zone.tab: Fix tzsetup(8) breakage.
[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{
c070746a 380 globaldata_t gd = mycpu;
99df837e 381 void *stack;
7d0bac62 382
c070746a
MD
383 /*
384 * If static thread storage is not supplied allocate a thread. Reuse
385 * a cached free thread if possible. gd_freetd is used to keep an exiting
386 * thread intact through the exit.
387 */
ef0fdad1 388 if (td == NULL) {
cf709dd2
MD
389 crit_enter_gd(gd);
390 if ((td = gd->gd_freetd) != NULL) {
391 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
392 TDF_RUNQ)) == 0);
c070746a 393 gd->gd_freetd = NULL;
cf709dd2 394 } else {
c070746a 395 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
396 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
397 TDF_RUNQ)) == 0);
398 }
399 crit_exit_gd(gd);
40aaf5fc
NT
400 KASSERT((td->td_flags &
401 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
402 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
403 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 404 }
c070746a
MD
405
406 /*
407 * Try to reuse cached stack.
408 */
f470d0c8
MD
409 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
410 if (flags & TDF_ALLOCATED_STACK) {
e4846942 411 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
412 stack = NULL;
413 }
414 }
415 if (stack == NULL) {
e40cfbd7 416 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 417 flags |= TDF_ALLOCATED_STACK;
99df837e 418 }
75cdbe6c 419 if (cpu < 0)
c070746a 420 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 421 else
f470d0c8 422 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 423 return(td);
7d0bac62
MD
424}
425
426/*
427 * Initialize a preexisting thread structure. This function is used by
428 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
429 *
f8c3996b
MD
430 * All threads start out in a critical section at a priority of
431 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
432 * appropriate. This function may send an IPI message when the
433 * requested cpu is not the current cpu and consequently gd_tdallq may
434 * not be initialized synchronously from the point of view of the originating
435 * cpu.
436 *
437 * NOTE! we have to be careful in regards to creating threads for other cpus
438 * if SMP has not yet been activated.
7d0bac62 439 */
41a01a4d
MD
440#ifdef SMP
441
75cdbe6c
MD
442static void
443lwkt_init_thread_remote(void *arg)
444{
445 thread_t td = arg;
446
52eedfb5
MD
447 /*
448 * Protected by critical section held by IPI dispatch
449 */
75cdbe6c
MD
450 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
451}
452
41a01a4d
MD
453#endif
454
fdce8919
MD
455/*
456 * lwkt core thread structural initialization.
457 *
458 * NOTE: All threads are initialized as mpsafe threads.
459 */
7d0bac62 460void
f470d0c8
MD
461lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
462 struct globaldata *gd)
7d0bac62 463{
37af14fe
MD
464 globaldata_t mygd = mycpu;
465
99df837e
MD
466 bzero(td, sizeof(struct thread));
467 td->td_kstack = stack;
f470d0c8 468 td->td_kstack_size = stksize;
d3d32139 469 td->td_flags = flags;
26a0694b 470 td->td_gd = gd;
f9235b6d
MD
471 td->td_pri = TDPRI_KERN_DAEMON;
472 td->td_critcount = 1;
3b998fa9 473 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
474 if (lwkt_use_spin_port)
475 lwkt_initport_spin(&td->td_msgport);
476 else
477 lwkt_initport_thread(&td->td_msgport, td);
99df837e 478 pmap_init_thread(td);
0f7a3396 479#ifdef SMP
5d21b981
MD
480 /*
481 * Normally initializing a thread for a remote cpu requires sending an
482 * IPI. However, the idlethread is setup before the other cpus are
483 * activated so we have to treat it as a special case. XXX manipulation
484 * of gd_tdallq requires the BGL.
485 */
486 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 487 crit_enter_gd(mygd);
75cdbe6c 488 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 489 crit_exit_gd(mygd);
75cdbe6c 490 } else {
2db3b277 491 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 492 }
0f7a3396 493#else
37af14fe 494 crit_enter_gd(mygd);
0f7a3396 495 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 496 crit_exit_gd(mygd);
0f7a3396 497#endif
8c72e3d5
AH
498
499 dsched_new_thread(td);
73e4f7b9
MD
500}
501
502void
503lwkt_set_comm(thread_t td, const char *ctl, ...)
504{
e2565a42 505 __va_list va;
73e4f7b9 506
e2565a42 507 __va_start(va, ctl);
379210cb 508 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 509 __va_end(va);
e7c0dbba 510 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
511}
512
99df837e 513void
73e4f7b9 514lwkt_hold(thread_t td)
99df837e 515{
74c9628e 516 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
517}
518
519void
520lwkt_rele(thread_t td)
521{
522 KKASSERT(td->td_refs > 0);
74c9628e 523 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
524}
525
526void
527lwkt_wait_free(thread_t td)
528{
529 while (td->td_refs)
377d4740 530 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
531}
532
533void
534lwkt_free_thread(thread_t td)
535{
74c9628e 536 KKASSERT(td->td_refs == 0);
cf709dd2 537 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
538 if (td->td_flags & TDF_ALLOCATED_THREAD) {
539 objcache_put(thread_cache, td);
540 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
541 /* client-allocated struct with internally allocated stack */
542 KASSERT(td->td_kstack && td->td_kstack_size > 0,
543 ("lwkt_free_thread: corrupted stack"));
544 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
545 td->td_kstack = NULL;
546 td->td_kstack_size = 0;
99df837e 547 }
e7c0dbba 548 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
549}
550
551
8ad65e08
MD
552/*
553 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
554 * switch to the idlethread. Switching must occur within a critical
555 * section to avoid races with the scheduling queue.
