Introduce sys/cpputil.h for our cpp metaprogramming needs
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
b12defdc 2 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
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
8c10bfcf
MD
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.
60f60350 20 *
8c10bfcf
MD
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
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
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.
8ad65e08
MD
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>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
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>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
85946b6c 96static __int64_t token_contention_count[TDPRI_MAX+1] __debugvar;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 102static void lwkt_setcpu_remote(void *arg);
88ebb169 103#endif
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
fb0f29c4
MD
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 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
0f7a3396 116#ifdef INVARIANTS
0c52fa62
SG
117SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 119#endif
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
38717797 128#ifdef INVARIANTS
85946b6c
MD
129SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_00, CTLFLAG_RW,
130 &token_contention_count[0], 0, "spinning due to token contention");
131SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_01, CTLFLAG_RW,
132 &token_contention_count[1], 0, "spinning due to token contention");
133SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_02, CTLFLAG_RW,
134 &token_contention_count[2], 0, "spinning due to token contention");
135SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_03, CTLFLAG_RW,
136 &token_contention_count[3], 0, "spinning due to token contention");
137SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_04, CTLFLAG_RW,
138 &token_contention_count[4], 0, "spinning due to token contention");
139SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_05, CTLFLAG_RW,
140 &token_contention_count[5], 0, "spinning due to token contention");
141SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_06, CTLFLAG_RW,
142 &token_contention_count[6], 0, "spinning due to token contention");
143SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_07, CTLFLAG_RW,
144 &token_contention_count[7], 0, "spinning due to token contention");
145SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_08, CTLFLAG_RW,
146 &token_contention_count[8], 0, "spinning due to token contention");
147SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_09, CTLFLAG_RW,
148 &token_contention_count[9], 0, "spinning due to token contention");
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_10, CTLFLAG_RW,
150 &token_contention_count[10], 0, "spinning due to token contention");
151SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_11, CTLFLAG_RW,
152 &token_contention_count[11], 0, "spinning due to token contention");
153SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_12, CTLFLAG_RW,
154 &token_contention_count[12], 0, "spinning due to token contention");
155SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_13, CTLFLAG_RW,
156 &token_contention_count[13], 0, "spinning due to token contention");
157SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_14, CTLFLAG_RW,
158 &token_contention_count[14], 0, "spinning due to token contention");
159SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_15, CTLFLAG_RW,
160 &token_contention_count[15], 0, "spinning due to token contention");
161SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_16, CTLFLAG_RW,
162 &token_contention_count[16], 0, "spinning due to token contention");
163SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_17, CTLFLAG_RW,
164 &token_contention_count[17], 0, "spinning due to token contention");
165SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_18, CTLFLAG_RW,
166 &token_contention_count[18], 0, "spinning due to token contention");
167SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_19, CTLFLAG_RW,
168 &token_contention_count[19], 0, "spinning due to token contention");
169SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_20, CTLFLAG_RW,
170 &token_contention_count[20], 0, "spinning due to token contention");
171SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_21, CTLFLAG_RW,
172 &token_contention_count[21], 0, "spinning due to token contention");
173SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_22, CTLFLAG_RW,
174 &token_contention_count[22], 0, "spinning due to token contention");
175SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_23, CTLFLAG_RW,
176 &token_contention_count[23], 0, "spinning due to token contention");
177SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_24, CTLFLAG_RW,
178 &token_contention_count[24], 0, "spinning due to token contention");
179SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_25, CTLFLAG_RW,
180 &token_contention_count[25], 0, "spinning due to token contention");
181SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_26, CTLFLAG_RW,
182 &token_contention_count[26], 0, "spinning due to token contention");
183SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_27, CTLFLAG_RW,
184 &token_contention_count[27], 0, "spinning due to token contention");
185SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_28, CTLFLAG_RW,
186 &token_contention_count[28], 0, "spinning due to token contention");
187SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_29, CTLFLAG_RW,
188 &token_contention_count[29], 0, "spinning due to token contention");
189SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_30, CTLFLAG_RW,
190 &token_contention_count[30], 0, "spinning due to token contention");
191SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count_31, CTLFLAG_RW,
192 &token_contention_count[31], 0, "spinning due to token contention");
38717797 193#endif
b12defdc 194static int fairq_enable = 0;
2a418930
MD
195SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
196 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 197static int fairq_bypass = -1;
b12defdc
MD
198SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
199 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
200extern int lwkt_sched_debug;
201int lwkt_sched_debug = 0;
202SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
203 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
204static int lwkt_spin_loops = 10;
205SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc
MD
206 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
207static int lwkt_spin_reseq = 0;
208SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
209 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
210static int lwkt_spin_monitor = 0;
211SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
212 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
d5b2d319
MD
213static int lwkt_spin_fatal = 0; /* disabled */
214SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
215 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 216static int preempt_enable = 1;
2a418930
MD
217SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
218 &preempt_enable, 0, "Enable preemption");
7b234d8c 219static int lwkt_cache_threads = 0;
765b1ae0
MD
220SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
221 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 222
2a418930
MD
223static __cachealign int lwkt_cseq_rindex;
224static __cachealign int lwkt_cseq_windex;
05220613 225
4b5f931b
MD
226/*
227 * These helper procedures handle the runq, they can only be called from
228 * within a critical section.
