kernel - Fix DEBUG_LOCKS races
[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");
765b1ae0
MD
219static int lwkt_cache_threads = 32;
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;
294 return (1);
295}
296
297static void
298_lwkt_thread_dtor(void *obj, void *privdata)
299{
300 struct thread *td = (struct thread *)obj;
301
302 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
303 ("_lwkt_thread_dtor: not allocated from objcache"));
304 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
305 td->td_kstack_size > 0,
306 ("_lwkt_thread_dtor: corrupted stack"));
307 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
308}
309
310/*
311 * Initialize the lwkt s/system.
765b1ae0
MD
312 *
313 * Nominally cache up to 32 thread + kstack structures.
40aaf5fc
NT
314 */
315void
316lwkt_init(void)
317{
765b1ae0
MD
318 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
319 thread_cache = objcache_create_mbacked(
320 M_THREAD, sizeof(struct thread),
321 NULL, lwkt_cache_threads,
322 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
323}
324
37af14fe
MD
325/*
326 * Schedule a thread to run. As the current thread we can always safely
327 * schedule ourselves, and a shortcut procedure is provided for that
328 * function.
329 *
330 * (non-blocking, self contained on a per cpu basis)
331 */
332void
333lwkt_schedule_self(thread_t td)
334{
cfaeae2a 335 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 336 crit_enter_quick(td);
f9235b6d
MD
337 KASSERT(td != &td->td_gd->gd_idlethread,
338 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 339 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 340 _lwkt_enqueue(td);
37af14fe
MD
341 crit_exit_quick(td);
342}
343
344/*
345 * Deschedule a thread.
346 *
347 * (non-blocking, self contained on a per cpu basis)
348 */
349void
350lwkt_deschedule_self(thread_t td)
351{
352 crit_enter_quick(td);
37af14fe
MD
353 _lwkt_dequeue(td);
354 crit_exit_quick(td);
355}
356
8ad65e08
MD
357/*
358 * LWKTs operate on a per-cpu basis
359 *
73e4f7b9 360 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
361 */
362void
363lwkt_gdinit(struct globaldata *gd)
364{
f9235b6d 365 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 366 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
367}
368
7d0bac62
MD
369/*
370 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
371 * or LWKT start address before it can be scheduled. If the target cpu is
372 * -1 the thread will be created on the current cpu.
0cfcada1
MD
373 *
374 * If you intend to create a thread without a process context this function
375 * does everything except load the startup and switcher function.
7d0bac62
MD
376 */
377thread_t
d3d32139 378lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 379{
d2d8515b 380 static int cpu_rotator;
c070746a 381 globaldata_t gd = mycpu;
99df837e 382 void *stack;
7d0bac62 383
c070746a
MD
384 /*
385 * If static thread storage is not supplied allocate a thread. Reuse
386 * a cached free thread if possible. gd_freetd is used to keep an exiting
387 * thread intact through the exit.
388 */
ef0fdad1 389 if (td == NULL) {
cf709dd2
MD
390 crit_enter_gd(gd);
391 if ((td = gd->gd_freetd) != NULL) {
392 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
393 TDF_RUNQ)) == 0);
c070746a 394 gd->gd_freetd = NULL;
cf709dd2 395 } else {
c070746a 396 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
397 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
398 TDF_RUNQ)) == 0);
399 }
400 crit_exit_gd(gd);
40aaf5fc
NT
401 KASSERT((td->td_flags &
402 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
403 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
404 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 405 }
c070746a
MD
406
407 /*
408 * Try to reuse cached stack.
409 */
f470d0c8
MD
410 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
411 if (flags & TDF_ALLOCATED_STACK) {
e4846942 412 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
413 stack = NULL;
414 }
415 }
416 if (stack == NULL) {
e40cfbd7 417 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 418 flags |= TDF_ALLOCATED_STACK;
99df837e 419 }
d2d8515b
MD
420 if (cpu < 0) {
421 cpu = ++cpu_rotator;
422 cpu_ccfence();
423 cpu %= ncpus;
424 }
425 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 426 return(td);
7d0bac62
MD
427}
428
429/*
430 * Initialize a preexisting thread structure. This function is used by
431 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
432 *
f8c3996b
MD
433 * All threads start out in a critical section at a priority of
434 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
435 * appropriate. This function may send an IPI message when the
436 * requested cpu is not the current cpu and consequently gd_tdallq may
437 * not be initialized synchronously from the point of view of the originating
438 * cpu.
439 *
440 * NOTE! we have to be careful in regards to creating threads for other cpus
441 * if SMP has not yet been activated.
7d0bac62 442 */
41a01a4d
MD
443#ifdef SMP
444
75cdbe6c
MD
445static void
446lwkt_init_thread_remote(void *arg)
447{
448 thread_t td = arg;
449
52eedfb5
MD
450 /*
451 * Protected by critical section held by IPI dispatch
452 */
75cdbe6c
MD
453 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
454}
455
41a01a4d
MD
456#endif
457
fdce8919
MD
458/*
459 * lwkt core thread structural initialization.
460 *
461 * NOTE: All threads are initialized as mpsafe threads.
462 */
7d0bac62 463void
f470d0c8
MD
464lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
465 struct globaldata *gd)
7d0bac62 466{
37af14fe
MD
467 globaldata_t mygd = mycpu;
468
99df837e
MD
469 bzero(td, sizeof(struct thread));
470 td->td_kstack = stack;
f470d0c8 471 td->td_kstack_size = stksize;
d3d32139 472 td->td_flags = flags;
26a0694b 473 td->td_gd = gd;
f9235b6d
MD
474 td->td_pri = TDPRI_KERN_DAEMON;
475 td->td_critcount = 1;
54341a3b 476 td->td_toks_have = NULL;
3b998fa9 477 td->td_toks_stop = &td->td_toks_base;
392cd266 478 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
fb0f29c4
MD
479 lwkt_initport_spin(&td->td_msgport);
480 else
481 lwkt_initport_thread(&td->td_msgport, td);
99df837e 482 pmap_init_thread(td);
0f7a3396 483#ifdef SMP
5d21b981
MD
484 /*
485 * Normally initializing a thread for a remote cpu requires sending an
486 * IPI. However, the idlethread is setup before the other cpus are
487 * activated so we have to treat it as a special case. XXX manipulation
488 * of gd_tdallq requires the BGL.