556 *
557 * We always have full control over our cpu's run queue. Other cpus
558 * that wish to manipulate our queue must use the cpu_*msg() calls to
559 * talk to our cpu, so a critical section is all that is needed and
560 * the result is very, very fast thread switching.
561 *
96728c05
MD
562 * The LWKT scheduler uses a fixed priority model and round-robins at
563 * each priority level. User process scheduling is a totally
564 * different beast and LWKT priorities should not be confused with
565 * user process priorities.
f1d1c3fa 566 *
69d78e99
MD
567 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
568 * is not called by the current thread in the preemption case, only when
569 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
570 *
571 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
572 * migration and tsleep deschedule the current lwkt thread and call
573 * lwkt_switch(). In particular, the target cpu of the migration fully
574 * expects the thread to become non-runnable and can deadlock against
575 * cpusync operations if we run any IPIs prior to switching the thread out.
576 *
577 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 578 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
579 */
580void
581lwkt_switch(void)
582{
37af14fe
MD
583 globaldata_t gd = mycpu;
584 thread_t td = gd->gd_curthread;
8ad65e08 585 thread_t ntd;
f9235b6d 586 thread_t xtd;
b12defdc 587 int spinning = 0;
8ad65e08 588
da0b0e8b
MD
589 KKASSERT(gd->gd_processing_ipiq == 0);
590
46a3f46d 591 /*
27e88a6e
MD
592 * Switching from within a 'fast' (non thread switched) interrupt or IPI
593 * is illegal. However, we may have to do it anyway if we hit a fatal
594 * kernel trap or we have paniced.
595 *
596 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 597 */
27e88a6e
MD
598 if (gd->gd_intr_nesting_level) {
599 int savegdnest;
600 int savegdtrap;
601
5fddbda2 602 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
603 panic("lwkt_switch: Attempt to switch from a "
604 "a fast interrupt, ipi, or hard code section, "
605 "td %p\n",
606 td);
27e88a6e
MD
607 } else {
608 savegdnest = gd->gd_intr_nesting_level;
609 savegdtrap = gd->gd_trap_nesting_level;
610 gd->gd_intr_nesting_level = 0;
611 gd->gd_trap_nesting_level = 0;
a7422615
MD
612 if ((td->td_flags & TDF_PANICWARN) == 0) {
613 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
614 kprintf("Warning: thread switch from interrupt, IPI, "
615 "or hard code section.\n"
a7422615 616 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 617 print_backtrace(-1);
a7422615 618 }
27e88a6e
MD
619 lwkt_switch();
620 gd->gd_intr_nesting_level = savegdnest;
621 gd->gd_trap_nesting_level = savegdtrap;
622 return;
623 }
96728c05 624 }
ef0fdad1 625
cb973d15 626 /*
85946b6c
MD
627 * Release our current user process designation if we are blocking
628 * or if a user reschedule was requested.
629 *
630 * NOTE: This function is NOT called if we are switching into or
631 * returning from a preemption.
632 *
633 * NOTE: Releasing our current user process designation may cause
634 * it to be assigned to another thread, which in turn will
635 * cause us to block in the usched acquire code when we attempt
636 * to return to userland.
637 *
638 * NOTE: On SMP systems this can be very nasty when heavy token
639 * contention is present so we want to be careful not to
640 * release the designation gratuitously.
cb973d15 641 */
85946b6c
MD
642 if (td->td_release &&
643 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 644 td->td_release(td);
85946b6c 645 }
cb973d15 646
85946b6c
MD
647 /*
648 * Release all tokens
649 */
37af14fe 650 crit_enter_gd(gd);
3b998fa9 651 if (TD_TOKS_HELD(td))
9d265729
MD
652 lwkt_relalltokens(td);
653
654 /*
b02926de
MD
655 * We had better not be holding any spin locks, but don't get into an
656 * endless panic loop.
9d265729 657 */
d666840a
MD
658 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
659 ("lwkt_switch: still holding %d exclusive spinlocks!",
660 gd->gd_spinlocks_wr));
9d265729 661
8a8d5d85
MD
662
663#ifdef SMP
0f7a3396
MD
664#ifdef INVARIANTS
665 if (td->td_cscount) {
6ea70f76 666 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
667 td);
668 if (panic_on_cscount)
669 panic("switching while mastering cpusync");
670 }
671#endif
8a8d5d85 672#endif
f9235b6d
MD
673
674 /*
675 * If we had preempted another thread on this cpu, resume the preempted
676 * thread. This occurs transparently, whether the preempted thread
677 * was scheduled or not (it may have been preempted after descheduling
678 * itself).
679 *
680 * We have to setup the MP lock for the original thread after backing
681 * out the adjustment that was made to curthread when the original
682 * was preempted.
683 */
99df837e 684 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
685 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
686 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
687
688 /*
b9eb1c19
MD
689 * The interrupt may have woken a thread up, we need to properly
690 * set the reschedule flag if the originally interrupted thread is
691 * at a lower priority.
85946b6c
MD
692 *
693 * The interrupt may not have descheduled.
8ec60c3f 694 */
85946b6c 695 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 696 need_lwkt_resched();
f9235b6d
MD
697 goto havethread_preempted;
698 }
699
b12defdc 700 /*
f9235b6d 701 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
702 * schedule the target thread.
703 *
704 * Reminder: Again, we cannot afford to run any IPIs in this path if
705 * the current thread has been descheduled.
f9235b6d
MD
706 */
707 for (;;) {
b12defdc 708 clear_lwkt_resched();
f9235b6d 709
4b5f931b 710 /*
2a418930 711 * Hotpath - pull the head of the run queue and attempt to schedule
85946b6c 712 * it.
