kernel - Fix excessive ipiq recursion (3)
[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>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
8cee56f4
MD
74#ifdef _KERNEL_VIRTUAL
75#include <pthread.h>
76#endif
77
d850923c
AE
78#if !defined(KTR_CTXSW)
79#define KTR_CTXSW KTR_ALL
80#endif
81KTR_INFO_MASTER(ctxsw);
5bf48697
AE
82KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
83KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
84KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
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
e28c8ef4
SW
92static int64_t switch_count = 0;
93static int64_t preempt_hit = 0;
94static int64_t preempt_miss = 0;
95static int64_t preempt_weird = 0;
fb0f29c4 96static int lwkt_use_spin_port;
40aaf5fc 97static struct objcache *thread_cache;
a46b4a23 98int cpu_mwait_spin = 0;
05220613 99
e381e77c 100static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 101static void lwkt_setcpu_remote(void *arg);
e381e77c 102
fb0f29c4
MD
103/*
104 * We can make all thread ports use the spin backend instead of the thread
105 * backend. This should only be set to debug the spin backend.
106 */
107TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
108
0f7a3396 109#ifdef INVARIANTS
0c52fa62
SG
110SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
111 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 112#endif
0c52fa62
SG
113SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
114 "Number of switched threads");
9733f757 115SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 116 "Successful preemption events");
9733f757 117SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
118 "Failed preemption events");
119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
120 "Number of preempted threads.");
b12defdc 121static int fairq_enable = 0;
2a418930
MD
122SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
123 &fairq_enable, 0, "Turn on fairq priority accumulators");
85946b6c 124static int fairq_bypass = -1;
b12defdc
MD
125SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
127extern int lwkt_sched_debug;
128int lwkt_sched_debug = 0;
129SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
130 &lwkt_sched_debug, 0, "Scheduler debug");
2a418930
MD
131static int lwkt_spin_loops = 10;
132SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc 133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
fbc024e4 134static int preempt_enable = 1;
2a418930
MD
135SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
136 &preempt_enable, 0, "Enable preemption");
7b234d8c 137static int lwkt_cache_threads = 0;
765b1ae0
MD
138SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
139 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 140
4b5f931b
MD
141/*
142 * These helper procedures handle the runq, they can only be called from
143 * within a critical section.
75cdbe6c
MD
144 *
145 * WARNING! Prior to SMP being brought up it is possible to enqueue and
146 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
147 * instead of 'mycpu' when referencing the globaldata structure. Once
148 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 149 */
f1d1c3fa
MD
150static __inline
151void
152_lwkt_dequeue(thread_t td)
153{
154 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 155 struct globaldata *gd = td->td_gd;
4b5f931b 156
f1d1c3fa 157 td->td_flags &= ~TDF_RUNQ;
f9235b6d 158 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 159 --gd->gd_tdrunqcount;
f9235b6d 160 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 161 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
162 }
163}
164
f9235b6d
MD
165/*
166 * Priority enqueue.
167 *
d992c377
MD
168 * There are a limited number of lwkt threads runnable since user
169 * processes only schedule one at a time per cpu. However, there can
170 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 171 * become runnable.
d992c377
MD
172 *
173 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
174 * will ignore user priority. This is to ensure that user threads in
175 * kernel mode get cpu at some point regardless of what the user
176 * scheduler thinks.
f9235b6d 177 */
f1d1c3fa
MD
178static __inline
179void
180_lwkt_enqueue(thread_t td)
181{
f9235b6d
MD
182 thread_t xtd;
183
7f5d7ed7 184 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 185 struct globaldata *gd = td->td_gd;
4b5f931b 186
f1d1c3fa 187 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
188 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
189 if (xtd == NULL) {
85946b6c
MD
190 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
191 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 192 } else {
e3e6be1f
MD
193 /*
194 * NOTE: td_upri - higher numbers more desireable, same sense
195 * as td_pri (typically reversed from lwp_upri).
196 *
197 * In the equal priority case we want the best selection
198 * at the beginning so the less desireable selections know
199 * that they have to setrunqueue/go-to-another-cpu, even
200 * though it means switching back to the 'best' selection.
201 * This also avoids degenerate situations when many threads
202 * are runnable or waking up at the same time.
203 *
204 * If upri matches exactly place at end/round-robin.