489 */
490 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 491 crit_enter_gd(mygd);
75cdbe6c 492 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 493 crit_exit_gd(mygd);
75cdbe6c 494 } else {
2db3b277 495 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 496 }
0f7a3396 497#else
37af14fe 498 crit_enter_gd(mygd);
0f7a3396 499 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 500 crit_exit_gd(mygd);
0f7a3396 501#endif
8c72e3d5
AH
502
503 dsched_new_thread(td);
73e4f7b9
MD
504}
505
506void
507lwkt_set_comm(thread_t td, const char *ctl, ...)
508{
e2565a42 509 __va_list va;
73e4f7b9 510
e2565a42 511 __va_start(va, ctl);
379210cb 512 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 513 __va_end(va);
e7c0dbba 514 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
515}
516
99df837e 517void
73e4f7b9 518lwkt_hold(thread_t td)
99df837e 519{
74c9628e 520 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
521}
522
523void
524lwkt_rele(thread_t td)
525{
526 KKASSERT(td->td_refs > 0);
74c9628e 527 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
528}
529
530void
531lwkt_wait_free(thread_t td)
532{
533 while (td->td_refs)
377d4740 534 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
535}
536
537void
538lwkt_free_thread(thread_t td)
539{
74c9628e 540 KKASSERT(td->td_refs == 0);
c17a6852
MD
541 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
542 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
543 if (td->td_flags & TDF_ALLOCATED_THREAD) {
544 objcache_put(thread_cache, td);
545 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
546 /* client-allocated struct with internally allocated stack */
547 KASSERT(td->td_kstack && td->td_kstack_size > 0,
548 ("lwkt_free_thread: corrupted stack"));
549 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
550 td->td_kstack = NULL;
551 td->td_kstack_size = 0;
99df837e 552 }
e7c0dbba 553 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
554}
555
556
8ad65e08
MD
557/*
558 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
559 * switch to the idlethread. Switching must occur within a critical
560 * section to avoid races with the scheduling queue.
561 *
562 * We always have full control over our cpu's run queue. Other cpus
563 * that wish to manipulate our queue must use the cpu_*msg() calls to
564 * talk to our cpu, so a critical section is all that is needed and
565 * the result is very, very fast thread switching.
566 *
96728c05
MD
567 * The LWKT scheduler uses a fixed priority model and round-robins at
568 * each priority level. User process scheduling is a totally
569 * different beast and LWKT priorities should not be confused with
570 * user process priorities.
f1d1c3fa 571 *
69d78e99
MD
572 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
573 * is not called by the current thread in the preemption case, only when
574 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
575 *
576 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
577 * migration and tsleep deschedule the current lwkt thread and call
578 * lwkt_switch(). In particular, the target cpu of the migration fully
579 * expects the thread to become non-runnable and can deadlock against
580 * cpusync operations if we run any IPIs prior to switching the thread out.
581 *
582 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 583 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
584 */
585void
586lwkt_switch(void)
587{
37af14fe
MD
588 globaldata_t gd = mycpu;
589 thread_t td = gd->gd_curthread;
8ad65e08 590 thread_t ntd;
f9235b6d 591 thread_t xtd;
b12defdc 592 int spinning = 0;
8ad65e08 593
da0b0e8b
MD
594 KKASSERT(gd->gd_processing_ipiq == 0);
595
46a3f46d 596 /*
27e88a6e
MD
597 * Switching from within a 'fast' (non thread switched) interrupt or IPI
598 * is illegal. However, we may have to do it anyway if we hit a fatal
599 * kernel trap or we have paniced.
600 *
601 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 602 */
27e88a6e
MD
603 if (gd->gd_intr_nesting_level) {
604 int savegdnest;
605 int savegdtrap;
606
5fddbda2 607 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
608 panic("lwkt_switch: Attempt to switch from a "
609 "a fast interrupt, ipi, or hard code section, "
610 "td %p\n",
611 td);
27e88a6e
MD
612 } else {
613 savegdnest = gd->gd_intr_nesting_level;
614 savegdtrap = gd->gd_trap_nesting_level;
615 gd->gd_intr_nesting_level = 0;
616 gd->gd_trap_nesting_level = 0;
a7422615
MD
617 if ((td->td_flags & TDF_PANICWARN) == 0) {
618 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
619 kprintf("Warning: thread switch from interrupt, IPI, "
620 "or hard code section.\n"
a7422615 621 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 622 print_backtrace(-1);
a7422615 623 }
27e88a6e
MD
624 lwkt_switch();
625 gd->gd_intr_nesting_level = savegdnest;
626 gd->gd_trap_nesting_level = savegdtrap;
627 return;
628 }
96728c05 629 }
ef0fdad1 630
cb973d15 631 /*
85946b6c
MD
632 * Release our current user process designation if we are blocking
633 * or if a user reschedule was requested.
634 *
635 * NOTE: This function is NOT called if we are switching into or
636 * returning from a preemption.
637 *
638 * NOTE: Releasing our current user process designation may cause
639 * it to be assigned to another thread, which in turn will
640 * cause us to block in the usched acquire code when we attempt
641 * to return to userland.
642 *
643 * NOTE: On SMP systems this can be very nasty when heavy token
644 * contention is present so we want to be careful not to
645 * release the designation gratuitously.
cb973d15 646 */
85946b6c
MD
647 if (td->td_release &&
648 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 649 td->td_release(td);
85946b6c 650 }
cb973d15 651
85946b6c
MD
652 /*
653 * Release all tokens
654 */
37af14fe 655 crit_enter_gd(gd);
3b998fa9 656 if (TD_TOKS_HELD(td))
9d265729
MD
657 lwkt_relalltokens(td);
658
659 /*
b02926de
MD
660 * We had better not be holding any spin locks, but don't get into an
661 * endless panic loop.
9d265729 662 */
d666840a
MD
663 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
664 ("lwkt_switch: still holding %d exclusive spinlocks!",
665 gd->gd_spinlocks_wr));
9d265729 666
8a8d5d85
MD
667
668#ifdef SMP
0f7a3396
MD
669#ifdef INVARIANTS
670 if (td->td_cscount) {
6ea70f76 671 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
672 td);
673 if (panic_on_cscount)
674 panic("switching while mastering cpusync");
675 }
676#endif
8a8d5d85 677#endif
f9235b6d
MD
678
679 /*
680 * If we had preempted another thread on this cpu, resume the preempted
681 * thread. This occurs transparently, whether the preempted thread
682 * was scheduled or not (it may have been preempted after descheduling
683 * itself).