41a01a4d 713 */
2a418930
MD
714 for (;;) {
715 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
716
717 if (ntd == NULL) {
718 /*
b12defdc 719 * Runq is empty, switch to idle to allow it to halt.
2a418930
MD
720 */
721 ntd = &gd->gd_idlethread;
6f207a2c 722#ifdef SMP
2a418930 723 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 724 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 725#endif
2a418930
MD
726 cpu_time.cp_msg[0] = 0;
727 cpu_time.cp_stallpc = 0;
2a418930
MD
728 goto haveidle;
729 }
b12defdc 730 break;
f9235b6d 731 }
41a01a4d 732
8ec60c3f 733 /*
b12defdc 734 * Hotpath - schedule ntd.
6f207a2c
MD
735 *
736 * NOTE: For UP there is no mplock and lwkt_getalltokens()
737 * always succeeds.
8ec60c3f 738 */
b12defdc
MD
739 if (TD_TOKS_NOT_HELD(ntd) ||
740 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
741 {
f9235b6d 742 goto havethread;
b12defdc 743 }
f9235b6d 744
f9235b6d 745 /*
2a418930
MD
746 * Coldpath (SMP only since tokens always succeed on UP)
747 *
748 * We had some contention on the thread we wanted to schedule.
749 * What we do now is try to find a thread that we can schedule
b12defdc 750 * in its stead.
2a418930 751 *
85946b6c
MD
752 * The coldpath scan does NOT rearrange threads in the run list.
753 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
754 * the next tick whenever the current head is not the current thread.
f9235b6d 755 */
b12defdc 756#ifdef INVARIANTS
85946b6c
MD
757 ++token_contention_count[ntd->td_pri];
758 ++ntd->td_contended;
b12defdc
MD
759#endif
760
85946b6c 761 if (fairq_bypass > 0)
b12defdc
MD
762 goto skip;
763
b12defdc
MD
764 xtd = NULL;
765 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
85946b6c
MD
766 /*
767 * Never schedule threads returning to userland or the
768 * user thread scheduler helper thread when higher priority
769 * threads are present.
770 */
771 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
772 ntd = NULL;
773 break;
774 }
775
776 /*
777 * Try this one.
778 */
b12defdc
MD
779 if (TD_TOKS_NOT_HELD(ntd) ||
780 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
781 goto havethread;
782 }
85946b6c
MD
783#ifdef INVARIANTS
784 ++token_contention_count[ntd->td_pri];
785 ++ntd->td_contended;
b12defdc 786#endif
2a418930
MD
787 }
788
b12defdc 789skip:
2a418930
MD
790 /*
791 * We exhausted the run list, meaning that all runnable threads
b12defdc 792 * are contested.
2a418930
MD
793 */
794 cpu_pause();
795 ntd = &gd->gd_idlethread;
796#ifdef SMP
797 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
798 ASSERT_NO_TOKENS_HELD(ntd);
799 /* contention case, do not clear contention mask */
800#endif
801
802 /*
b12defdc
MD
803 * We are going to have to retry but if the current thread is not
804 * on the runq we instead switch through the idle thread to get away
805 * from the current thread. We have to flag for lwkt reschedule
806 * to prevent the idle thread from halting.
2a418930 807 *
b12defdc
MD
808 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
809 * instruct it to deal with the potential for deadlocks by
810 * ordering the tokens by address.
2a418930 811 */
b12defdc 812 if ((td->td_flags & TDF_RUNQ) == 0) {
85946b6c 813 need_lwkt_resched(); /* prevent hlt */
2a418930 814 goto haveidle;
4b5f931b 815 }
bd52bedf 816#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
817 if ((read_rflags() & PSL_I) == 0) {
818 cpu_enable_intr();
819 panic("lwkt_switch() called with interrupts disabled");
820 }
821#endif
b12defdc
MD
822
823 /*
824 * Number iterations so far. After a certain point we switch to
825 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
826 */
827 if (spinning < 0x7FFFFFFF)
828 ++spinning;
829
830#ifdef SMP
831 /*
832 * lwkt_getalltokens() failed in sorted token mode, we can use
833 * monitor/mwait in this case.
834 */
835 if (spinning >= lwkt_spin_loops &&
836 (cpu_mi_feature & CPU_MI_MONITOR) &&
837 lwkt_spin_monitor)
838 {
839 cpu_mmw_pause_int(&gd->gd_reqflags,
840 (gd->gd_reqflags | RQF_SPINNING) &
841 ~RQF_IDLECHECK_WK_MASK);
842 }
843#endif
844
845 /*
846 * We already checked that td is still scheduled so this should be
847 * safe.
848 */
849 splz_check();
850
851 /*
852 * This experimental resequencer is used as a fall-back to reduce
853 * hw cache line contention by placing each core's scheduler into a
854 * time-domain-multplexed slot.
855 *
856 * The resequencer is disabled by default. It's functionality has
857 * largely been superceeded by the token algorithm which limits races
858 * to a subset of cores.
859 *
860 * The resequencer algorithm tends to break down when more than
861 * 20 cores are contending. What appears to happen is that new
862 * tokens can be obtained out of address-sorted order by new cores
863 * while existing cores languish in long delays between retries and
864 * wind up being starved-out of the token acquisition.