205 */
d992c377 206 while (xtd &&
e3e6be1f 207 (xtd->td_pri >= td->td_pri ||
d992c377 208 (xtd->td_pri == td->td_pri &&
e3e6be1f 209 xtd->td_upri >= td->td_upri))) {
85946b6c 210 xtd = TAILQ_NEXT(xtd, td_threadq);
d992c377 211 }
85946b6c
MD
212 if (xtd)
213 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
214 else
215 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 216 }
de4d4cb0 217 ++gd->gd_tdrunqcount;
b12defdc
MD
218
219 /*
85946b6c 220 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 221 */
85946b6c
MD
222 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
223 need_lwkt_resched();
f1d1c3fa
MD
224 }
225}
8ad65e08 226
e28c8ef4 227static boolean_t
40aaf5fc
NT
228_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
229{
230 struct thread *td = (struct thread *)obj;
231
232 td->td_kstack = NULL;
233 td->td_kstack_size = 0;
234 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 235 td->td_mpflags = 0;
40aaf5fc
NT
236 return (1);
237}
238
239static void
240_lwkt_thread_dtor(void *obj, void *privdata)
241{
242 struct thread *td = (struct thread *)obj;
243
244 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
245 ("_lwkt_thread_dtor: not allocated from objcache"));
246 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
247 td->td_kstack_size > 0,
248 ("_lwkt_thread_dtor: corrupted stack"));
249 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
250 td->td_kstack = NULL;
251 td->td_flags = 0;
40aaf5fc
NT
252}
253
254/*
255 * Initialize the lwkt s/system.
765b1ae0 256 *
7b234d8c
MD
257 * Nominally cache up to 32 thread + kstack structures. Cache more on
258 * systems with a lot of cpu cores.
40aaf5fc 259 */
ced589cb 260static void
40aaf5fc
NT
261lwkt_init(void)
262{
765b1ae0 263 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
264 if (lwkt_cache_threads == 0) {
265 lwkt_cache_threads = ncpus * 4;
266 if (lwkt_cache_threads < 32)
267 lwkt_cache_threads = 32;
268 }
765b1ae0
MD
269 thread_cache = objcache_create_mbacked(
270 M_THREAD, sizeof(struct thread),
2fce2579 271 0, lwkt_cache_threads,
765b1ae0 272 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc 273}
ced589cb 274SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
40aaf5fc 275
37af14fe
MD
276/*
277 * Schedule a thread to run. As the current thread we can always safely
278 * schedule ourselves, and a shortcut procedure is provided for that
279 * function.
280 *
281 * (non-blocking, self contained on a per cpu basis)
282 */
283void
284lwkt_schedule_self(thread_t td)
285{
cfaeae2a 286 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 287 crit_enter_quick(td);
f9235b6d
MD
288 KASSERT(td != &td->td_gd->gd_idlethread,
289 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
290 KKASSERT(td->td_lwp == NULL ||
291 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 292 _lwkt_enqueue(td);
37af14fe
MD
293 crit_exit_quick(td);
294}
295
296/*
297 * Deschedule a thread.
298 *
299 * (non-blocking, self contained on a per cpu basis)
300 */
301void
302lwkt_deschedule_self(thread_t td)
303{
304 crit_enter_quick(td);
37af14fe
MD
305 _lwkt_dequeue(td);
306 crit_exit_quick(td);
307}
308
8ad65e08
MD
309/*
310 * LWKTs operate on a per-cpu basis
311 *
73e4f7b9 312 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
313 */
314void
315lwkt_gdinit(struct globaldata *gd)
316{
f9235b6d 317 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 318 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
319}
320
7d0bac62
MD
321/*
322 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
323 * or LWKT start address before it can be scheduled. If the target cpu is
324 * -1 the thread will be created on the current cpu.
0cfcada1
MD
325 *
326 * If you intend to create a thread without a process context this function
327 * does everything except load the startup and switcher function.
7d0bac62
MD
328 */
329thread_t
d3d32139 330lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 331{
d2d8515b 332 static int cpu_rotator;
c070746a 333 globaldata_t gd = mycpu;
99df837e 334 void *stack;
7d0bac62 335
c070746a
MD
336 /*
337 * If static thread storage is not supplied allocate a thread. Reuse
338 * a cached free thread if possible. gd_freetd is used to keep an exiting
339 * thread intact through the exit.
340 */
ef0fdad1 341 if (td == NULL) {
cf709dd2
MD
342 crit_enter_gd(gd);
343 if ((td = gd->gd_freetd) != NULL) {
344 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
345 TDF_RUNQ)) == 0);
c070746a 346 gd->gd_freetd = NULL;
cf709dd2 347 } else {
c070746a 348 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
349 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
350 TDF_RUNQ)) == 0);
351 }
352 crit_exit_gd(gd);
40aaf5fc 353 KASSERT((td->td_flags &
2af9d75d
MD
354 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
355 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
356 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
357 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 358 }
c070746a
MD
359
360 /*
361 * Try to reuse cached stack.
362 */
f470d0c8
MD
363 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
364 if (flags & TDF_ALLOCATED_STACK) {
e4846942 365 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
366 stack = NULL;
367 }
368 }
369 if (stack == NULL) {
e40cfbd7 370 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 371 flags |= TDF_ALLOCATED_STACK;
99df837e 372 }
d2d8515b
MD
373 if (cpu < 0) {
374 cpu = ++cpu_rotator;
375 cpu_ccfence();
376 cpu %= ncpus;
377 }
378 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 379 return(td);
7d0bac62
MD
380}
381
382/*
383 * Initialize a preexisting thread structure. This function is used by
384 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
385 *
f8c3996b
MD
386 * All threads start out in a critical section at a priority of
387 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
388 * appropriate. This function may send an IPI message when the
389 * requested cpu is not the current cpu and consequently gd_tdallq may
390 * not be initialized synchronously from the point of view of the originating
391 * cpu.
