority threads and request a LWKT reschedule
1517 	 * if it moved.
1518 	 */
1519 	xtd = TAILQ_NEXT(td, td_threadq);
1520 	if (xtd && xtd->td_pri == td->td_pri) {
1521 	    TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1522 	    while (xtd && xtd->td_pri == td->td_pri)
1523 		xtd = TAILQ_NEXT(xtd, td_threadq);
1524 	    if (xtd)
1525 		TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1526 	    else
1527 		TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1528 	    need_lwkt_resched();
1529 	}
1530     } else {
1531 	/*
1532 	 * If we scheduled a thread other than the one at the head of the
1533 	 * queue always request a reschedule every tick.
1534 	 */
1535 	need_lwkt_resched();
1536     }
1537 }
1538 
1539 /*
1540  * Migrate the current thread to the specified cpu.
1541  *
1542  * This is accomplished by descheduling ourselves from the current cpu
1543  * and setting td_migrate_gd.  The lwkt_switch() code will detect that the
1544  * 'old' thread wants to migrate after it has been completely switched out
1545  * and will complete the migration.
1546  *
1547  * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1548  *
1549  * We must be sure to release our current process designation (if a user
1550  * process) before clearing out any tsleepq we are on because the release
1551  * code may re-add us.
1552  *
1553  * We must be sure to remove ourselves from the current cpu's tsleepq
1554  * before potentially moving to another queue.  The thread can be on
1555  * a tsleepq due to a left-over tsleep_interlock().
1556  */
1557 
1558 void
1559 lwkt_setcpu_self(globaldata_t rgd)
1560 {
1561 #ifdef SMP
1562     thread_t td = curthread;
1563 
1564     if (td->td_gd != rgd) {
1565 	crit_enter_quick(td);
1566 
1567 	if (td->td_release)
1568 	    td->td_release(td);
1569 	if (td->td_flags & TDF_TSLEEPQ)
1570 	    tsleep_remove(td);
1571 
1572 	/*
1573 	 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1574 	 * trying to deschedule ourselves and switch away, then deschedule
1575 	 * ourself, remove us from tdallq, and set td_migrate_gd.  Finally,
1576 	 * call lwkt_switch() to complete the operation.
1577 	 */
1578 	td->td_flags |= TDF_MIGRATING;
1579 	lwkt_deschedule_self(td);
1580 	TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1581 	td->td_migrate_gd = rgd;
1582 	lwkt_switch();
1583 
1584 	/*
1585 	 * We are now on the target cpu
1586 	 */
1587 	KKASSERT(rgd == mycpu);
1588 	TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1589 	crit_exit_quick(td);
1590     }
1591 #endif
1592 }
1593 
1594 void
1595 lwkt_migratecpu(int cpuid)
1596 {
1597 #ifdef SMP
1598 	globaldata_t rgd;
1599 
1600 	rgd = globaldata_find(cpuid);
1601 	lwkt_setcpu_self(rgd);
1602 #endif
1603 }
1604 
1605 #ifdef SMP
1606 /*
1607  * Remote IPI for cpu migration (called while in a critical section so we
1608  * do not have to enter another one).
1609  *
1610  * The thread (td) has already been completely descheduled from the
1611  * originating cpu and we can simply assert the case.  The thread is
1612  * assigned to the new cpu and enqueued.
1613  *
1614  * The thread will re-add itself to tdallq when it resumes execution.
1615  */
1616 static void
1617 lwkt_setcpu_remote(void *arg)
1618 {
1619     thread_t td = arg;
1620     globaldata_t gd = mycpu;
1621 
1622     KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1623     td->td_gd = gd;
1624     cpu_mfence();
1625     td->td_flags &= ~TDF_MIGRATING;
1626     KKASSERT(td->td_migrate_gd == NULL);
1627     KKASSERT(td->td_lwp == NULL ||
1628 	    (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1629     _lwkt_enqueue(td);
1630 }
1631 #endif
1632 
1633 struct lwp *
1634 lwkt_preempted_proc(void)
1635 {
1636     thread_t td = curthread;
1637     while (td->td_preempted)
1638 	td = td->td_preempted;
1639     return(td->td_lwp);
1640 }
1641 
1642 /*
1643  * Create a kernel process/thread/whatever.  It shares it's address space
1644  * with proc0 - ie: kernel only.
1645  *
1646  * If the cpu is not specified one will be selected.  In the future
1647  * specifying a cpu of -1 will enable kernel thread migration between
1648  * cpus.
1649  */
1650 int
1651 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1652 	    thread_t template, int tdflags, int cpu, const char *fmt, ...)