684 *
685 * We have to setup the MP lock for the original thread after backing
686 * out the adjustment that was made to curthread when the original
687 * was preempted.
688 */
99df837e 689 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
690 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
691 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
692
693 /*
b9eb1c19
MD
694 * The interrupt may have woken a thread up, we need to properly
695 * set the reschedule flag if the originally interrupted thread is
696 * at a lower priority.
85946b6c
MD
697 *
698 * The interrupt may not have descheduled.
8ec60c3f 699 */
85946b6c 700 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 701 need_lwkt_resched();
f9235b6d
MD
702 goto havethread_preempted;
703 }
704
b12defdc 705 /*
f9235b6d 706 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
707 * schedule the target thread.
708 *
709 * Reminder: Again, we cannot afford to run any IPIs in this path if
710 * the current thread has been descheduled.
f9235b6d
MD
711 */
712 for (;;) {
b12defdc 713 clear_lwkt_resched();
f9235b6d 714
4b5f931b 715 /*
2a418930 716 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 717 * it.
41a01a4d 718 */
2a418930
MD
719 for (;;) {
720 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
721
722 if (ntd == NULL) {
723 /*
b12defdc 724 * Runq is empty, switch to idle to allow it to halt.
2a418930
MD
725 */
726 ntd = &gd->gd_idlethread;
6f207a2c 727#ifdef SMP
2a418930 728 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 729 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 730#endif
2a418930
MD
731 cpu_time.cp_msg[0] = 0;
732 cpu_time.cp_stallpc = 0;
2a418930
MD
733 goto haveidle;
734 }
b12defdc 735 break;
f9235b6d 736 }
41a01a4d 737
8ec60c3f 738 /*
b12defdc 739 * Hotpath - schedule ntd.
6f207a2c
MD
740 *
741 * NOTE: For UP there is no mplock and lwkt_getalltokens()
742 * always succeeds.
8ec60c3f 743 */
b12defdc
MD
744 if (TD_TOKS_NOT_HELD(ntd) ||
745 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
746 {
f9235b6d 747 goto havethread;
b12defdc 748 }
f9235b6d 749
f9235b6d 750 /*
2a418930
MD
751 * Coldpath (SMP only since tokens always succeed on UP)
752 *
753 * We had some contention on the thread we wanted to schedule.
754 * What we do now is try to find a thread that we can schedule
b12defdc 755 * in its stead.
2a418930 756 *
85946b6c
MD
757 * The coldpath scan does NOT rearrange threads in the run list.
758 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
759 * the next tick whenever the current head is not the current thread.
f9235b6d 760 */
b12defdc 761#ifdef INVARIANTS
85946b6c
MD
762 ++token_contention_count[ntd->td_pri];
763 ++ntd->td_contended;
b12defdc
MD
764#endif
765
85946b6c 766 if (fairq_bypass > 0)
b12defdc
MD
767 goto skip;
768
b12defdc
MD
769 xtd = NULL;
770 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
85946b6c
MD
771 /*
772 * Never schedule threads returning to userland or the
773 * user thread scheduler helper thread when higher priority
774 * threads are present.
775 */
776 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
777 ntd = NULL;
778 break;
779 }
780
781 /*
782 * Try this one.
783 */
b12defdc
MD
784 if (TD_TOKS_NOT_HELD(ntd) ||
785 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
786 goto havethread;
787 }
85946b6c
MD
788#ifdef INVARIANTS
789 ++token_contention_count[ntd->td_pri];
790 ++ntd->td_contended;
b12defdc 791#endif
2a418930
MD
792 }
793
b12defdc 794skip:
2a418930
MD
795 /*
796 * We exhausted the run list, meaning that all runnable threads
b12defdc 797 * are contested.
2a418930
MD
798 */
799 cpu_pause();
800 ntd = &gd->gd_idlethread;
801#ifdef SMP
802 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
803 ASSERT_NO_TOKENS_HELD(ntd);
804 /* contention case, do not clear contention mask */
805#endif
806
807 /*
b12defdc
MD
808 * We are going to have to retry but if the current thread is not
809 * on the runq we instead switch through the idle thread to get away
810 * from the current thread. We have to flag for lwkt reschedule
811 * to prevent the idle thread from halting.
2a418930 812 *
b12defdc
MD
813 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
814 * instruct it to deal with the potential for deadlocks by
815 * ordering the tokens by address.
2a418930 816 */
b12defdc 817 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 818 need_lwkt_resched(); /* prevent hlt */
2a418930 819 goto haveidle;
4b5f931b 820 }
bd52bedf 821#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
822 if ((read_rflags() & PSL_I) == 0) {
823 cpu_enable_intr();
824 panic("lwkt_switch() called with interrupts disabled");
825 }
826#endif
b12defdc
MD
827
828 /*
829 * Number iterations so far. After a certain point we switch to
830 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
831 */
832 if (spinning < 0x7FFFFFFF)
833 ++spinning;
834
835#ifdef SMP
836 /*
837 * lwkt_getalltokens() failed in sorted token mode, we can use
838 * monitor/mwait in this case.
839 */
840 if (spinning >= lwkt_spin_loops &&
841 (cpu_mi_feature & CPU_MI_MONITOR) &&
842 lwkt_spin_monitor)
843 {
844 cpu_mmw_pause_int(&gd->gd_reqflags,
845 (gd->gd_reqflags | RQF_SPINNING) &
846 ~RQF_IDLECHECK_WK_MASK);
847 }
848#endif
849
850 /*
851 * We already checked that td is still scheduled so this should be
852 * safe.
853 */
854 splz_check();
855
856 /*
857 * This experimental resequencer is used as a fall-back to reduce
858 * hw cache line contention by placing each core's scheduler into a
859 * time-domain-multplexed slot.
860 *
861 * The resequencer is disabled by default. It's functionality has
862 * largely been superceeded by the token algorithm which limits races
863 * to a subset of cores.