865 */
866 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
867 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
868 int oseq;
869
870 while ((oseq = lwkt_cseq_rindex) != cseq) {
871 cpu_ccfence();
872#if 1
873 if (cpu_mi_feature & CPU_MI_MONITOR) {
874 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
875 } else {
876#endif
877 cpu_pause();
878 cpu_lfence();
879#if 1
880 }
8b402283 881#endif
0f0466c0 882 }
b12defdc
MD
883 DELAY(1);
884 atomic_add_int(&lwkt_cseq_rindex, 1);
2a418930 885 }
2a418930 886 /* highest level for(;;) loop */
f1d1c3fa 887 }
8a8d5d85 888
2a418930 889havethread:
b12defdc 890 /*
f9235b6d
MD
891 * If the thread we came up with is a higher or equal priority verses
892 * the thread at the head of the queue we move our thread to the
893 * front. This way we can always check the front of the queue.
be71787b
MD
894 *
895 * Clear gd_idle_repeat when doing a normal switch to a non-idle
896 * thread.
f9235b6d 897 */
9ac1ee6e 898 ntd->td_wmesg = NULL;
b12defdc 899 ++gd->gd_cnt.v_swtch;
85946b6c 900#if 0
f9235b6d
MD
901 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
902 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
903 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
904 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
905 }
85946b6c 906#endif
be71787b 907 gd->gd_idle_repeat = 0;
2a418930 908
f9235b6d 909havethread_preempted:
f9235b6d
MD
910 /*
911 * If the new target does not need the MP lock and we are holding it,
912 * release the MP lock. If the new target requires the MP lock we have
913 * already acquired it for the target.
8a8d5d85 914 */
2a418930 915 ;
f9235b6d
MD
916haveidle:
917 KASSERT(ntd->td_critcount,
b5d16701
MD
918 ("priority problem in lwkt_switch %d %d",
919 td->td_critcount, ntd->td_critcount));
920
94f6d86e 921 if (td != ntd) {
cc9b6223
MD
922 /*
923 * Execute the actual thread switch operation. This function
924 * returns to the current thread and returns the previous thread
925 * (which may be different from the thread we switched to).
926 *
927 * We are responsible for marking ntd as TDF_RUNNING.
928 */
94f6d86e 929 ++switch_count;
a1f0fb66 930 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
931 ntd->td_flags |= TDF_RUNNING;
932 lwkt_switch_return(td->td_switch(ntd));
933 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 934 }
b12defdc
MD
935
936#if 1
937 /*
938 * catch-all
939 */
940 splz_check();
941#endif
37af14fe
MD
942 /* NOTE: current cpu may have changed after switch */
943 crit_exit_quick(td);
8ad65e08
MD
944}
945
cc9b6223
MD
946/*
947 * Called by assembly in the td_switch (thread restore path) for thread
948 * bootstrap cases which do not 'return' to lwkt_switch().
949 */
950void
951lwkt_switch_return(thread_t otd)
952{
953#ifdef SMP
954 globaldata_t rgd;
955
956 /*
957 * Check if otd was migrating. Now that we are on ntd we can finish
958 * up the migration. This is a bit messy but it is the only place
959 * where td is known to be fully descheduled.
960 *
961 * We can only activate the migration if otd was migrating but not
962 * held on the cpu due to a preemption chain. We still have to
963 * clear TDF_RUNNING on the old thread either way.
964 *
965 * We are responsible for clearing the previously running thread's
966 * TDF_RUNNING.
967 */
968 if ((rgd = otd->td_migrate_gd) != NULL &&
969 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
970 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
971 (TDF_MIGRATING | TDF_RUNNING));
972 otd->td_migrate_gd = NULL;
973 otd->td_flags &= ~TDF_RUNNING;
974 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
975 } else {
976 otd->td_flags &= ~TDF_RUNNING;
977 }
978#else
979 otd->td_flags &= ~TDF_RUNNING;
980#endif
981}
982
b68b7282 983/*
96728c05
MD
984 * Request that the target thread preempt the current thread. Preemption
985 * only works under a specific set of conditions:
b68b7282 986 *
96728c05
MD
987 * - We are not preempting ourselves
988 * - The target thread is owned by the current cpu
989 * - We are not currently being preempted
990 * - The target is not currently being preempted
d3d1cbc8
MD
991 * - We are not holding any spin locks
992 * - The target thread is not holding any tokens
96728c05
MD
993 * - We are able to satisfy the target's MP lock requirements (if any).
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 *
26a0694b
MD
1001 * XXX at the moment we run the target thread in a critical section during
1002 * the preemption in order to prevent the target from taking interrupts
1003 * that *WE* can't. Preemption is strictly limited to interrupt threads
1004 * and interrupt-like threads, outside of a critical section, and the
1005 * preempted source thread will be resumed the instant the target blocks
1006 * whether or not the source is scheduled (i.e. preemption is supposed to
1007 * be as transparent as possible).
b68b7282
MD
1008 */
1009void
f9235b6d 1010lwkt_preempt(thread_t ntd, int critcount)
b68b7282 1011{
46a3f46d 1012 struct globaldata *gd = mycpu;
cc9b6223 1013 thread_t xtd;
0a3f9b47 1014 thread_t td;
2d910aaf 1015 int save_gd_intr_nesting_level;
b68b7282 1016
26a0694b 1017 /*
96728c05
MD
1018 * The caller has put us in a critical section. We can only preempt
1019 * if the caller of the caller was not in a critical section (basically
f9235b6d 1020 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1021 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1022 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1023 *
1024 * YYY The target thread must be in a critical section (else it must
1025 * inherit our critical section? I dunno yet).