392 *
393 * NOTE! we have to be careful in regards to creating threads for other cpus
394 * if SMP has not yet been activated.
7d0bac62 395 */
75cdbe6c
MD
396static void
397lwkt_init_thread_remote(void *arg)
398{
399 thread_t td = arg;
400
52eedfb5
MD
401 /*
402 * Protected by critical section held by IPI dispatch
403 */
75cdbe6c
MD
404 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
405}
406
fdce8919
MD
407/*
408 * lwkt core thread structural initialization.
409 *
410 * NOTE: All threads are initialized as mpsafe threads.
411 */
7d0bac62 412void
f470d0c8
MD
413lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
414 struct globaldata *gd)
7d0bac62 415{
37af14fe
MD
416 globaldata_t mygd = mycpu;
417
99df837e
MD
418 bzero(td, sizeof(struct thread));
419 td->td_kstack = stack;
f470d0c8 420 td->td_kstack_size = stksize;
d3d32139 421 td->td_flags = flags;
4643740a 422 td->td_mpflags = 0;
f256b6c0 423 td->td_type = TD_TYPE_GENERIC;
26a0694b 424 td->td_gd = gd;
f9235b6d
MD
425 td->td_pri = TDPRI_KERN_DAEMON;
426 td->td_critcount = 1;
54341a3b 427 td->td_toks_have = NULL;
3b998fa9 428 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
429 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
430 lwkt_initport_spin(&td->td_msgport, td,
431 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
432 } else {
fb0f29c4 433 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 434 }
99df837e 435 pmap_init_thread(td);
5d21b981
MD
436 /*
437 * Normally initializing a thread for a remote cpu requires sending an
438 * IPI. However, the idlethread is setup before the other cpus are
439 * activated so we have to treat it as a special case. XXX manipulation
440 * of gd_tdallq requires the BGL.
441 */
442 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 443 crit_enter_gd(mygd);
75cdbe6c 444 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 445 crit_exit_gd(mygd);
75cdbe6c 446 } else {
2db3b277 447 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 448 }
3573cf7b 449 dsched_enter_thread(td);
73e4f7b9
MD
450}
451
452void
453lwkt_set_comm(thread_t td, const char *ctl, ...)
454{
e2565a42 455 __va_list va;
73e4f7b9 456
e2565a42 457 __va_start(va, ctl);
379210cb 458 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 459 __va_end(va);
5bf48697 460 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
461}
462
eb2adbf5
MD
463/*
464 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
465 * this does not prevent the thread from migrating to another cpu so the
466 * gd_tdallq state is not protected by this.
467 */
99df837e 468void
73e4f7b9 469lwkt_hold(thread_t td)
99df837e 470{
74c9628e 471 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
472}
473
474void
475lwkt_rele(thread_t td)
476{
477 KKASSERT(td->td_refs > 0);
74c9628e 478 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
479}
480
73e4f7b9
MD
481void
482lwkt_free_thread(thread_t td)
483{
74c9628e 484 KKASSERT(td->td_refs == 0);
c17a6852
MD
485 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
486 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
487 if (td->td_flags & TDF_ALLOCATED_THREAD) {
488 objcache_put(thread_cache, td);
489 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
490 /* client-allocated struct with internally allocated stack */
491 KASSERT(td->td_kstack && td->td_kstack_size > 0,
492 ("lwkt_free_thread: corrupted stack"));
493 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
494 td->td_kstack = NULL;
495 td->td_kstack_size = 0;
99df837e 496 }
a86ce0cd 497
e7c0dbba 498 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
499}
500
501
8ad65e08
MD
502/*
503 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
504 * switch to the idlethread. Switching must occur within a critical
505 * section to avoid races with the scheduling queue.
506 *
507 * We always have full control over our cpu's run queue. Other cpus
508 * that wish to manipulate our queue must use the cpu_*msg() calls to
509 * talk to our cpu, so a critical section is all that is needed and
510 * the result is very, very fast thread switching.
511 *
96728c05
MD
512 * The LWKT scheduler uses a fixed priority model and round-robins at
513 * each priority level. User process scheduling is a totally
514 * different beast and LWKT priorities should not be confused with
515 * user process priorities.
f1d1c3fa 516 *
69d78e99
MD
517 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
518 * is not called by the current thread in the preemption case, only when
519 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
520 *
521 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
522 * migration and tsleep deschedule the current lwkt thread and call
523 * lwkt_switch(). In particular, the target cpu of the migration fully
524 * expects the thread to become non-runnable and can deadlock against
525 * cpusync operations if we run any IPIs prior to switching the thread out.
526 *
527 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 528 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
529 */
530void
531lwkt_switch(void)
532{
37af14fe
MD
533 globaldata_t gd = mycpu;
534 thread_t td = gd->gd_curthread;
8ad65e08 535 thread_t ntd;
5411d8f1 536 int upri;
8ad65e08 537
da0b0e8b 538 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 539 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 540
46a3f46d 541 /*
27e88a6e
MD
542 * Switching from within a 'fast' (non thread switched) interrupt or IPI
543 * is illegal. However, we may have to do it anyway if we hit a fatal
544 * kernel trap or we have paniced.