75cdbe6c
MD
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.
4b5f931b 234 */
f1d1c3fa
MD
235static __inline
236void
237_lwkt_dequeue(thread_t td)
238{
239 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 240 struct globaldata *gd = td->td_gd;
4b5f931b 241
f1d1c3fa 242 td->td_flags &= ~TDF_RUNQ;
f9235b6d 243 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 244 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 245 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
246 }
247}
248
f9235b6d
MD
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 */
f1d1c3fa
MD
255static __inline
256void
257_lwkt_enqueue(thread_t td)
258{
f9235b6d
MD
259 thread_t xtd;
260
7f5d7ed7 261 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 262 struct globaldata *gd = td->td_gd;
4b5f931b 263
f1d1c3fa 264 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
265 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
266 if (xtd == NULL) {
85946b6c
MD
267 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
268 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 269 } else {
85946b6c
MD
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);
f9235b6d 276 }
b12defdc
MD
277
278 /*
85946b6c 279 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 280 */
85946b6c
MD
281 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
282 need_lwkt_resched();
f1d1c3fa
MD
283 }
284}
8ad65e08 285
40aaf5fc
NT
286static __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;
4643740a 294 td->td_mpflags = 0;
40aaf5fc
NT
295 return (1);
296}
297
298static 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);
7b234d8c
MD
309 td->td_kstack = NULL;
310 td->td_flags = 0;
40aaf5fc
NT
311}
312
313/*
314 * Initialize the lwkt s/system.
765b1ae0 315 *
7b234d8c
MD
316 * Nominally cache up to 32 thread + kstack structures. Cache more on
317 * systems with a lot of cpu cores.
40aaf5fc
NT
318 */
319void
320lwkt_init(void)
321{
765b1ae0 322 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
323 if (lwkt_cache_threads == 0) {
324 lwkt_cache_threads = ncpus * 4;
325 if (lwkt_cache_threads < 32)
326 lwkt_cache_threads = 32;
327 }
765b1ae0
MD
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);
40aaf5fc
NT
332}
333
37af14fe
MD
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 */
341void
342lwkt_schedule_self(thread_t td)
343{
cfaeae2a 344 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 345 crit_enter_quick(td);
f9235b6d
MD
346 KASSERT(td != &td->td_gd->gd_idlethread,
347 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
348 KKASSERT(td->td_lwp == NULL ||
349 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 350 _lwkt_enqueue(td);
37af14fe
MD
351 crit_exit_quick(td);
352}
353
354/*
355 * Deschedule a thread.
356 *
357 * (non-blocking, self contained on a per cpu basis)
358 */
359void
360lwkt_deschedule_self(thread_t td)
361{
362 crit_enter_quick(td);
37af14fe
MD
363 _lwkt_dequeue(td);
364 crit_exit_quick(td);
365}
366
8ad65e08
MD
367/*
368 * LWKTs operate on a per-cpu basis
369 *
73e4f7b9 370 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
371 */
372void
373lwkt_gdinit(struct globaldata *gd)
374{
f9235b6d 375 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 376 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
377}
378
7d0bac62
MD
379/*
380 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
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.
0cfcada1
MD
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.
7d0bac62
MD
386 */
387thread_t
d3d32139 388lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 389{
d2d8515b 390 static int cpu_rotator;
c070746a 391 globaldata_t gd = mycpu;
99df837e 392 void *stack;
7d0bac62 393
c070746a
MD
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 */
ef0fdad1 399 if (td == NULL) {
cf709dd2
MD
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);
c070746a 404 gd->gd_freetd = NULL;
cf709dd2 405 } else {
c070746a 406 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
407 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
408 TDF_RUNQ)) == 0);
409 }
410 crit_exit_gd(gd);
40aaf5fc 411 KASSERT((td->td_flags &
2af9d75d
MD
412 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
413 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
414 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
415 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 416 }
c070746a
MD
417
418 /*
419 * Try to reuse cached stack.
420 */
f470d0c8
MD
421 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
422 if (flags & TDF_ALLOCATED_STACK) {
e4846942 423 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
424 stack = NULL;
425 }
426 }
427 if (stack == NULL) {
e40cfbd7 428 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 429 flags |= TDF_ALLOCATED_STACK;
99df837e 430 }
d2d8515b
MD
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));
99df837e 437 return(td);
7d0bac62
MD
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 *
f8c3996b
MD
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
75cdbe6c
MD
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.