26a0694b 1026 */
f9235b6d 1027 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1028
b12defdc 1029 td = gd->gd_curthread;
fbc024e4 1030 if (preempt_enable == 0) {
85946b6c 1031#if 0
b12defdc
MD
1032 if (ntd->td_pri > td->td_pri)
1033 need_lwkt_resched();
85946b6c 1034#endif
fbc024e4
MD
1035 ++preempt_miss;
1036 return;
1037 }
f9235b6d 1038 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1039 ++preempt_miss;
1040 return;
1041 }
f9235b6d 1042 if (td->td_critcount > critcount) {
96728c05 1043 ++preempt_miss;
85946b6c 1044#if 0
8ec60c3f 1045 need_lwkt_resched();
85946b6c 1046#endif
96728c05
MD
1047 return;
1048 }
1049#ifdef SMP
46a3f46d 1050 if (ntd->td_gd != gd) {
96728c05 1051 ++preempt_miss;
85946b6c 1052#if 0
8ec60c3f 1053 need_lwkt_resched();
85946b6c 1054#endif
96728c05
MD
1055 return;
1056 }
1057#endif
41a01a4d 1058 /*
77912481
MD
1059 * We don't have to check spinlocks here as they will also bump
1060 * td_critcount.
d3d1cbc8
MD
1061 *
1062 * Do not try to preempt if the target thread is holding any tokens.
1063 * We could try to acquire the tokens but this case is so rare there
1064 * is no need to support it.
41a01a4d 1065 */
77912481
MD
1066 KKASSERT(gd->gd_spinlocks_wr == 0);
1067
3b998fa9 1068 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 1069 ++preempt_miss;
85946b6c 1070#if 0
d3d1cbc8 1071 need_lwkt_resched();
85946b6c 1072#endif
d3d1cbc8
MD
1073 return;
1074 }
26a0694b
MD
1075 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1076 ++preempt_weird;
85946b6c 1077#if 0
8ec60c3f 1078 need_lwkt_resched();
85946b6c 1079#endif
26a0694b
MD
1080 return;
1081 }
1082 if (ntd->td_preempted) {
4b5f931b 1083 ++preempt_hit;
85946b6c 1084#if 0
8ec60c3f 1085 need_lwkt_resched();
85946b6c 1086#endif
26a0694b 1087 return;
b68b7282 1088 }
da0b0e8b 1089 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1090
8ec60c3f
MD
1091 /*
1092 * Since we are able to preempt the current thread, there is no need to
1093 * call need_lwkt_resched().
2d910aaf
MD
1094 *
1095 * We must temporarily clear gd_intr_nesting_level around the switch
1096 * since switchouts from the target thread are allowed (they will just
1097 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1098 *
1099 * A preemption must switch back to the original thread, assert the
1100 * case.
8ec60c3f 1101 */
26a0694b
MD
1102 ++preempt_hit;
1103 ntd->td_preempted = td;
1104 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1105 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1106 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1107 gd->gd_intr_nesting_level = 0;
cc9b6223
MD
1108 ntd->td_flags |= TDF_RUNNING;
1109 xtd = td->td_switch(ntd);
1110 KKASSERT(xtd == ntd);
1111 lwkt_switch_return(xtd);
2d910aaf 1112 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1113
26a0694b
MD
1114 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1115 ntd->td_preempted = NULL;
1116 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b12defdc
MD
1117#if 1
1118 /*
1119 * catch-all
1120 */
1121 splz_check();
1122#endif
b68b7282
MD
1123}
1124
f1d1c3fa 1125/*
faaeffac 1126 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1127 * This will work inside a critical section but not inside a hard code
1128 * section.
ef0fdad1 1129 *
f1d1c3fa
MD
1130 * (self contained on a per cpu basis)
1131 */
1132void
faaeffac 1133splz_check(void)
f1d1c3fa 1134{
7966cb69
MD
1135 globaldata_t gd = mycpu;
1136 thread_t td = gd->gd_curthread;
ef0fdad1 1137
4a28fe22
MD
1138 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1139 gd->gd_intr_nesting_level == 0 &&
1140 td->td_nest_count < 2)
1141 {
f1d1c3fa 1142 splz();
4a28fe22
MD
1143 }
1144}
1145
1146/*
1147 * This version is integrated into crit_exit, reqflags has already
1148 * been tested but td_critcount has not.
1149 *
1150 * We only want to execute the splz() on the 1->0 transition of
1151 * critcount and not in a hard code section or if too deeply nested.
1152 */
1153void
1154lwkt_maybe_splz(thread_t td)
1155{
1156 globaldata_t gd = td->td_gd;
1157
1158 if (td->td_critcount == 0 &&
1159 gd->gd_intr_nesting_level == 0 &&
1160 td->td_nest_count < 2)
1161 {
1162 splz();
1163 }
f1d1c3fa
MD
1164}
1165
8ad65e08 1166/*
f9235b6d
MD
1167 * This function is used to negotiate a passive release of the current
1168 * process/lwp designation with the user scheduler, allowing the user
1169 * scheduler to schedule another user thread. The related kernel thread
1170 * (curthread) continues running in the released state.
8ad65e08
MD
1171 */
1172void
f9235b6d 1173lwkt_passive_release(struct thread *td)
8ad65e08 1174{
f9235b6d
MD
1175 struct lwp *lp = td->td_lwp;
1176
1177 td->td_release = NULL;
1178 lwkt_setpri_self(TDPRI_KERN_USER);
1179 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1180}
1181
f9235b6d 1182
3824f392 1183/*
f9235b6d
MD
1184 * This implements a normal yield. This routine is virtually a nop if
1185 * there is nothing to yield to but it will always run any pending interrupts
1186 * if called from a critical section.
1187 *
1188 * This yield is designed for kernel threads without a user context.