545 *
546 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 547 */
27e88a6e
MD
548 if (gd->gd_intr_nesting_level) {
549 int savegdnest;
550 int savegdtrap;
551
5fddbda2 552 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 553 panic("lwkt_switch: Attempt to switch from a "
5a8df152 554 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
555 "td %p\n",
556 td);
27e88a6e
MD
557 } else {
558 savegdnest = gd->gd_intr_nesting_level;
559 savegdtrap = gd->gd_trap_nesting_level;
560 gd->gd_intr_nesting_level = 0;
561 gd->gd_trap_nesting_level = 0;
a7422615
MD
562 if ((td->td_flags & TDF_PANICWARN) == 0) {
563 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
564 kprintf("Warning: thread switch from interrupt, IPI, "
565 "or hard code section.\n"
a7422615 566 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 567 print_backtrace(-1);
a7422615 568 }
27e88a6e
MD
569 lwkt_switch();
570 gd->gd_intr_nesting_level = savegdnest;
571 gd->gd_trap_nesting_level = savegdtrap;
572 return;
573 }
96728c05 574 }
ef0fdad1 575
cb973d15 576 /*
85946b6c
MD
577 * Release our current user process designation if we are blocking
578 * or if a user reschedule was requested.
579 *
580 * NOTE: This function is NOT called if we are switching into or
581 * returning from a preemption.
582 *
583 * NOTE: Releasing our current user process designation may cause
584 * it to be assigned to another thread, which in turn will
585 * cause us to block in the usched acquire code when we attempt
586 * to return to userland.
587 *
588 * NOTE: On SMP systems this can be very nasty when heavy token
589 * contention is present so we want to be careful not to
590 * release the designation gratuitously.
cb973d15 591 */
85946b6c
MD
592 if (td->td_release &&
593 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 594 td->td_release(td);
85946b6c 595 }
cb973d15 596
85946b6c 597 /*
5411d8f1
MD
598 * Release all tokens. Once we do this we must remain in the critical
599 * section and cannot run IPIs or other interrupts until we switch away
600 * because they may implode if they try to get a token using our thread
601 * context.
85946b6c 602 */
37af14fe 603 crit_enter_gd(gd);
3b998fa9 604 if (TD_TOKS_HELD(td))
9d265729
MD
605 lwkt_relalltokens(td);
606
607 /*
b02926de
MD
608 * We had better not be holding any spin locks, but don't get into an
609 * endless panic loop.
9d265729 610 */
0846e4ce 611 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 612 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 613 gd->gd_spinlocks));
9d265729 614
0f7a3396
MD
615#ifdef INVARIANTS
616 if (td->td_cscount) {
6ea70f76 617 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
618 td);
619 if (panic_on_cscount)
620 panic("switching while mastering cpusync");
621 }
8a8d5d85 622#endif
f9235b6d
MD
623
624 /*
625 * If we had preempted another thread on this cpu, resume the preempted
626 * thread. This occurs transparently, whether the preempted thread
627 * was scheduled or not (it may have been preempted after descheduling
628 * itself).
629 *
630 * We have to setup the MP lock for the original thread after backing
631 * out the adjustment that was made to curthread when the original
632 * was preempted.
633 */
99df837e 634 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
635 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
636 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
637
638 /*
b9eb1c19
MD
639 * The interrupt may have woken a thread up, we need to properly
640 * set the reschedule flag if the originally interrupted thread is
641 * at a lower priority.
85946b6c
MD
642 *
643 * The interrupt may not have descheduled.
8ec60c3f 644 */
85946b6c 645 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
8ec60c3f 646 need_lwkt_resched();
f9235b6d
MD
647 goto havethread_preempted;
648 }
649
b12defdc 650 /*
5411d8f1
MD
651 * Figure out switch target. If we cannot switch to our desired target
652 * look for a thread that we can switch to.
cfaeae2a 653 *
5411d8f1
MD
654 * NOTE! The limited spin loop and related parameters are extremely
655 * important for system performance, particularly for pipes and
656 * concurrent conflicting VM faults.
f9235b6d 657 */
5411d8f1
MD
658 clear_lwkt_resched();
659 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
660
661 if (ntd) {
662 do {
663 if (TD_TOKS_NOT_HELD(ntd) ||
664 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
665 {
666 goto havethread;
667 }
668 ++gd->gd_cnt.v_lock_colls;
669 ++ntd->td_contended;
670 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
671 upri = ntd->td_upri;
f9235b6d 672
f9235b6d 673 /*
5411d8f1
MD
674 * Bleh, the thread we wanted to switch to has a contended token.
675 * See if we can switch to another thread.