864 *
865 * The resequencer algorithm tends to break down when more than
866 * 20 cores are contending. What appears to happen is that new
867 * tokens can be obtained out of address-sorted order by new cores
868 * while existing cores languish in long delays between retries and
869 * wind up being starved-out of the token acquisition.
870 */
871 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
872 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
873 int oseq;
874
875 while ((oseq = lwkt_cseq_rindex) != cseq) {
876 cpu_ccfence();
877#if 1
878 if (cpu_mi_feature & CPU_MI_MONITOR) {
879 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
880 } else {
881#endif
882 cpu_pause();
883 cpu_lfence();
884#if 1
885 }
8b402283 886#endif
0f0466c0 887 }
b12defdc
MD
888 DELAY(1);
889 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 890 }
2a418930 891 /* highest level for(;;) loop */
f1d1c3fa 892 }
8a8d5d85 893
2a418930 894havethread:
b12defdc 895 /*
be71787b
MD
896 * Clear gd_idle_repeat when doing a normal switch to a non-idle
897 * thread.
f9235b6d 898 */
9ac1ee6e 899 ntd->td_wmesg = NULL;
b12defdc 900 ++gd->gd_cnt.v_swtch;
be71787b 901 gd->gd_idle_repeat = 0;
2a418930 902
f9235b6d 903havethread_preempted:
f9235b6d
MD
904 /*
905 * If the new target does not need the MP lock and we are holding it,
906 * release the MP lock. If the new target requires the MP lock we have
907 * already acquired it for the target.
8a8d5d85 908 */
2a418930 909 ;
f9235b6d
MD
910haveidle:
911 KASSERT(ntd->td_critcount,
b5d16701
MD
912 ("priority problem in lwkt_switch %d %d",
913 td->td_critcount, ntd->td_critcount));
914
94f6d86e 915 if (td != ntd) {
cc9b6223
MD
916 /*
917 * Execute the actual thread switch operation. This function
918 * returns to the current thread and returns the previous thread
919 * (which may be different from the thread we switched to).
920 *
921 * We are responsible for marking ntd as TDF_RUNNING.
922 */
94f6d86e 923 ++switch_count;
a1f0fb66 924 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
925 ntd->td_flags |= TDF_RUNNING;
926 lwkt_switch_return(td->td_switch(ntd));
927 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 928 }
b12defdc 929
b12defdc 930 /*
54341a3b 931 * catch-all. XXX is this strictly needed?
b12defdc
MD
932 */
933 splz_check();
54341a3b 934
37af14fe
MD
935 /* NOTE: current cpu may have changed after switch */
936 crit_exit_quick(td);
8ad65e08
MD
937}
938
cc9b6223
MD
939/*
940 * Called by assembly in the td_switch (thread restore path) for thread
941 * bootstrap cases which do not 'return' to lwkt_switch().
942 */
943void
944lwkt_switch_return(thread_t otd)
945{
946#ifdef SMP
947 globaldata_t rgd;
948
949 /*
950 * Check if otd was migrating. Now that we are on ntd we can finish
951 * up the migration. This is a bit messy but it is the only place
952 * where td is known to be fully descheduled.
953 *
954 * We can only activate the migration if otd was migrating but not
955 * held on the cpu due to a preemption chain. We still have to
956 * clear TDF_RUNNING on the old thread either way.
957 *
958 * We are responsible for clearing the previously running thread's
959 * TDF_RUNNING.
960 */
961 if ((rgd = otd->td_migrate_gd) != NULL &&
962 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
963 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
964 (TDF_MIGRATING | TDF_RUNNING));
965 otd->td_migrate_gd = NULL;
966 otd->td_flags &= ~TDF_RUNNING;
967 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
968 } else {
969 otd->td_flags &= ~TDF_RUNNING;
970 }
971#else
972 otd->td_flags &= ~TDF_RUNNING;
973#endif
974}
975
b68b7282 976/*
96728c05 977 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
978 * can only occur if our only critical section is the one that we were called
979 * with, the relative priority of the target thread is higher, and the target
980 * thread holds no tokens. This also only works if we are not holding any
981 * spinlocks (obviously).
96728c05
MD
982 *
983 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
984 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 985 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
986 * to determine whether preemption is possible (aka usually just the crit
987 * priority of lwkt_schedule() itself).
b68b7282 988 *
54341a3b
MD
989 * Preemption is typically limited to interrupt threads.
990 *
991 * Operation works in a fairly straight-forward manner. The normal
992 * scheduling code is bypassed and we switch directly to the target
993 * thread. When the target thread attempts to block or switch away
994 * code at the base of lwkt_switch() will switch directly back to our
995 * thread. Our thread is able to retain whatever tokens it holds and
996 * if the target needs one of them the target will switch back to us
997 * and reschedule itself normally.
b68b7282
MD
998 */
999void
f9235b6d 1000lwkt_preempt(thread_t ntd, int critcount)
b68b7282 1001{
46a3f46d 1002 struct globaldata *gd = mycpu;
cc9b6223 1003 thread_t xtd;
0a3f9b47 1004 thread_t td;
2d910aaf 1005 int save_gd_intr_nesting_level;
b68b7282 1006
26a0694b 1007 /*
96728c05
MD
1008 * The caller has put us in a critical section. We can only preempt
1009 * if the caller of the caller was not in a critical section (basically
f9235b6d 1010 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1011 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1012 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1013 *
1014 * YYY The target thread must be in a critical section (else it must
1015 * inherit our critical section? I dunno yet).
26a0694b 1016 */
f9235b6d 1017 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1018
b12defdc 1019 td = gd->gd_curthread;
fbc024e4
MD
1020 if (preempt_enable == 0) {
1021 ++preempt_miss;
1022 return;
1023 }
f9235b6d 1024 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1025 ++preempt_miss;
1026 return;
1027 }
f9235b6d 1028 if (td->td_critcount > critcount) {
96728c05
MD
1029 ++preempt_miss;
1030 return;
1031 }
1032#ifdef SMP
46a3f46d 1033 if (ntd->td_gd != gd) {
96728c05
MD
1034 ++preempt_miss;
1035 return;
1036 }
1037#endif
41a01a4d 1038 /*
77912481
MD
1039 * We don't have to check spinlocks here as they will also bump
1040 * td_critcount.
d3d1cbc8
MD
1041 *
1042 * Do not try to preempt if the target thread is holding any tokens.