7d0bac62 453 */
41a01a4d
MD
454#ifdef SMP
455
75cdbe6c
MD
456static void
457lwkt_init_thread_remote(void *arg)
458{
459 thread_t td = arg;
460
52eedfb5
MD
461 /*
462 * Protected by critical section held by IPI dispatch
463 */
75cdbe6c
MD
464 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
465}
466
41a01a4d
MD
467#endif
468
fdce8919
MD
469/*
470 * lwkt core thread structural initialization.
471 *
472 * NOTE: All threads are initialized as mpsafe threads.
473 */
7d0bac62 474void
f470d0c8
MD
475lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
476 struct globaldata *gd)
7d0bac62 477{
37af14fe
MD
478 globaldata_t mygd = mycpu;
479
99df837e
MD
480 bzero(td, sizeof(struct thread));
481 td->td_kstack = stack;
f470d0c8 482 td->td_kstack_size = stksize;
d3d32139 483 td->td_flags = flags;
4643740a 484 td->td_mpflags = 0;
26a0694b 485 td->td_gd = gd;
f9235b6d
MD
486 td->td_pri = TDPRI_KERN_DAEMON;
487 td->td_critcount = 1;
54341a3b 488 td->td_toks_have = NULL;
3b998fa9 489 td->td_toks_stop = &td->td_toks_base;
392cd266 490 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
fb0f29c4
MD
491 lwkt_initport_spin(&td->td_msgport);
492 else
493 lwkt_initport_thread(&td->td_msgport, td);
99df837e 494 pmap_init_thread(td);
0f7a3396 495#ifdef SMP
5d21b981
MD
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) {
37af14fe 503 crit_enter_gd(mygd);
75cdbe6c 504 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 505 crit_exit_gd(mygd);
75cdbe6c 506 } else {
2db3b277 507 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 508 }
0f7a3396 509#else
37af14fe 510 crit_enter_gd(mygd);
0f7a3396 511 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 512 crit_exit_gd(mygd);
0f7a3396 513#endif
8c72e3d5
AH
514
515 dsched_new_thread(td);
73e4f7b9
MD
516}
517
518void
519lwkt_set_comm(thread_t td, const char *ctl, ...)
520{
e2565a42 521 __va_list va;
73e4f7b9 522
e2565a42 523 __va_start(va, ctl);
379210cb 524 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 525 __va_end(va);
e7c0dbba 526 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
527}
528
eb2adbf5
MD
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 */
99df837e 534void
73e4f7b9 535lwkt_hold(thread_t td)
99df837e 536{
74c9628e 537 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
538}
539
540void
541lwkt_rele(thread_t td)
542{
543 KKASSERT(td->td_refs > 0);
74c9628e 544 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
545}
546
73e4f7b9
MD
547void
548lwkt_free_thread(thread_t td)
549{
74c9628e 550 KKASSERT(td->td_refs == 0);
c17a6852
MD
551 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
552 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
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;
99df837e 562 }
e7c0dbba 563 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
564}
565
566
8ad65e08
MD
567/*
568 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
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 *
96728c05
MD
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.
f1d1c3fa 581 *
69d78e99
MD
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).
cfaeae2a
MD
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
95858b91 593 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
594 */
595void
596lwkt_switch(void)
597{
37af14fe
MD
598 globaldata_t gd = mycpu;
599 thread_t td = gd->gd_curthread;
8ad65e08 600 thread_t ntd;
f9235b6d 601 thread_t xtd;
b12defdc 602 int spinning = 0;
8ad65e08 603
da0b0e8b 604 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 605 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 606
46a3f46d 607 /*
27e88a6e
MD
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.
46a3f46d 613 */
27e88a6e
MD
614 if (gd->gd_intr_nesting_level) {
615 int savegdnest;
616 int savegdtrap;
617
5fddbda2 618 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
619 panic("lwkt_switch: Attempt to switch from a "
620 "a fast interrupt, ipi, or hard code section, "
621 "td %p\n",
622 td);
27e88a6e
MD
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;
a7422615
MD
628 if ((td->td_flags & TDF_PANICWARN) == 0) {
629 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
630 kprintf("Warning: thread switch from interrupt, IPI, "
631 "or hard code section.\n"
a7422615 632 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 633 print_backtrace(-1);
a7422615 634 }
27e88a6e
MD
635 lwkt_switch();
636 gd->gd_intr_nesting_level = savegdnest;
637 gd->gd_trap_nesting_level = savegdtrap;
638 return;
639 }
96728c05 640 }
ef0fdad1 641
cb973d15 642 /*
85946b6c
MD
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.
cb973d15 657 */
85946b6c
MD
658 if (td->td_release &&
659 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 660 td->td_release(td);
85946b6c 661 }
cb973d15 662
85946b6c
MD
663 /*
664 * Release all tokens
665 */
37af14fe 666 crit_enter_gd(gd);
3b998fa9 667 if (TD_TOKS_HELD(td))
9d265729
MD
668 lwkt_relalltokens(td);
669
670 /*
b02926de
MD
671 * We had better not be holding any spin locks, but don't get into an
672 * endless panic loop.