2a418930 676 *
5411d8f1
MD
677 * We generally don't want to do this because it represents a
678 * priority inversion. Do not allow the case if the thread
679 * is returning to userland (not a kernel thread) AND the thread
680 * has a lower upri.
f9235b6d 681 */
b12defdc 682 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5411d8f1
MD
683 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri)
684 break;
685 upri = ntd->td_upri;
b12defdc 686
5411d8f1
MD
687 /*
688 * Try this one.
689 */
690 if (TD_TOKS_NOT_HELD(ntd) ||
691 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
692 goto havethread;
693 }
694 ++ntd->td_contended;
695 ++gd->gd_cnt.v_lock_colls;
b12defdc 696 }
b12defdc
MD
697
698 /*
5411d8f1
MD
699 * Fall through, switch to idle thread to get us out of the current
700 * context. Since we were contended, prevent HLT by flagging a
701 * LWKT reschedule.
b12defdc 702 */
5411d8f1 703 need_lwkt_resched();
f1d1c3fa 704 }
8a8d5d85 705
5411d8f1
MD
706 /*
707 * We either contended on ntd or the runq is empty. We must switch
708 * through the idle thread to get out of the current context.
709 */
710 ntd = &gd->gd_idlethread;
711 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
712 ASSERT_NO_TOKENS_HELD(ntd);
713 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
714 goto haveidle;
715
2a418930 716havethread:
b12defdc 717 /*
be71787b
MD
718 * Clear gd_idle_repeat when doing a normal switch to a non-idle
719 * thread.
f9235b6d 720 */
9ac1ee6e 721 ntd->td_wmesg = NULL;
844cfcd6 722 ntd->td_contended = 0;
b12defdc 723 ++gd->gd_cnt.v_swtch;
be71787b 724 gd->gd_idle_repeat = 0;
2a418930 725
f9235b6d 726havethread_preempted:
f9235b6d
MD
727 /*
728 * If the new target does not need the MP lock and we are holding it,
729 * release the MP lock. If the new target requires the MP lock we have
730 * already acquired it for the target.
8a8d5d85 731 */
2a418930 732 ;
f9235b6d
MD
733haveidle:
734 KASSERT(ntd->td_critcount,
b5d16701
MD
735 ("priority problem in lwkt_switch %d %d",
736 td->td_critcount, ntd->td_critcount));
737
94f6d86e 738 if (td != ntd) {
cc9b6223
MD
739 /*
740 * Execute the actual thread switch operation. This function
741 * returns to the current thread and returns the previous thread
742 * (which may be different from the thread we switched to).
743 *
744 * We are responsible for marking ntd as TDF_RUNNING.
745 */
121f93bc 746 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
94f6d86e 747 ++switch_count;
a1f0fb66 748 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
749 ntd->td_flags |= TDF_RUNNING;
750 lwkt_switch_return(td->td_switch(ntd));
751 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 752 }
b12defdc 753
b12defdc 754 /*
54341a3b 755 * catch-all. XXX is this strictly needed?
b12defdc
MD
756 */
757 splz_check();
54341a3b 758
37af14fe
MD
759 /* NOTE: current cpu may have changed after switch */
760 crit_exit_quick(td);
8ad65e08
MD
761}
762
cc9b6223
MD
763/*
764 * Called by assembly in the td_switch (thread restore path) for thread
765 * bootstrap cases which do not 'return' to lwkt_switch().
766 */
767void
768lwkt_switch_return(thread_t otd)
769{
cc9b6223
MD
770 globaldata_t rgd;
771
772 /*
773 * Check if otd was migrating. Now that we are on ntd we can finish
774 * up the migration. This is a bit messy but it is the only place
775 * where td is known to be fully descheduled.
776 *
777 * We can only activate the migration if otd was migrating but not
778 * held on the cpu due to a preemption chain. We still have to
779 * clear TDF_RUNNING on the old thread either way.
780 *
781 * We are responsible for clearing the previously running thread's
782 * TDF_RUNNING.
783 */
784 if ((rgd = otd->td_migrate_gd) != NULL &&
785 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
786 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
787 (TDF_MIGRATING | TDF_RUNNING));
788 otd->td_migrate_gd = NULL;
789 otd->td_flags &= ~TDF_RUNNING;
790 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
791 } else {
792 otd->td_flags &= ~TDF_RUNNING;
793 }
2b07d9aa
MD
794
795 /*
796 * Final exit validations (see lwp_wait()). Note that otd becomes
797 * invalid the *instant* we set TDF_MP_EXITSIG.
798 */
799 while (otd->td_flags & TDF_EXITING) {
800 u_int mpflags;
801
802 mpflags = otd->td_mpflags;
803 cpu_ccfence();
804
805 if (mpflags & TDF_MP_EXITWAIT) {
806 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
807 mpflags | TDF_MP_EXITSIG)) {
808 wakeup(otd);
809 break;
810 }
811 } else {
812 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
813 mpflags | TDF_MP_EXITSIG)) {
814 wakeup(otd);
815 break;
816 }
817 }
818 }
cc9b6223
MD
819}
820
b68b7282 821/*
96728c05 822 * Request that the target thread preempt the current thread. Preemption
54341a3b
MD
823 * can only occur if our only critical section is the one that we were called
824 * with, the relative priority of the target thread is higher, and the target
825 * thread holds no tokens. This also only works if we are not holding any
826 * spinlocks (obviously).