1043 * We could try to acquire the tokens but this case is so rare there
1044 * is no need to support it.
41a01a4d 1045 */
77912481
MD
1046 KKASSERT(gd->gd_spinlocks_wr == 0);
1047
3b998fa9 1048 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1049 ++preempt_miss;
d3d1cbc8
MD
1050 return;
1051 }
26a0694b
MD
1052 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1053 ++preempt_weird;
1054 return;
1055 }
1056 if (ntd->td_preempted) {
4b5f931b 1057 ++preempt_hit;
26a0694b 1058 return;
b68b7282 1059 }
da0b0e8b 1060 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1061
8ec60c3f
MD
1062 /*
1063 * Since we are able to preempt the current thread, there is no need to
1064 * call need_lwkt_resched().
2d910aaf
MD
1065 *
1066 * We must temporarily clear gd_intr_nesting_level around the switch
1067 * since switchouts from the target thread are allowed (they will just
1068 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1069 *
1070 * A preemption must switch back to the original thread, assert the
1071 * case.
8ec60c3f 1072 */
26a0694b
MD
1073 ++preempt_hit;
1074 ntd->td_preempted = td;
1075 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1076 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1077 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1078 gd->gd_intr_nesting_level = 0;
cc9b6223
MD
1079 ntd->td_flags |= TDF_RUNNING;
1080 xtd = td->td_switch(ntd);
1081 KKASSERT(xtd == ntd);
1082 lwkt_switch_return(xtd);
2d910aaf 1083 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1084
26a0694b
MD
1085 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1086 ntd->td_preempted = NULL;
1087 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1088}
1089
f1d1c3fa 1090/*
faaeffac 1091 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1092 * This will work inside a critical section but not inside a hard code
1093 * section.
ef0fdad1 1094 *
f1d1c3fa
MD
1095 * (self contained on a per cpu basis)
1096 */
1097void
faaeffac 1098splz_check(void)
f1d1c3fa 1099{
7966cb69
MD
1100 globaldata_t gd = mycpu;
1101 thread_t td = gd->gd_curthread;
ef0fdad1 1102
4a28fe22
MD
1103 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1104 gd->gd_intr_nesting_level == 0 &&
1105 td->td_nest_count < 2)
1106 {
f1d1c3fa 1107 splz();
4a28fe22
MD
1108 }
1109}
1110
1111/*
1112 * This version is integrated into crit_exit, reqflags has already
1113 * been tested but td_critcount has not.
1114 *
1115 * We only want to execute the splz() on the 1->0 transition of
1116 * critcount and not in a hard code section or if too deeply nested.
1117 */
1118void
1119lwkt_maybe_splz(thread_t td)
1120{
1121 globaldata_t gd = td->td_gd;
1122
1123 if (td->td_critcount == 0 &&
1124 gd->gd_intr_nesting_level == 0 &&
1125 td->td_nest_count < 2)
1126 {
1127 splz();
1128 }
f1d1c3fa
MD
1129}
1130
e6546af9
MD
1131/*
1132 * Drivers which set up processing co-threads can call this function to
1133 * run the co-thread at a higher priority and to allow it to preempt
1134 * normal threads.
1135 */
1136void
1137lwkt_set_interrupt_support_thread(void)
1138{
1139 thread_t td = curthread;
1140
1141 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1142 td->td_flags |= TDF_INTTHREAD;
1143 td->td_preemptable = lwkt_preempt;
1144}
1145
1146
8ad65e08 1147/*
f9235b6d
MD
1148 * This function is used to negotiate a passive release of the current
1149 * process/lwp designation with the user scheduler, allowing the user
1150 * scheduler to schedule another user thread. The related kernel thread
1151 * (curthread) continues running in the released state.
8ad65e08
MD
1152 */
1153void
f9235b6d 1154lwkt_passive_release(struct thread *td)
8ad65e08 1155{
f9235b6d
MD
1156 struct lwp *lp = td->td_lwp;
1157
1158 td->td_release = NULL;
1159 lwkt_setpri_self(TDPRI_KERN_USER);
1160 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1161}
1162
f9235b6d 1163
3824f392 1164/*
d2d8515b
MD
1165 * This implements a LWKT yield, allowing a kernel thread to yield to other
1166 * kernel threads at the same or higher priority. This function can be
1167 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1168 *
d2d8515b
MD
1169 * Most kernel threads run at the same priority in order to allow equal
1170 * sharing.
f9235b6d
MD
1171 *
1172 * (self contained on a per cpu basis)
3824f392
MD
1173 */
1174void
f9235b6d 1175lwkt_yield(void)
3824f392 1176{
f9235b6d
MD
1177 globaldata_t gd = mycpu;
1178 thread_t td = gd->gd_curthread;
3824f392 1179
f9235b6d
MD
1180 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1181 splz();
85946b6c 1182 if (lwkt_resched_wanted()) {
f9235b6d
MD
1183 lwkt_schedule_self(curthread);
1184 lwkt_switch();
f9235b6d 1185 }
3824f392
MD
1186}
1187
1188/*
f9235b6d
MD
1189 * This yield is designed for kernel threads with a user context.
1190 *
1191 * The kernel acting on behalf of the user is potentially cpu-bound,
1192 * this function will efficiently allow other threads to run and also
1193 * switch to other processes by releasing.
3824f392
MD
1194 *
1195 * The lwkt_user_yield() function is designed to have very low overhead
1196 * if no yield is determined to be needed.
1197 */
1198void
1199lwkt_user_yield(void)
1200{
f9235b6d
MD
1201 globaldata_t gd = mycpu;
1202 thread_t td = gd->gd_curthread;
1203
1204 /*
1205 * Always run any pending interrupts in case we are in a critical
1206 * section.
1207 */
1208 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1209 splz();
3824f392 1210
3824f392 1211 /*
f9235b6d
MD
1212 * Switch (which forces a release) if another kernel thread needs
1213 * the cpu, if userland wants us to resched, or if our kernel
1214 * quantum has run out.