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);
9388413d 1289 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1290 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1291 _lwkt_enqueue(td);
1292 } else {
f1d1c3fa 1293 /*
7cd8d145
MD
1294 * If we own the thread, there is no race (since we are in a
1295 * critical section). If we do not own the thread there might
1296 * be a race but the target cpu will deal with it.
f1d1c3fa 1297 */
0f7a3396 1298#ifdef SMP
7cd8d145 1299 if (td->td_gd == mygd) {
9d265729 1300 _lwkt_enqueue(td);
85946b6c 1301 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1302 } else {
e381e77c 1303 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1304 }
0f7a3396 1305#else
7cd8d145 1306 _lwkt_enqueue(td);
85946b6c 1307 _lwkt_schedule_post(mygd, td, 1);
0f7a3396 1308#endif
8ad65e08 1309 }
37af14fe 1310 crit_exit_gd(mygd);
8ad65e08
MD
1311}
1312
361d01dd
MD
1313void
1314lwkt_schedule(thread_t td)
1315{
85946b6c 1316 _lwkt_schedule(td);
361d01dd
MD
1317}
1318
1319void
85946b6c 1320lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1321{
85946b6c 1322 _lwkt_schedule(td);
361d01dd
MD
1323}
1324
88ebb169
SW
1325#ifdef SMP
1326
e381e77c
MD
1327/*
1328 * When scheduled remotely if frame != NULL the IPIQ is being
1329 * run via doreti or an interrupt then preemption can be allowed.
1330 *
1331 * To allow preemption we have to drop the critical section so only
1332 * one is present in _lwkt_schedule_post.
1333 */
1334static void
1335lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1336{
1337 thread_t td = curthread;
1338 thread_t ntd = arg;
1339
1340 if (frame && ntd->td_preemptable) {
1341 crit_exit_noyield(td);
85946b6c 1342 _lwkt_schedule(ntd);
e381e77c
MD
1343 crit_enter_quick(td);
1344 } else {
85946b6c 1345 _lwkt_schedule(ntd);
e381e77c
MD
1346 }
1347}
1348
d9eea1a5 1349/*
52eedfb5
MD
1350 * Thread migration using a 'Pull' method. The thread may or may not be
1351 * the current thread. It MUST be descheduled and in a stable state.
1352 * lwkt_giveaway() must be called on the cpu owning the thread.
1353 *
1354 * At any point after lwkt_giveaway() is called, the target cpu may
1355 * 'pull' the thread by calling lwkt_acquire().
1356 *
ae8e83e6
MD
1357 * We have to make sure the thread is not sitting on a per-cpu tsleep
1358 * queue or it will blow up when it moves to another cpu.
1359 *
52eedfb5 1360 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1361 */
52eedfb5
MD
1362void
1363lwkt_giveaway(thread_t td)
1364{
3b4192fb 1365 globaldata_t gd = mycpu;
52eedfb5 1366
3b4192fb
MD
1367 crit_enter_gd(gd);
1368 if (td->td_flags & TDF_TSLEEPQ)
1369 tsleep_remove(td);
1370 KKASSERT(td->td_gd == gd);
1371 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1372 td->td_flags |= TDF_MIGRATING;
1373 crit_exit_gd(gd);
52eedfb5
MD
1374}
1375
a2a5ad0d
MD
1376void
1377lwkt_acquire(thread_t td)
1378{
37af14fe
MD
1379 globaldata_t gd;
1380 globaldata_t mygd;
cc9b6223 1381 int retry = 10000000;
a2a5ad0d 1382
52eedfb5 1383 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1384 gd = td->td_gd;
37af14fe 1385 mygd = mycpu;
52eedfb5 1386 if (gd != mycpu) {
35238fa5 1387 cpu_lfence();
52eedfb5 1388 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1389 crit_enter_gd(mygd);
cfaeae2a 1390 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1391 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1392#ifdef SMP
1393 lwkt_process_ipiq();
1394#endif
52eedfb5 1395 cpu_lfence();
cc9b6223
MD
1396 if (--retry == 0) {
1397 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1398 td, td->td_flags);
1399 retry = 10000000;
1400 }
df910c23 1401 }
cfaeae2a 1402 DEBUG_POP_INFO();
562273ea 1403 cpu_mfence();
37af14fe 1404 td->td_gd = mygd;
52eedfb5
MD
1405 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1406 td->td_flags &= ~TDF_MIGRATING;
1407 crit_exit_gd(mygd);
1408 } else {
1409 crit_enter_gd(mygd);
1410 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1411 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1412 crit_exit_gd(mygd);
a2a5ad0d
MD
1413 }
1414}
1415
52eedfb5
MD
1416#endif
1417
f1d1c3fa
MD
1418/*
1419 * Generic deschedule. Descheduling threads other then your own should be
1420 * done only in carefully controlled circumstances. Descheduling is
1421 * asynchronous.
1422 *
1423 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1424 */
1425void
1426lwkt_deschedule(thread_t td)
1427{
f1d1c3fa 1428 crit_enter();
b8a98473 1429#ifdef SMP
f1d1c3fa
MD
1430 if (td == curthread) {
1431 _lwkt_dequeue(td);
1432 } else {
a72187e9 1433 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1434 _lwkt_dequeue(td);
1435 } else {
b8a98473 1436 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1437 }
1438 }
b8a98473
MD
1439#else
1440 _lwkt_dequeue(td);
1441#endif
f1d1c3fa
MD
1442 crit_exit();
1443}
1444
4b5f931b
MD
1445/*
1446 * Set the target thread's priority. This routine does not automatically
1447 * switch to a higher priority thread, LWKT threads are not designed for
1448 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1449 */
1450void
1451lwkt_setpri(thread_t td, int pri)
1452{
a72187e9 1453 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1454 if (td->td_pri != pri) {
1455 KKASSERT(pri >= 0);
1456 crit_enter();
1457 if (td->td_flags & TDF_RUNQ) {
1458 _lwkt_dequeue(td);
1459 td->td_pri = pri;
1460 _lwkt_enqueue(td);
1461 } else {
1462 td->td_pri = pri;
1463 }
1464 crit_exit();
26a0694b 1465 }
26a0694b
MD
1466}
1467
03bd0a5e
MD
1468/*
1469 * Set the initial priority for a thread prior to it being scheduled for
1470 * the first time. The thread MUST NOT be scheduled before or during
1471 * this call. The thread may be assigned to a cpu other then the current
1472 * cpu.