96728c05
MD
827 *
828 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
829 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 830 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
831 * to determine whether preemption is possible (aka usually just the crit
832 * priority of lwkt_schedule() itself).
b68b7282 833 *
54341a3b
MD
834 * Preemption is typically limited to interrupt threads.
835 *
836 * Operation works in a fairly straight-forward manner. The normal
837 * scheduling code is bypassed and we switch directly to the target
838 * thread. When the target thread attempts to block or switch away
839 * code at the base of lwkt_switch() will switch directly back to our
840 * thread. Our thread is able to retain whatever tokens it holds and
841 * if the target needs one of them the target will switch back to us
842 * and reschedule itself normally.
b68b7282
MD
843 */
844void
f9235b6d 845lwkt_preempt(thread_t ntd, int critcount)
b68b7282 846{
46a3f46d 847 struct globaldata *gd = mycpu;
cc9b6223 848 thread_t xtd;
0a3f9b47 849 thread_t td;
2d910aaf 850 int save_gd_intr_nesting_level;
b68b7282 851
26a0694b 852 /*
96728c05
MD
853 * The caller has put us in a critical section. We can only preempt
854 * if the caller of the caller was not in a critical section (basically
f9235b6d 855 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 856 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 857 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
858 *
859 * YYY The target thread must be in a critical section (else it must
860 * inherit our critical section? I dunno yet).
26a0694b 861 */
f9235b6d 862 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 863
b12defdc 864 td = gd->gd_curthread;
fbc024e4
MD
865 if (preempt_enable == 0) {
866 ++preempt_miss;
867 return;
868 }
f9235b6d 869 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
870 ++preempt_miss;
871 return;
872 }
f9235b6d 873 if (td->td_critcount > critcount) {
96728c05
MD
874 ++preempt_miss;
875 return;
876 }
121f93bc
MD
877 if (td->td_cscount) {
878 ++preempt_miss;
879 return;
880 }
46a3f46d 881 if (ntd->td_gd != gd) {
96728c05
MD
882 ++preempt_miss;
883 return;
884 }
41a01a4d 885 /*
77912481
MD
886 * We don't have to check spinlocks here as they will also bump
887 * td_critcount.
d3d1cbc8
MD
888 *
889 * Do not try to preempt if the target thread is holding any tokens.
890 * We could try to acquire the tokens but this case is so rare there
891 * is no need to support it.
41a01a4d 892 */
0846e4ce 893 KKASSERT(gd->gd_spinlocks == 0);
77912481 894
3b998fa9 895 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8 896 ++preempt_miss;
d3d1cbc8
MD
897 return;
898 }
26a0694b
MD
899 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
900 ++preempt_weird;
901 return;
902 }
903 if (ntd->td_preempted) {
4b5f931b 904 ++preempt_hit;
26a0694b 905 return;
b68b7282 906 }
da0b0e8b 907 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 908
8ec60c3f
MD
909 /*
910 * Since we are able to preempt the current thread, there is no need to
911 * call need_lwkt_resched().
2d910aaf
MD
912 *
913 * We must temporarily clear gd_intr_nesting_level around the switch
914 * since switchouts from the target thread are allowed (they will just
915 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
916 *
917 * A preemption must switch back to the original thread, assert the
918 * case.
8ec60c3f 919 */
26a0694b
MD
920 ++preempt_hit;
921 ntd->td_preempted = td;
922 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 923 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
924 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
925 gd->gd_intr_nesting_level = 0;
121f93bc
MD
926
927 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
928 ntd->td_flags |= TDF_RUNNING;
929 xtd = td->td_switch(ntd);
930 KKASSERT(xtd == ntd);
931 lwkt_switch_return(xtd);
2d910aaf 932 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 933
26a0694b
MD
934 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
935 ntd->td_preempted = NULL;
936 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
937}
938
f1d1c3fa 939/*
faaeffac 940 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
941 * This will work inside a critical section but not inside a hard code
942 * section.
ef0fdad1 943 *
f1d1c3fa
MD
944 * (self contained on a per cpu basis)
945 */
946void
faaeffac 947splz_check(void)
f1d1c3fa 948{
7966cb69
MD
949 globaldata_t gd = mycpu;
950 thread_t td = gd->gd_curthread;
ef0fdad1 951
4a28fe22
MD
952 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
953 gd->gd_intr_nesting_level == 0 &&
954 td->td_nest_count < 2)
955 {
f1d1c3fa 956 splz();
4a28fe22
MD
957 }
958}
959
960/*
961 * This version is integrated into crit_exit, reqflags has already
962 * been tested but td_critcount has not.
963 *
964 * We only want to execute the splz() on the 1->0 transition of
965 * critcount and not in a hard code section or if too deeply nested.