3824f392 1215 */
f9235b6d 1216 if (lwkt_resched_wanted() ||
85946b6c 1217 user_resched_wanted())
f9235b6d 1218 {
3824f392 1219 lwkt_switch();
3824f392
MD
1220 }
1221
f9235b6d 1222#if 0
3824f392 1223 /*
f9235b6d
MD
1224 * Reacquire the current process if we are released.
1225 *
1226 * XXX not implemented atm. The kernel may be holding locks and such,
1227 * so we want the thread to continue to receive cpu.
3824f392 1228 */
f9235b6d
MD
1229 if (td->td_release == NULL && lp) {
1230 lp->lwp_proc->p_usched->acquire_curproc(lp);
1231 td->td_release = lwkt_passive_release;
1232 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1233 }
f9235b6d 1234#endif
b9eb1c19
MD
1235}
1236
8ad65e08 1237/*
f1d1c3fa
MD
1238 * Generic schedule. Possibly schedule threads belonging to other cpus and
1239 * deal with threads that might be blocked on a wait queue.
1240 *
0a3f9b47
MD
1241 * We have a little helper inline function which does additional work after
1242 * the thread has been enqueued, including dealing with preemption and
1243 * setting need_lwkt_resched() (which prevents the kernel from returning
1244 * to userland until it has processed higher priority threads).
6330a558
MD
1245 *
1246 * It is possible for this routine to be called after a failed _enqueue
1247 * (due to the target thread migrating, sleeping, or otherwise blocked).
1248 * We have to check that the thread is actually on the run queue!
8ad65e08 1249 */
0a3f9b47
MD
1250static __inline
1251void
85946b6c 1252_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1253{
6330a558 1254 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1255 if (ntd->td_preemptable) {
f9235b6d 1256 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1257 }
0a3f9b47
MD
1258 }
1259}
1260
361d01dd 1261static __inline
8ad65e08 1262void
85946b6c 1263_lwkt_schedule(thread_t td)
8ad65e08 1264{
37af14fe
MD
1265 globaldata_t mygd = mycpu;
1266
cf709dd2
MD
1267 KASSERT(td != &td->td_gd->gd_idlethread,
1268 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1269 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1270 crit_enter_gd(mygd);
9388413d 1271 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1272 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1273 _lwkt_enqueue(td);
1274 } else {
f1d1c3fa 1275 /*
7cd8d145
MD
1276 * If we own the thread, there is no race (since we are in a
1277 * critical section). If we do not own the thread there might
1278 * be a race but the target cpu will deal with it.
f1d1c3fa 1279 */
0f7a3396 1280#ifdef SMP
7cd8d145 1281 if (td->td_gd == mygd) {
9d265729 1282 _lwkt_enqueue(td);
85946b6c 1283 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1284 } else {
e381e77c 1285 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1286 }
0f7a3396 1287#else
7cd8d145 1288 _lwkt_enqueue(td);
85946b6c 1289 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1290#endif
8ad65e08 1291 }
37af14fe 1292 crit_exit_gd(mygd);
8ad65e08
MD
1293}
1294
361d01dd
MD
1295void
1296lwkt_schedule(thread_t td)
1297{
85946b6c 1298 _lwkt_schedule(td);
361d01dd
MD
1299}
1300
1301void
85946b6c 1302lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1303{
85946b6c 1304 _lwkt_schedule(td);
361d01dd
MD
1305}
1306
88ebb169
SW
1307#ifdef SMP
1308
e381e77c
MD
1309/*
1310 * When scheduled remotely if frame != NULL the IPIQ is being
1311 * run via doreti or an interrupt then preemption can be allowed.
1312 *
1313 * To allow preemption we have to drop the critical section so only
1314 * one is present in _lwkt_schedule_post.
1315 */
1316static void
1317lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1318{
1319 thread_t td = curthread;
1320 thread_t ntd = arg;
1321
1322 if (frame && ntd->td_preemptable) {
1323 crit_exit_noyield(td);
85946b6c 1324 _lwkt_schedule(ntd);
e381e77c
MD
1325 crit_enter_quick(td);
1326 } else {
85946b6c 1327 _lwkt_schedule(ntd);
e381e77c
MD
1328 }
1329}
1330
d9eea1a5 1331/*
52eedfb5
MD
1332 * Thread migration using a 'Pull' method. The thread may or may not be
1333 * the current thread. It MUST be descheduled and in a stable state.
1334 * lwkt_giveaway() must be called on the cpu owning the thread.
1335 *
1336 * At any point after lwkt_giveaway() is called, the target cpu may
1337 * 'pull' the thread by calling lwkt_acquire().
1338 *
ae8e83e6
MD
1339 * We have to make sure the thread is not sitting on a per-cpu tsleep
1340 * queue or it will blow up when it moves to another cpu.
1341 *
52eedfb5 1342 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1343 */
52eedfb5
MD
1344void
1345lwkt_giveaway(thread_t td)
1346{
3b4192fb 1347 globaldata_t gd = mycpu;
52eedfb5 1348
3b4192fb
MD
1349 crit_enter_gd(gd);
1350 if (td->td_flags & TDF_TSLEEPQ)
1351 tsleep_remove(td);
1352 KKASSERT(td->td_gd == gd);
1353 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1354 td->td_flags |= TDF_MIGRATING;
1355 crit_exit_gd(gd);
52eedfb5
MD
1356}
1357
a2a5ad0d
MD
1358void
1359lwkt_acquire(thread_t td)
1360{
37af14fe
MD
1361 globaldata_t gd;
1362 globaldata_t mygd;
cc9b6223 1363 int retry = 10000000;
a2a5ad0d 1364
52eedfb5 1365 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1366 gd = td->td_gd;
37af14fe 1367 mygd = mycpu;
52eedfb5 1368 if (gd != mycpu) {
35238fa5 1369 cpu_lfence();
52eedfb5 1370 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1371 crit_enter_gd(mygd);
cfaeae2a 1372 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1373 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1374#ifdef SMP
1375 lwkt_process_ipiq();
1376#endif
52eedfb5 1377 cpu_lfence();
cc9b6223
MD
1378 if (--retry == 0) {
1379 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1380 td, td->td_flags);
1381 retry = 10000000;
1382 }
df910c23 1383 }
cfaeae2a 1384 DEBUG_POP_INFO();
562273ea 1385 cpu_mfence();
37af14fe 1386 td->td_gd = mygd;
52eedfb5
MD
1387 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1388 td->td_flags &= ~TDF_MIGRATING;
1389 crit_exit_gd(mygd);
1390 } else {
1391 crit_enter_gd(mygd);
1392 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1393 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1394 crit_exit_gd(mygd);
a2a5ad0d
MD
1395 }
1396}
1397
52eedfb5
MD
1398#endif
1399
f1d1c3fa
MD
1400/*
1401 * Generic deschedule. Descheduling threads other then your own should be
1402 * done only in carefully controlled circumstances. Descheduling is
1403 * asynchronous.