1473 *
1474 * Typically used after a thread has been created with TDF_STOPPREQ,
1475 * and before the thread is initially scheduled.
1476 */
1477void
1478lwkt_setpri_initial(thread_t td, int pri)
1479{
1480 KKASSERT(pri >= 0);
1481 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1482 td->td_pri = pri;
03bd0a5e
MD
1483}
1484
26a0694b
MD
1485void
1486lwkt_setpri_self(int pri)
1487{
1488 thread_t td = curthread;
1489
4b5f931b
MD
1490 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1491 crit_enter();
1492 if (td->td_flags & TDF_RUNQ) {
1493 _lwkt_dequeue(td);
f9235b6d 1494 td->td_pri = pri;
4b5f931b
MD
1495 _lwkt_enqueue(td);
1496 } else {
f9235b6d 1497 td->td_pri = pri;
4b5f931b
MD
1498 }
1499 crit_exit();
1500}
1501
f9235b6d 1502/*
85946b6c 1503 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1504 */
1505void
85946b6c 1506lwkt_schedulerclock(thread_t td)
f9235b6d 1507{
85946b6c
MD
1508 globaldata_t gd = td->td_gd;
1509 thread_t xtd;
2a418930 1510
85946b6c
MD
1511 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1512 /*
1513 * If the current thread is at the head of the runq shift it to the
1514 * end of any equal-priority threads and request a LWKT reschedule
1515 * if it moved.
1516 */
1517 xtd = TAILQ_NEXT(td, td_threadq);
1518 if (xtd && xtd->td_pri == td->td_pri) {
1519 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1520 while (xtd && xtd->td_pri == td->td_pri)
1521 xtd = TAILQ_NEXT(xtd, td_threadq);
1522 if (xtd)
1523 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1524 else
1525 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1526 need_lwkt_resched();
f9235b6d 1527 }
85946b6c
MD
1528 } else {
1529 /*
1530 * If we scheduled a thread other than the one at the head of the
1531 * queue always request a reschedule every tick.
1532 */
1533 need_lwkt_resched();
f9235b6d
MD
1534 }
1535}
1536
5d21b981 1537/*
52eedfb5
MD
1538 * Migrate the current thread to the specified cpu.
1539 *
cc9b6223
MD
1540 * This is accomplished by descheduling ourselves from the current cpu
1541 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1542 * 'old' thread wants to migrate after it has been completely switched out
1543 * and will complete the migration.
1544 *
1545 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1546 *
1547 * We must be sure to release our current process designation (if a user
1548 * process) before clearing out any tsleepq we are on because the release
1549 * code may re-add us.
ae8e83e6
MD
1550 *
1551 * We must be sure to remove ourselves from the current cpu's tsleepq
1552 * before potentially moving to another queue. The thread can be on
1553 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1554 */
5d21b981
MD
1555
1556void
1557lwkt_setcpu_self(globaldata_t rgd)
1558{
1559#ifdef SMP
1560 thread_t td = curthread;
1561
1562 if (td->td_gd != rgd) {
1563 crit_enter_quick(td);
cc9b6223 1564
95858b91
MD
1565 if (td->td_release)
1566 td->td_release(td);
ae8e83e6 1567 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1568 tsleep_remove(td);
cc9b6223
MD
1569
1570 /*
1571 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1572 * trying to deschedule ourselves and switch away, then deschedule
1573 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1574 * call lwkt_switch() to complete the operation.
1575 */
5d21b981
MD
1576 td->td_flags |= TDF_MIGRATING;
1577 lwkt_deschedule_self(td);
52eedfb5 1578 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1579 td->td_migrate_gd = rgd;
5d21b981 1580 lwkt_switch();
cc9b6223
MD
1581
1582 /*
1583 * We are now on the target cpu
1584 */
1585 KKASSERT(rgd == mycpu);
52eedfb5 1586 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1587 crit_exit_quick(td);
1588 }
1589#endif
1590}
1591
ecdefdda
MD
1592void
1593lwkt_migratecpu(int cpuid)
1594{
1595#ifdef SMP
1596 globaldata_t rgd;
1597
1598 rgd = globaldata_find(cpuid);
1599 lwkt_setcpu_self(rgd);
1600#endif
1601}
1602
cc9b6223 1603#ifdef SMP
5d21b981
MD
1604/*
1605 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1606 * do not have to enter another one).
1607 *
1608 * The thread (td) has already been completely descheduled from the
1609 * originating cpu and we can simply assert the case. The thread is
1610 * assigned to the new cpu and enqueued.