925040f2 966 *
0846e4ce 967 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
968 */
969void
970lwkt_maybe_splz(thread_t td)
971{
972 globaldata_t gd = td->td_gd;
973
974 if (td->td_critcount == 0 &&
975 gd->gd_intr_nesting_level == 0 &&
976 td->td_nest_count < 2)
977 {
978 splz();
979 }
f1d1c3fa
MD
980}
981
e6546af9
MD
982/*
983 * Drivers which set up processing co-threads can call this function to
984 * run the co-thread at a higher priority and to allow it to preempt
985 * normal threads.
986 */
987void
988lwkt_set_interrupt_support_thread(void)
989{
990 thread_t td = curthread;
991
992 lwkt_setpri_self(TDPRI_INT_SUPPORT);
993 td->td_flags |= TDF_INTTHREAD;
994 td->td_preemptable = lwkt_preempt;
995}
996
997
8ad65e08 998/*
f9235b6d
MD
999 * This function is used to negotiate a passive release of the current
1000 * process/lwp designation with the user scheduler, allowing the user
1001 * scheduler to schedule another user thread. The related kernel thread
1002 * (curthread) continues running in the released state.
8ad65e08
MD
1003 */
1004void
f9235b6d 1005lwkt_passive_release(struct thread *td)
8ad65e08 1006{
f9235b6d
MD
1007 struct lwp *lp = td->td_lwp;
1008
1009 td->td_release = NULL;
1010 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1011
f9235b6d 1012 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1013}
1014
f9235b6d 1015
3824f392 1016/*
d2d8515b
MD
1017 * This implements a LWKT yield, allowing a kernel thread to yield to other
1018 * kernel threads at the same or higher priority. This function can be
1019 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1020 *
d2d8515b
MD
1021 * Most kernel threads run at the same priority in order to allow equal
1022 * sharing.
f9235b6d
MD
1023 *
1024 * (self contained on a per cpu basis)
3824f392
MD
1025 */
1026void
f9235b6d 1027lwkt_yield(void)
3824f392 1028{
f9235b6d
MD
1029 globaldata_t gd = mycpu;
1030 thread_t td = gd->gd_curthread;
3824f392 1031
f9235b6d
MD
1032 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1033 splz();
85946b6c 1034 if (lwkt_resched_wanted()) {
f9235b6d
MD
1035 lwkt_schedule_self(curthread);
1036 lwkt_switch();
f9235b6d 1037 }
3824f392
MD
1038}
1039
40504122
MD
1040/*
1041 * The quick version processes pending interrupts and higher-priority
1042 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1043 *
1044 * When called while attempting to return to userland the only same-pri
1045 * threads are the ones which have already tried to become the current
1046 * user process.
40504122
MD
1047 */
1048void
1049lwkt_yield_quick(void)
1050{
1051 globaldata_t gd = mycpu;
1052 thread_t td = gd->gd_curthread;
1053
1054 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1055 splz();
1056 if (lwkt_resched_wanted()) {
9c99cb33 1057 crit_enter();
40504122
MD
1058 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1059 clear_lwkt_resched();
1060 } else {
1061 lwkt_schedule_self(curthread);
1062 lwkt_switch();
1063 }
9c99cb33 1064 crit_exit();
40504122
MD
1065 }
1066}
1067
3824f392 1068/*
f9235b6d
MD
1069 * This yield is designed for kernel threads with a user context.
1070 *
1071 * The kernel acting on behalf of the user is potentially cpu-bound,
1072 * this function will efficiently allow other threads to run and also
1073 * switch to other processes by releasing.
3824f392
MD
1074 *
1075 * The lwkt_user_yield() function is designed to have very low overhead
1076 * if no yield is determined to be needed.
1077 */
1078void
1079lwkt_user_yield(void)
1080{
f9235b6d
MD
1081 globaldata_t gd = mycpu;
1082 thread_t td = gd->gd_curthread;
1083
1084 /*
1085 * Always run any pending interrupts in case we are in a critical
1086 * section.
1087 */
1088 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1089 splz();
3824f392 1090
3824f392 1091 /*
f9235b6d
MD
1092 * Switch (which forces a release) if another kernel thread needs
1093 * the cpu, if userland wants us to resched, or if our kernel
1094 * quantum has run out.
3824f392 1095 */
f9235b6d 1096 if (lwkt_resched_wanted() ||
85946b6c 1097 user_resched_wanted())
f9235b6d 1098 {
3824f392 1099 lwkt_switch();
3824f392
MD
1100 }
1101
f9235b6d 1102#if 0
3824f392 1103 /*
f9235b6d
MD
1104 * Reacquire the current process if we are released.
1105 *
1106 * XXX not implemented atm. The kernel may be holding locks and such,
1107 * so we want the thread to continue to receive cpu.
3824f392 1108 */
f9235b6d
MD
1109 if (td->td_release == NULL && lp) {
1110 lp->lwp_proc->p_usched->acquire_curproc(lp);
1111 td->td_release = lwkt_passive_release;
1112 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1113 }
f9235b6d 1114#endif
b9eb1c19
MD
1115}
1116
8ad65e08 1117/*
f1d1c3fa
MD
1118 * Generic schedule. Possibly schedule threads belonging to other cpus and
1119 * deal with threads that might be blocked on a wait queue.