9d265729 673 */
d666840a
MD
674 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
675 ("lwkt_switch: still holding %d exclusive spinlocks!",
676 gd->gd_spinlocks_wr));
9d265729 677
8a8d5d85
MD
678
679#ifdef SMP
0f7a3396
MD
680#ifdef INVARIANTS
681 if (td->td_cscount) {
6ea70f76 682 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
683 td);
684 if (panic_on_cscount)
685 panic("switching while mastering cpusync");
686 }
687#endif
8a8d5d85 688#endif
f9235b6d
MD
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 */
99df837e 700 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
701 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
702 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
703
704 /*
b9eb1c19
MD
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.
85946b6c
MD
708 *
709 * The interrupt may not have descheduled.
8ec60c3f 710 */
85946b6c 711 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 712 need_lwkt_resched();
f9235b6d
MD
713 goto havethread_preempted;
714 }
715
b12defdc 716 /*
f9235b6d 717 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
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.
f9235b6d
MD
722 */
723 for (;;) {
b12defdc 724 clear_lwkt_resched();
f9235b6d 725
4b5f931b 726 /*
2a418930 727 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 728 * it.
41a01a4d 729 */
2a418930
MD
730 for (;;) {
731 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
732
733 if (ntd == NULL) {
734 /*
b12defdc 735 * Runq is empty, switch to idle to allow it to halt.
2a418930
MD
736 */
737 ntd = &gd->gd_idlethread;
6f207a2c 738#ifdef SMP
2a418930 739 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 740 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 741#endif
2a418930
MD
742 cpu_time.cp_msg[0] = 0;
743 cpu_time.cp_stallpc = 0;
2a418930
MD
744 goto haveidle;
745 }
b12defdc 746 break;
f9235b6d 747 }
41a01a4d 748
8ec60c3f 749 /*
b12defdc 750 * Hotpath - schedule ntd.
6f207a2c
MD
751 *
752 * NOTE: For UP there is no mplock and lwkt_getalltokens()
753 * always succeeds.
8ec60c3f 754 */
b12defdc
MD
755 if (TD_TOKS_NOT_HELD(ntd) ||
756 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
757 {
f9235b6d 758 goto havethread;
b12defdc 759 }
f9235b6d 760
f9235b6d 761 /*
2a418930
MD
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
b12defdc 766 * in its stead.
2a418930 767 *
85946b6c
MD
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.
f9235b6d 771 */
b12defdc 772#ifdef INVARIANTS
85946b6c
MD
773 ++token_contention_count[ntd->td_pri];
774 ++ntd->td_contended;
b12defdc
MD
775#endif
776
85946b6c 777 if (fairq_bypass > 0)
b12defdc
MD
778 goto skip;
779
b12defdc
MD
780 xtd = NULL;
781 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
85946b6c
MD
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 */
b12defdc
MD
795 if (TD_TOKS_NOT_HELD(ntd) ||
796 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
797 goto havethread;
798 }
85946b6c
MD
799#ifdef INVARIANTS
800 ++token_contention_count[ntd->td_pri];
801 ++ntd->td_contended;
b12defdc 802#endif
2a418930
MD
803 }
804
b12defdc 805skip:
2a418930
MD
806 /*
807 * We exhausted the run list, meaning that all runnable threads
b12defdc 808 * are contested.
2a418930
MD
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 /*
b12defdc
MD
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.
2a418930 823 *
b12defdc
MD
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.
2a418930 827 */
b12defdc 828 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 829 need_lwkt_resched(); /* prevent hlt */
2a418930 830 goto haveidle;
4b5f931b 831 }
bd52bedf 832#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
833 if ((read_rflags() & PSL_I) == 0) {
834 cpu_enable_intr();
835 panic("lwkt_switch() called with interrupts disabled");
836 }
837#endif
b12defdc
MD
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 }
8b402283 897#endif
0f0466c0 898 }
b12defdc
MD
899 DELAY(1);
900 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 901 }
2a418930 902 /* highest level for(;;) loop */
f1d1c3fa 903 }
8a8d5d85 904
2a418930 905havethread:
b12defdc 906 /*
be71787b
MD
907 * Clear gd_idle_repeat when doing a normal switch to a non-idle
908 * thread.
f9235b6d 909 */
9ac1ee6e 910 ntd->td_wmesg = NULL;
b12defdc 911 ++gd->gd_cnt.v_swtch;
be71787b 912 gd->gd_idle_repeat = 0;
2a418930 913
f9235b6d 914havethread_preempted:
f9235b6d
MD
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.