1653 {
1654     thread_t td;
1655     __va_list ap;
1656 
1657     td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1658 			   tdflags);
1659     if (tdp)
1660 	*tdp = td;
1661     cpu_set_thread_handler(td, lwkt_exit, func, arg);
1662 
1663     /*
1664      * Set up arg0 for 'ps' etc
1665      */
1666     __va_start(ap, fmt);
1667     kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1668     __va_end(ap);
1669 
1670     /*
1671      * Schedule the thread to run
1672      */
1673     if (td->td_flags & TDF_NOSTART)
1674 	td->td_flags &= ~TDF_NOSTART;
1675     else
1676 	lwkt_schedule(td);
1677     return 0;
1678 }
1679 
1680 /*
1681  * Destroy an LWKT thread.   Warning!  This function is not called when
1682  * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1683  * uses a different reaping mechanism.
1684  */
1685 void
1686 lwkt_exit(void)
1687 {
1688     thread_t td = curthread;
1689     thread_t std;
1690     globaldata_t gd;
1691 
1692     /*
1693      * Do any cleanup that might block here
1694      */
1695     if (td->td_flags & TDF_VERBOSE)
1696 	kprintf("kthread %p %s has exited\n", td, td->td_comm);
1697     caps_exit(td);
1698     biosched_done(td);
1699     dsched_exit_thread(td);
1700 
1701     /*
1702      * Get us into a critical section to interlock gd_freetd and loop
1703      * until we can get it freed.
1704      *
1705      * We have to cache the current td in gd_freetd because objcache_put()ing
1706      * it would rip it out from under us while our thread is still active.
1707      *
1708      * We are the current thread so of course our own TDF_RUNNING bit will
1709      * be set, so unlike the lwp reap code we don't wait for it to clear.
1710      */
1711     gd = mycpu;
1712     crit_enter_quick(td);
1713     for (;;) {
1714 	if (td->td_refs) {
1715 	    tsleep(td, 0, "tdreap", 1);
1716 	    continue;
1717 	}
1718 	if ((std = gd->gd_freetd) != NULL) {
1719 	    KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1720 	    gd->gd_freetd = NULL;
1721 	    objcache_put(thread_cache, std);
1722 	    continue;
1723 	}
1724 	break;
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      * None of this may block, the critical section is the only thing
1733      * protecting tdallq and the only thing preventing new lwkt_hold()
1734      * thread refs now.
1735      */
1736     if (td->td_flags & TDF_TSLEEPQ)
1737 	tsleep_remove(td);
1738     lwkt_deschedule_self(td);
1739     lwkt_remove_tdallq(td);
1740     KKASSERT(td->td_refs == 0);
1741 
1742     /*
1743      * Final cleanup
1744      */
1745     KKASSERT(gd->gd_freetd == NULL);
1746     if (td->td_flags & TDF_ALLOCATED_THREAD)
1747 	gd->gd_freetd = td;
1748     cpu_thread_exit();
1749 }
1750 
1751 void
1752 lwkt_remove_tdallq(thread_t td)
1753 {
1754     KKASSERT(td->td_gd == mycpu);
1755     TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1756 }
1757 
1758 /*
1759  * Code reduction and branch prediction improvements.  Call/return
1760  * overhead on modern cpus often degenerates into 0 cycles due to
1761  * the cpu's branch prediction hardware and return pc cache.  We
1762  * can take advantage of this by not inlining medium-complexity
1763  * functions and we can also reduce the branch prediction impact
1764  * by collapsing perfectly predictable branches into a single
1765  * procedure instead of duplicating it.
1766  *
1767  * Is any of this noticeable?  Probably not, so I'll take the
1768  * smaller code size.
1769  */
1770 void
1771 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1772 {
1773     _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1774 }
1775 
1776 void
1777 crit_panic(void)
1778 {
1779     thread_t td = curthread;
1780     int lcrit = td->td_critcount;
1781 
1782     td->td_critcount = 0;
1783     panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1784     /* NOT REACHED */
1785 }
1786 
1787 #ifdef SMP
1788 
1789 /*
1790  * Called from debugger/panic on cpus which have been stopped.  We must still
1791  * process the IPIQ while stopped, even if we were stopped while in a critical
1792  * section (XXX).
1793  *
1794  * If we are dumping also try to process any pending interrupts.  This may
1795  * or may not work depending on the state of the cpu at the point it was
1796  * stopped.
1797  */
1798 void
1799 lwkt_smp_stopped(void)
1800 {
1801     globaldata_t gd = mycpu;
1802 
1803     crit_enter_gd(gd);
1804     if (dumping) {
1805 	lwkt_process_ipiq();
1806 	splz();
1807     } else {
1808 	lwkt_process_ipiq();
1809     }
1810     crit_exit_gd(gd);
1811 }
1812 
1813 #endif
1814