1404 *
1405 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1406 */
1407void
1408lwkt_deschedule(thread_t td)
1409{
f1d1c3fa 1410 crit_enter();
b8a98473 1411#ifdef SMP
f1d1c3fa
MD
1412 if (td == curthread) {
1413 _lwkt_dequeue(td);
1414 } else {
a72187e9 1415 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1416 _lwkt_dequeue(td);
1417 } else {
b8a98473 1418 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1419 }
1420 }
b8a98473
MD
1421#else
1422 _lwkt_dequeue(td);
1423#endif
f1d1c3fa
MD
1424 crit_exit();
1425}
1426
4b5f931b
MD
1427/*
1428 * Set the target thread's priority. This routine does not automatically
1429 * switch to a higher priority thread, LWKT threads are not designed for
1430 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1431 */
1432void
1433lwkt_setpri(thread_t td, int pri)
1434{
f9235b6d
MD
1435 if (td->td_pri != pri) {
1436 KKASSERT(pri >= 0);
1437 crit_enter();
1438 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1439 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1440 _lwkt_dequeue(td);
1441 td->td_pri = pri;
1442 _lwkt_enqueue(td);
1443 } else {
1444 td->td_pri = pri;
1445 }
1446 crit_exit();
26a0694b 1447 }
26a0694b
MD
1448}
1449
03bd0a5e
MD
1450/*
1451 * Set the initial priority for a thread prior to it being scheduled for
1452 * the first time. The thread MUST NOT be scheduled before or during
1453 * this call. The thread may be assigned to a cpu other then the current
1454 * cpu.
1455 *
1456 * Typically used after a thread has been created with TDF_STOPPREQ,
1457 * and before the thread is initially scheduled.
1458 */
1459void
1460lwkt_setpri_initial(thread_t td, int pri)
1461{
1462 KKASSERT(pri >= 0);
1463 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1464 td->td_pri = pri;
03bd0a5e
MD
1465}
1466
26a0694b
MD
1467void
1468lwkt_setpri_self(int pri)
1469{
1470 thread_t td = curthread;
1471
4b5f931b
MD
1472 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1473 crit_enter();
1474 if (td->td_flags & TDF_RUNQ) {
1475 _lwkt_dequeue(td);
f9235b6d 1476 td->td_pri = pri;
4b5f931b
MD
1477 _lwkt_enqueue(td);
1478 } else {
f9235b6d 1479 td->td_pri = pri;
4b5f931b
MD
1480 }
1481 crit_exit();
1482}
1483
f9235b6d 1484/*
85946b6c 1485 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1486 */
1487void
85946b6c 1488lwkt_schedulerclock(thread_t td)
f9235b6d 1489{
85946b6c
MD
1490 globaldata_t gd = td->td_gd;
1491 thread_t xtd;
2a418930 1492
85946b6c
MD
1493 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1494 /*
1495 * If the current thread is at the head of the runq shift it to the
1496 * end of any equal-priority threads and request a LWKT reschedule
1497 * if it moved.
1498 */
1499 xtd = TAILQ_NEXT(td, td_threadq);
1500 if (xtd && xtd->td_pri == td->td_pri) {
1501 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1502 while (xtd && xtd->td_pri == td->td_pri)
1503 xtd = TAILQ_NEXT(xtd, td_threadq);
1504 if (xtd)
1505 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1506 else
1507 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1508 need_lwkt_resched();
f9235b6d 1509 }
85946b6c
MD
1510 } else {
1511 /*
1512 * If we scheduled a thread other than the one at the head of the
1513 * queue always request a reschedule every tick.
1514 */
1515 need_lwkt_resched();
f9235b6d
MD
1516 }
1517}
1518
5d21b981 1519/*
52eedfb5
MD
1520 * Migrate the current thread to the specified cpu.
1521 *
cc9b6223
MD
1522 * This is accomplished by descheduling ourselves from the current cpu
1523 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1524 * 'old' thread wants to migrate after it has been completely switched out
1525 * and will complete the migration.
1526 *
1527 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1528 *
1529 * We must be sure to release our current process designation (if a user
1530 * process) before clearing out any tsleepq we are on because the release
1531 * code may re-add us.
ae8e83e6
MD
1532 *
1533 * We must be sure to remove ourselves from the current cpu's tsleepq
1534 * before potentially moving to another queue. The thread can be on
1535 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1536 */
5d21b981
MD
1537
1538void
1539lwkt_setcpu_self(globaldata_t rgd)
1540{
1541#ifdef SMP
1542 thread_t td = curthread;
1543
1544 if (td->td_gd != rgd) {
1545 crit_enter_quick(td);
cc9b6223 1546
95858b91
MD
1547 if (td->td_release)
1548 td->td_release(td);
ae8e83e6 1549 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1550 tsleep_remove(td);
cc9b6223
MD
1551
1552 /*
1553 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1554 * trying to deschedule ourselves and switch away, then deschedule
1555 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1556 * call lwkt_switch() to complete the operation.
1557 */
5d21b981
MD
1558 td->td_flags |= TDF_MIGRATING;
1559 lwkt_deschedule_self(td);
52eedfb5 1560 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1561 td->td_migrate_gd = rgd;
5d21b981 1562 lwkt_switch();
cc9b6223
MD
1563
1564 /*
1565 * We are now on the target cpu
1566 */
1567 KKASSERT(rgd == mycpu);
52eedfb5 1568 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1569 crit_exit_quick(td);
1570 }
1571#endif
1572}
1573
ecdefdda
MD
1574void
1575lwkt_migratecpu(int cpuid)
1576{
1577#ifdef SMP
1578 globaldata_t rgd;
1579
1580 rgd = globaldata_find(cpuid);
1581 lwkt_setcpu_self(rgd);
1582#endif
1583}
1584
cc9b6223 1585#ifdef SMP
5d21b981
MD
1586/*
1587 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1588 * do not have to enter another one).