8a8d5d85 919 */
2a418930 920 ;
f9235b6d
MD
921haveidle:
922 KASSERT(ntd->td_critcount,
b5d16701
MD
923 ("priority problem in lwkt_switch %d %d",
924 td->td_critcount, ntd->td_critcount));
925
94f6d86e 926 if (td != ntd) {
cc9b6223
MD
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 */
121f93bc 934 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 935 ++switch_count;
a1f0fb66 936 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
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 */
94f6d86e 940 }
b12defdc 941
b12defdc 942 /*
54341a3b 943 * catch-all. XXX is this strictly needed?
b12defdc
MD
944 */
945 splz_check();
54341a3b 946
37af14fe
MD
947 /* NOTE: current cpu may have changed after switch */
948 crit_exit_quick(td);
8ad65e08
MD
949}
950
cc9b6223
MD
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 */
955void
956lwkt_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
b68b7282 988/*
96728c05 989 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
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).
96728c05
MD
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.
f9235b6d 997 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
998 * to determine whether preemption is possible (aka usually just the crit
999 * priority of lwkt_schedule() itself).
b68b7282 1000 *
54341a3b
MD
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.
b68b7282
MD
1010 */
1011void
f9235b6d 1012lwkt_preempt(thread_t ntd, int critcount)
b68b7282 1013{
46a3f46d 1014 struct globaldata *gd = mycpu;
cc9b6223 1015 thread_t xtd;
0a3f9b47 1016 thread_t td;
2d910aaf 1017 int save_gd_intr_nesting_level;
b68b7282 1018
26a0694b 1019 /*
96728c05
MD
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
f9235b6d 1022 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1023 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1024 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1025 *
1026 * YYY The target thread must be in a critical section (else it must
1027 * inherit our critical section? I dunno yet).
26a0694b 1028 */
f9235b6d 1029 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1030
b12defdc 1031 td = gd->gd_curthread;
fbc024e4
MD
1032 if (preempt_enable == 0) {
1033 ++preempt_miss;
1034 return;
1035 }
f9235b6d 1036 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1037 ++preempt_miss;
1038 return;
1039 }
f9235b6d 1040 if (td->td_critcount > critcount) {
96728c05
MD
1041 ++preempt_miss;
1042 return;
1043 }
1044#ifdef SMP
121f93bc
MD
1045 if (td->td_cscount) {
1046 ++preempt_miss;
1047 return;
1048 }
46a3f46d 1049 if (ntd->td_gd != gd) {
96728c05
MD
1050 ++preempt_miss;
1051 return;
1052 }
1053#endif
41a01a4d 1054 /*
77912481
MD
1055 * We don't have to check spinlocks here as they will also bump
1056 * td_critcount.
d3d1cbc8
MD
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.
41a01a4d 1061 */
77912481
MD
1062 KKASSERT(gd->gd_spinlocks_wr == 0);
1063
3b998fa9 1064 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1065 ++preempt_miss;
d3d1cbc8
MD
1066 return;
1067 }
26a0694b
MD
1068 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1069 ++preempt_weird;
1070 return;
1071 }
1072 if (ntd->td_preempted) {
4b5f931b 1073 ++preempt_hit;
26a0694b 1074 return;
b68b7282 1075 }
da0b0e8b 1076 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1077
8ec60c3f
MD
1078 /*
1079 * Since we are able to preempt the current thread, there is no need to
1080 * call need_lwkt_resched().
2d910aaf
MD
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.
cc9b6223
MD
1085 *
1086 * A preemption must switch back to the original thread, assert the
1087 * case.
8ec60c3f 1088 */
26a0694b
MD
1089 ++preempt_hit;
1090 ntd->td_preempted = td;
1091 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1092 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1093 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1094 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1095
1096 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1097 ntd->td_flags |= TDF_RUNNING;
1098 xtd = td->td_switch(ntd);
1099 KKASSERT(xtd == ntd);
1100 lwkt_switch_return(xtd);
2d910aaf 1101 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1102
26a0694b
MD
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);
b68b7282
MD
1106}
1107
f1d1c3fa 1108/*
faaeffac 1109 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1110 * This will work inside a critical section but not inside a hard code
1111 * section.
ef0fdad1 1112 *
f1d1c3fa
MD
1113 * (self contained on a per cpu basis)
1114 */
1115void
faaeffac 1116splz_check(void)
f1d1c3fa 1117{
7966cb69
MD
1118 globaldata_t gd = mycpu;
1119 thread_t td = gd->gd_curthread;
ef0fdad1 1120
4a28fe22
MD
1121 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1122 gd->gd_intr_nesting_level == 0 &&
1123 td->td_nest_count < 2)
1124 {
f1d1c3fa 1125 splz();
4a28fe22
MD
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 */
1136void
1137lwkt_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 }
f1d1c3fa
MD
1147}
1148
e6546af9
MD
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 */
1154void
1155lwkt_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
8ad65e08 1165/*
f9235b6d
MD
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.