5d21b981 1611 *
cc9b6223 1612 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1613 */
1614static void
1615lwkt_setcpu_remote(void *arg)
1616{
1617 thread_t td = arg;
1618 globaldata_t gd = mycpu;
1619
cc9b6223 1620 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1621 td->td_gd = gd;
562273ea 1622 cpu_mfence();
5d21b981 1623 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1624 KKASSERT(td->td_migrate_gd == NULL);
9388413d 1625 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1626 _lwkt_enqueue(td);
1627}
3d28ff59 1628#endif
5d21b981 1629
553ea3c8 1630struct lwp *
4b5f931b
MD
1631lwkt_preempted_proc(void)
1632{
73e4f7b9 1633 thread_t td = curthread;
4b5f931b
MD
1634 while (td->td_preempted)
1635 td = td->td_preempted;
553ea3c8 1636 return(td->td_lwp);
4b5f931b
MD
1637}
1638
99df837e
MD
1639/*
1640 * Create a kernel process/thread/whatever. It shares it's address space
1641 * with proc0 - ie: kernel only.
1642 *
365fa13f
MD
1643 * NOTE! By default new threads are created with the MP lock held. A
1644 * thread which does not require the MP lock should release it by calling
1645 * rel_mplock() at the start of the new thread.
99df837e
MD
1646 */
1647int
c9e9fb21
MD
1648lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1649 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1650{
73e4f7b9 1651 thread_t td;
e2565a42 1652 __va_list ap;
99df837e 1653
d3d32139 1654 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1655 tdflags);
a2a5ad0d
MD
1656 if (tdp)
1657 *tdp = td;
709799ea 1658 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1659
1660 /*
1661 * Set up arg0 for 'ps' etc
1662 */
e2565a42 1663 __va_start(ap, fmt);
379210cb 1664 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1665 __va_end(ap);
99df837e
MD
1666
1667 /*
1668 * Schedule the thread to run
1669 */
ef0fdad1
MD
1670 if ((td->td_flags & TDF_STOPREQ) == 0)
1671 lwkt_schedule(td);
1672 else
1673 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1674 return 0;
1675}
1676
1677/*
1678 * Destroy an LWKT thread. Warning! This function is not called when
1679 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1680 * uses a different reaping mechanism.
1681 */
1682void
1683lwkt_exit(void)
1684{
1685 thread_t td = curthread;
c070746a 1686 thread_t std;
8826f33a 1687 globaldata_t gd;
99df837e 1688
2883d2d8
MD
1689 /*
1690 * Do any cleanup that might block here
1691 */
99df837e 1692 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1693 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1694 caps_exit(td);
2883d2d8
MD
1695 biosched_done(td);
1696 dsched_exit_thread(td);
c070746a
MD
1697
1698 /*
1699 * Get us into a critical section to interlock gd_freetd and loop
1700 * until we can get it freed.
1701 *
1702 * We have to cache the current td in gd_freetd because objcache_put()ing
1703 * it would rip it out from under us while our thread is still active.
1704 */
1705 gd = mycpu;
37af14fe 1706 crit_enter_quick(td);
c070746a 1707 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1708 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1709 gd->gd_freetd = NULL;
1710 objcache_put(thread_cache, std);
1711 }
3b4192fb
MD
1712
1713 /*
1714 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1715 * the last time. We cannot block after this point or we may end
1716 * up with a stale td on the tsleepq.
3b4192fb
MD
1717 */
1718 if (td->td_flags & TDF_TSLEEPQ)
1719 tsleep_remove(td);
37af14fe 1720 lwkt_deschedule_self(td);
e56e4dea 1721 lwkt_remove_tdallq(td);
74c9628e 1722 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1723
1724 /*
1725 * Final cleanup
1726 */
1727 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1728 if (td->td_flags & TDF_ALLOCATED_THREAD)
1729 gd->gd_freetd = td;
99df837e
MD
1730 cpu_thread_exit();
1731}
1732
e56e4dea
MD
1733void
1734lwkt_remove_tdallq(thread_t td)
1735{
1736 KKASSERT(td->td_gd == mycpu);
1737 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1738}
1739
9cf43f91
MD
1740/*
1741 * Code reduction and branch prediction improvements. Call/return
1742 * overhead on modern cpus often degenerates into 0 cycles due to
1743 * the cpu's branch prediction hardware and return pc cache. We
1744 * can take advantage of this by not inlining medium-complexity
1745 * functions and we can also reduce the branch prediction impact
1746 * by collapsing perfectly predictable branches into a single
1747 * procedure instead of duplicating it.
1748 *
1749 * Is any of this noticeable? Probably not, so I'll take the
1750 * smaller code size.
1751 */
1752void
b6468f56 1753crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1754{
b6468f56 1755 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1756}
1757
2d93b37a
MD
1758void
1759crit_panic(void)
1760{
1761 thread_t td = curthread;
850634cc 1762 int lcrit = td->td_critcount;
2d93b37a 1763
850634cc
AH
1764 td->td_critcount = 0;
1765 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1766 /* NOT REACHED */
2d93b37a
MD
1767}
1768
d165e668
MD
1769#ifdef SMP
1770
bd8015ca
MD
1771/*
1772 * Called from debugger/panic on cpus which have been stopped. We must still
1773 * process the IPIQ while stopped, even if we were stopped while in a critical
1774 * section (XXX).
1775 *
1776 * If we are dumping also try to process any pending interrupts. This may
1777 * or may not work depending on the state of the cpu at the point it was
1778 * stopped.
1779 */
1780void
1781lwkt_smp_stopped(void)
1782{
1783 globaldata_t gd = mycpu;
1784
1785 crit_enter_gd(gd);
1786 if (dumping) {
1787 lwkt_process_ipiq();
1788 splz();
1789 } else {
1790 lwkt_process_ipiq();
1791 }
1792 crit_exit_gd(gd);
1793}
1794
d165e668 1795#endif