1589 *
1590 * The thread (td) has already been completely descheduled from the
1591 * originating cpu and we can simply assert the case. The thread is
1592 * assigned to the new cpu and enqueued.
5d21b981 1593 *
cc9b6223 1594 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1595 */
1596static void
1597lwkt_setcpu_remote(void *arg)
1598{
1599 thread_t td = arg;
1600 globaldata_t gd = mycpu;
1601
cc9b6223 1602 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1603 td->td_gd = gd;
562273ea 1604 cpu_mfence();
5d21b981 1605 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1606 KKASSERT(td->td_migrate_gd == NULL);
9388413d 1607 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1608 _lwkt_enqueue(td);
1609}
3d28ff59 1610#endif
5d21b981 1611
553ea3c8 1612struct lwp *
4b5f931b
MD
1613lwkt_preempted_proc(void)
1614{
73e4f7b9 1615 thread_t td = curthread;
4b5f931b
MD
1616 while (td->td_preempted)
1617 td = td->td_preempted;
553ea3c8 1618 return(td->td_lwp);
4b5f931b
MD
1619}
1620
99df837e
MD
1621/*
1622 * Create a kernel process/thread/whatever. It shares it's address space
1623 * with proc0 - ie: kernel only.
1624 *
d2d8515b
MD
1625 * If the cpu is not specified one will be selected. In the future
1626 * specifying a cpu of -1 will enable kernel thread migration between
1627 * cpus.
99df837e
MD
1628 */
1629int
c9e9fb21
MD
1630lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1631 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1632{
73e4f7b9 1633 thread_t td;
e2565a42 1634 __va_list ap;
99df837e 1635
d3d32139 1636 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1637 tdflags);
a2a5ad0d
MD
1638 if (tdp)
1639 *tdp = td;
709799ea 1640 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1641
1642 /*
1643 * Set up arg0 for 'ps' etc
1644 */
e2565a42 1645 __va_start(ap, fmt);
379210cb 1646 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1647 __va_end(ap);
99df837e
MD
1648
1649 /*
1650 * Schedule the thread to run
1651 */
ef0fdad1
MD
1652 if ((td->td_flags & TDF_STOPREQ) == 0)
1653 lwkt_schedule(td);
1654 else
1655 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1656 return 0;
1657}
1658
1659/*
1660 * Destroy an LWKT thread. Warning! This function is not called when
1661 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1662 * uses a different reaping mechanism.
1663 */
1664void
1665lwkt_exit(void)
1666{
1667 thread_t td = curthread;
c070746a 1668 thread_t std;
8826f33a 1669 globaldata_t gd;
99df837e 1670
2883d2d8
MD
1671 /*
1672 * Do any cleanup that might block here
1673 */
99df837e 1674 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1675 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1676 caps_exit(td);
2883d2d8
MD
1677 biosched_done(td);
1678 dsched_exit_thread(td);
c070746a
MD
1679
1680 /*
1681 * Get us into a critical section to interlock gd_freetd and loop
1682 * until we can get it freed.
1683 *
1684 * We have to cache the current td in gd_freetd because objcache_put()ing
1685 * it would rip it out from under us while our thread is still active.
1686 */
1687 gd = mycpu;
37af14fe 1688 crit_enter_quick(td);
c070746a 1689 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1690 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1691 gd->gd_freetd = NULL;
1692 objcache_put(thread_cache, std);
1693 }
3b4192fb
MD
1694
1695 /*
1696 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1697 * the last time. We cannot block after this point or we may end
1698 * up with a stale td on the tsleepq.
3b4192fb
MD
1699 */
1700 if (td->td_flags & TDF_TSLEEPQ)
1701 tsleep_remove(td);
37af14fe 1702 lwkt_deschedule_self(td);
e56e4dea 1703 lwkt_remove_tdallq(td);
74c9628e 1704 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1705
1706 /*
1707 * Final cleanup
1708 */
1709 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1710 if (td->td_flags & TDF_ALLOCATED_THREAD)
1711 gd->gd_freetd = td;
99df837e
MD
1712 cpu_thread_exit();
1713}
1714
e56e4dea
MD
1715void
1716lwkt_remove_tdallq(thread_t td)
1717{
1718 KKASSERT(td->td_gd == mycpu);
1719 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1720}
1721
9cf43f91
MD
1722/*
1723 * Code reduction and branch prediction improvements. Call/return
1724 * overhead on modern cpus often degenerates into 0 cycles due to
1725 * the cpu's branch prediction hardware and return pc cache. We
1726 * can take advantage of this by not inlining medium-complexity
1727 * functions and we can also reduce the branch prediction impact
1728 * by collapsing perfectly predictable branches into a single
1729 * procedure instead of duplicating it.
1730 *
1731 * Is any of this noticeable? Probably not, so I'll take the
1732 * smaller code size.
1733 */
1734void
b6468f56 1735crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1736{
b6468f56 1737 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1738}
1739
2d93b37a
MD
1740void
1741crit_panic(void)
1742{
1743 thread_t td = curthread;
850634cc 1744 int lcrit = td->td_critcount;
2d93b37a 1745
850634cc
AH
1746 td->td_critcount = 0;
1747 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1748 /* NOT REACHED */
2d93b37a
MD
1749}
1750
d165e668
MD
1751#ifdef SMP
1752
bd8015ca
MD
1753/*
1754 * Called from debugger/panic on cpus which have been stopped. We must still
1755 * process the IPIQ while stopped, even if we were stopped while in a critical
1756 * section (XXX).
1757 *
1758 * If we are dumping also try to process any pending interrupts. This may
1759 * or may not work depending on the state of the cpu at the point it was
1760 * stopped.
1761 */
1762void
1763lwkt_smp_stopped(void)
1764{
1765 globaldata_t gd = mycpu;
1766
1767 crit_enter_gd(gd);
1768 if (dumping) {
1769 lwkt_process_ipiq();
1770 splz();
1771 } else {
1772 lwkt_process_ipiq();
1773 }
1774 crit_exit_gd(gd);
1775}
1776
d165e668 1777#endif