8ad65e08
MD
1170 */
1171void
f9235b6d 1172lwkt_passive_release(struct thread *td)
8ad65e08 1173{
f9235b6d
MD
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);
f1d1c3fa
MD
1179}
1180
f9235b6d 1181
3824f392 1182/*
d2d8515b
MD
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.
f9235b6d 1186 *
d2d8515b
MD
1187 * Most kernel threads run at the same priority in order to allow equal
1188 * sharing.
f9235b6d
MD
1189 *
1190 * (self contained on a per cpu basis)
3824f392
MD
1191 */
1192void
f9235b6d 1193lwkt_yield(void)
3824f392 1194{
f9235b6d
MD
1195 globaldata_t gd = mycpu;
1196 thread_t td = gd->gd_curthread;
3824f392 1197
f9235b6d
MD
1198 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1199 splz();
85946b6c 1200 if (lwkt_resched_wanted()) {
f9235b6d
MD
1201 lwkt_schedule_self(curthread);
1202 lwkt_switch();
f9235b6d 1203 }
3824f392
MD
1204}
1205
1206/*
f9235b6d
MD
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.
3824f392
MD
1212 *
1213 * The lwkt_user_yield() function is designed to have very low overhead
1214 * if no yield is determined to be needed.
1215 */
1216void
1217lwkt_user_yield(void)
1218{
f9235b6d
MD
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();
3824f392 1228
3824f392 1229 /*
f9235b6d
MD
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.
3824f392 1233 */
f9235b6d 1234 if (lwkt_resched_wanted() ||
85946b6c 1235 user_resched_wanted())
f9235b6d 1236 {
3824f392 1237 lwkt_switch();
3824f392
MD
1238 }
1239
f9235b6d 1240#if 0
3824f392 1241 /*
f9235b6d
MD
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.
3824f392 1246 */
f9235b6d
MD
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);
3824f392 1251 }
f9235b6d 1252#endif
b9eb1c19
MD
1253}
1254
8ad65e08 1255/*
f1d1c3fa
MD
1256 * Generic schedule. Possibly schedule threads belonging to other cpus and
1257 * deal with threads that might be blocked on a wait queue.
1258 *
0a3f9b47
MD
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).
6330a558
MD
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!
8ad65e08 1267 */
0a3f9b47
MD
1268static __inline
1269void
85946b6c 1270_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1271{
6330a558 1272 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1273 if (ntd->td_preemptable) {
f9235b6d 1274 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1275 }
0a3f9b47
MD
1276 }
1277}
1278
361d01dd 1279static __inline
8ad65e08 1280void
85946b6c 1281_lwkt_schedule(thread_t td)
8ad65e08 1282{
37af14fe
MD
1283 globaldata_t mygd = mycpu;
1284
cf709dd2
MD
1285 KASSERT(td != &td->td_gd->gd_idlethread,
1286 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1287 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1288 crit_enter_gd(mygd);
4643740a
MD
1289 KKASSERT(td->td_lwp == NULL ||
1290 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1291
37af14fe 1292 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1293 _lwkt_enqueue(td);
1294 } else {
f1d1c3fa 1295 /*
7cd8d145
MD
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.
f1d1c3fa 1299 */
0f7a3396 1300#ifdef SMP
7cd8d145 1301 if (td->td_gd == mygd) {
9d265729 1302 _lwkt_enqueue(td);
85946b6c 1303 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1304 } else {
e381e77c 1305 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1306 }
0f7a3396 1307#else
7cd8d145 1308 _lwkt_enqueue(td);
85946b6c 1309 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1310#endif
8ad65e08 1311 }
37af14fe 1312 crit_exit_gd(mygd);
8ad65e08
MD
1313}
1314
361d01dd
MD
1315void
1316lwkt_schedule(thread_t td)
1317{
85946b6c 1318 _lwkt_schedule(td);
361d01dd
MD
1319}
1320
1321void
85946b6c 1322lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1323{
85946b6c 1324 _lwkt_schedule(td);
361d01dd
MD
1325}
1326
88ebb169
SW
1327#ifdef SMP
1328
e381e77c
MD
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 */
1336static void
1337lwkt_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);
85946b6c 1344 _lwkt_schedule(ntd);
e381e77c
MD
1345 crit_enter_quick(td);
1346 } else {
85946b6c 1347 _lwkt_schedule(ntd);
e381e77c
MD
1348 }
1349}
1350
d9eea1a5 1351/*
52eedfb5
MD
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 *
ae8e83e6
MD
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 *
52eedfb5 1362 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1363 */
52eedfb5
MD
1364void
1365lwkt_giveaway(thread_t td)
1366{
3b4192fb 1367 globaldata_t gd = mycpu;
52eedfb5 1368
3b4192fb
MD
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);
52eedfb5
MD
1376}
1377
a2a5ad0d
MD
1378void
1379lwkt_acquire(thread_t td)
1380{
37af14fe
MD
1381 globaldata_t gd;
1382 globaldata_t mygd;
cc9b6223 1383 int retry = 10000000;
a2a5ad0d 1384
52eedfb5 1385 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1386 gd = td->td_gd;
37af14fe 1387 mygd = mycpu;
52eedfb5 1388 if (gd != mycpu) {
35238fa5 1389 cpu_lfence();
52eedfb5 1390 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1391 crit_enter_gd(mygd);
cfaeae2a 1392 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1393 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1394#ifdef SMP
1395 lwkt_process_ipiq();
1396#endif
52eedfb5 1397 cpu_lfence();
cc9b6223
MD
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 }
df910c23 1403 }
cfaeae2a 1404 DEBUG_POP_INFO();
562273ea 1405 cpu_mfence();
37af14fe 1406 td->td_gd = mygd;
52eedfb5
MD
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;
37af14fe 1414 crit_exit_gd(mygd);
a2a5ad0d
MD
1415 }
1416}
1417
52eedfb5
MD
1418#endif
1419
f1d1c3fa
MD
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.
8ad65e08
MD
1426 */
1427void
1428lwkt_deschedule(thread_t td)
1429{
f1d1c3fa 1430 crit_enter();
b8a98473 1431#ifdef SMP
f1d1c3fa
MD
1432 if (td == curthread) {
1433 _lwkt_dequeue(td);
1434 } else {
a72187e9 1435 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1436 _lwkt_dequeue(td);
1437 } else {
b8a98473 1438 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1439 }
1440 }
b8a98473
MD
1441#else
1442 _lwkt_dequeue(td);
1443#endif
f1d1c3fa
MD
1444 crit_exit();
1445}
1446
4b5f931b
MD
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.
4b5f931b
MD
1451 */
1452void
1453lwkt_setpri(thread_t td, int pri)
1454{
f9235b6d
MD
1455 if (td->td_pri != pri) {
1456 KKASSERT(pri >= 0);
1457 crit_enter();
1458 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1459 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1460 _lwkt_dequeue(td);
1461 td->td_pri = pri;
1462 _lwkt_enqueue(td);
1463 } else {
1464 td->td_pri = pri;
1465 }
1466 crit_exit();
26a0694b 1467 }
26a0694b
MD
1468}
1469
03bd0a5e
MD
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 */
1479void
1480lwkt_setpri_initial(thread_t td, int pri)
1481{
1482 KKASSERT(pri >= 0);
1483 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1484 td->td_pri = pri;
03bd0a5e
MD
1485}
1486
26a0694b
MD
1487void
1488lwkt_setpri_self(int pri)
1489{
1490 thread_t td = curthread;
1491
4b5f931b
MD
1492 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1493 crit_enter();
1494 if (td->td_flags & TDF_RUNQ) {
1495 _lwkt_dequeue(td);
f9235b6d 1496 td->td_pri = pri;
4b5f931b
MD
1497 _lwkt_enqueue(td);
1498 } else {
f9235b6d 1499 td->td_pri = pri;
4b5f931b
MD
1500 }
1501 crit_exit();
1502}
1503
f9235b6d 1504/*
85946b6c 1505 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1506 */
1507void
85946b6c 1508lwkt_schedulerclock(thread_t td)
f9235b6d 1509{
85946b6c
MD
1510 globaldata_t gd = td->td_gd;
1511 thread_t xtd;
2a418930 1512
85946b6c
MD
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();
f9235b6d 1529 }
85946b6c
MD
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();
f9235b6d
MD
1536 }
1537}
1538
5d21b981 1539/*
52eedfb5
MD
1540 * Migrate the current thread to the specified cpu.
1541 *
cc9b6223
MD
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.
ae8e83e6
MD
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().
5d21b981 1556 */
5d21b981
MD
1557
1558void
1559lwkt_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);
cc9b6223 1566
95858b91
MD
1567 if (td->td_release)
1568 td->td_release(td);
ae8e83e6 1569 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1570 tsleep_remove(td);
cc9b6223
MD
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 */
5d21b981
MD
1578 td->td_flags |= TDF_MIGRATING;
1579 lwkt_deschedule_self(td);
52eedfb5 1580 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1581 td->td_migrate_gd = rgd;
5d21b981 1582 lwkt_switch();
cc9b6223
MD
1583
1584 /*
1585 * We are now on the target cpu
1586 */
1587 KKASSERT(rgd == mycpu);
52eedfb5 1588 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1589 crit_exit_quick(td);
1590 }
1591#endif
1592}
1593
ecdefdda
MD
1594void
1595lwkt_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
cc9b6223 1605#ifdef SMP
5d21b981
MD
1606/*
1607 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
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.
5d21b981 1613 *
cc9b6223 1614 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1615 */
1616static void
1617lwkt_setcpu_remote(void *arg)
1618{
1619 thread_t td = arg;
1620 globaldata_t gd = mycpu;
1621
cc9b6223 1622 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1623 td->td_gd = gd;
562273ea 1624 cpu_mfence();
5d21b981 1625 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1626 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1627 KKASSERT(td->td_lwp == NULL ||
1628 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1629 _lwkt_enqueue(td);
1630}
3d28ff59 1631#endif
5d21b981 1632
553ea3c8 1633struct lwp *
4b5f931b
MD
1634lwkt_preempted_proc(void)
1635{
73e4f7b9 1636 thread_t td = curthread;
4b5f931b
MD
1637 while (td->td_preempted)
1638 td = td->td_preempted;
553ea3c8 1639 return(td->td_lwp);
4b5f931b
MD
1640}
1641
99df837e
MD
1642/*
1643 * Create a kernel process/thread/whatever. It shares it's address space
1644 * with proc0 - ie: kernel only.
1645 *
d2d8515b
MD
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.
99df837e
MD
1649 */
1650int
c9e9fb21
MD
1651lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1652 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1653{
73e4f7b9 1654 thread_t td;
e2565a42 1655 __va_list ap;
99df837e 1656
d3d32139 1657 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1658 tdflags);
a2a5ad0d
MD
1659 if (tdp)
1660 *tdp = td;
709799ea 1661 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1662
1663 /*
1664 * Set up arg0 for 'ps' etc
1665 */
e2565a42 1666 __va_start(ap, fmt);
379210cb 1667 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1668 __va_end(ap);
99df837e
MD
1669
1670 /*
1671 * Schedule the thread to run
1672 */
4643740a
MD
1673 if (td->td_flags & TDF_NOSTART)
1674 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1675 else
4643740a 1676 lwkt_schedule(td);
99df837e
MD
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 */
1685void
1686lwkt_exit(void)
1687{
1688 thread_t td = curthread;
c070746a 1689 thread_t std;
8826f33a 1690 globaldata_t gd;
99df837e 1691
2883d2d8
MD
1692 /*
1693 * Do any cleanup that might block here
1694 */
99df837e 1695 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1696 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1697 caps_exit(td);
2883d2d8
MD
1698 biosched_done(td);
1699 dsched_exit_thread(td);
c070746a
MD
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.
2af9d75d
MD
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.
c070746a
MD
1710 */
1711 gd = mycpu;
37af14fe 1712 crit_enter_quick(td);
2af9d75d
MD
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;
c070746a 1725 }
3b4192fb
MD
1726
1727 /*
1728 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1729 * the last time. We cannot block after this point or we may end
1730 * up with a stale td on the tsleepq.
eb2adbf5
MD
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.
3b4192fb
MD
1735 */
1736 if (td->td_flags & TDF_TSLEEPQ)
1737 tsleep_remove(td);
37af14fe 1738 lwkt_deschedule_self(td);
e56e4dea 1739 lwkt_remove_tdallq(td);
74c9628e 1740 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1741
1742 /*
1743 * Final cleanup
1744 */
1745 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1746 if (td->td_flags & TDF_ALLOCATED_THREAD)
1747 gd->gd_freetd = td;
99df837e
MD
1748 cpu_thread_exit();
1749}
1750
e56e4dea
MD
1751void
1752lwkt_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
9cf43f91
MD
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 */
1770void
b6468f56 1771crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1772{
b6468f56 1773 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1774}
1775
2d93b37a
MD
1776void
1777crit_panic(void)
1778{
1779 thread_t td = curthread;
850634cc 1780 int lcrit = td->td_critcount;
2d93b37a 1781
850634cc
AH
1782 td->td_critcount = 0;
1783 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1784 /* NOT REACHED */
2d93b37a
MD
1785}
1786
d165e668
MD
1787#ifdef SMP
1788
bd8015ca
MD
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 */
1798void
1799lwkt_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
d165e668 1813#endif