1120 *
0a3f9b47
MD
1121 * We have a little helper inline function which does additional work after
1122 * the thread has been enqueued, including dealing with preemption and
1123 * setting need_lwkt_resched() (which prevents the kernel from returning
1124 * to userland until it has processed higher priority threads).
6330a558
MD
1125 *
1126 * It is possible for this routine to be called after a failed _enqueue
1127 * (due to the target thread migrating, sleeping, or otherwise blocked).
1128 * We have to check that the thread is actually on the run queue!
8ad65e08 1129 */
0a3f9b47
MD
1130static __inline
1131void
85946b6c 1132_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1133{
6330a558 1134 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1135 if (ntd->td_preemptable) {
f9235b6d 1136 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1137 }
0a3f9b47
MD
1138 }
1139}
1140
361d01dd 1141static __inline
8ad65e08 1142void
85946b6c 1143_lwkt_schedule(thread_t td)
8ad65e08 1144{
37af14fe
MD
1145 globaldata_t mygd = mycpu;
1146
cf709dd2
MD
1147 KASSERT(td != &td->td_gd->gd_idlethread,
1148 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1149 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1150 crit_enter_gd(mygd);
4643740a
MD
1151 KKASSERT(td->td_lwp == NULL ||
1152 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1153
37af14fe 1154 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1155 _lwkt_enqueue(td);
1156 } else {
f1d1c3fa 1157 /*
7cd8d145
MD
1158 * If we own the thread, there is no race (since we are in a
1159 * critical section). If we do not own the thread there might
1160 * be a race but the target cpu will deal with it.
f1d1c3fa 1161 */
7cd8d145 1162 if (td->td_gd == mygd) {
9d265729 1163 _lwkt_enqueue(td);
85946b6c 1164 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1165 } else {
e381e77c 1166 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1167 }
8ad65e08 1168 }
37af14fe 1169 crit_exit_gd(mygd);
8ad65e08
MD
1170}
1171
361d01dd
MD
1172void
1173lwkt_schedule(thread_t td)
1174{
85946b6c 1175 _lwkt_schedule(td);
361d01dd
MD
1176}
1177
1178void
85946b6c 1179lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1180{
85946b6c 1181 _lwkt_schedule(td);
361d01dd
MD
1182}
1183
e381e77c
MD
1184/*
1185 * When scheduled remotely if frame != NULL the IPIQ is being
1186 * run via doreti or an interrupt then preemption can be allowed.
1187 *
1188 * To allow preemption we have to drop the critical section so only
1189 * one is present in _lwkt_schedule_post.
1190 */
1191static void
1192lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1193{
1194 thread_t td = curthread;
1195 thread_t ntd = arg;
1196
1197 if (frame && ntd->td_preemptable) {
1198 crit_exit_noyield(td);
85946b6c 1199 _lwkt_schedule(ntd);
e381e77c
MD
1200 crit_enter_quick(td);
1201 } else {
85946b6c 1202 _lwkt_schedule(ntd);
e381e77c
MD
1203 }
1204}
1205
d9eea1a5 1206/*
52eedfb5
MD
1207 * Thread migration using a 'Pull' method. The thread may or may not be
1208 * the current thread. It MUST be descheduled and in a stable state.
1209 * lwkt_giveaway() must be called on the cpu owning the thread.
1210 *
1211 * At any point after lwkt_giveaway() is called, the target cpu may
1212 * 'pull' the thread by calling lwkt_acquire().
1213 *
ae8e83e6
MD
1214 * We have to make sure the thread is not sitting on a per-cpu tsleep
1215 * queue or it will blow up when it moves to another cpu.
1216 *
52eedfb5 1217 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1218 */
52eedfb5
MD
1219void
1220lwkt_giveaway(thread_t td)
1221{
3b4192fb 1222 globaldata_t gd = mycpu;
52eedfb5 1223
3b4192fb
MD
1224 crit_enter_gd(gd);
1225 if (td->td_flags & TDF_TSLEEPQ)
1226 tsleep_remove(td);
1227 KKASSERT(td->td_gd == gd);
1228 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1229 td->td_flags |= TDF_MIGRATING;
1230 crit_exit_gd(gd);
52eedfb5
MD
1231}
1232
a2a5ad0d
MD
1233void
1234lwkt_acquire(thread_t td)
1235{
37af14fe
MD
1236 globaldata_t gd;
1237 globaldata_t mygd;
cc9b6223 1238 int retry = 10000000;
a2a5ad0d 1239
52eedfb5 1240 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1241 gd = td->td_gd;
37af14fe 1242 mygd = mycpu;
52eedfb5 1243 if (gd != mycpu) {
35238fa5 1244 cpu_lfence();
52eedfb5 1245 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1246 crit_enter_gd(mygd);
cfaeae2a 1247 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1248 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1249 lwkt_process_ipiq();
52eedfb5 1250 cpu_lfence();
cc9b6223
MD
1251 if (--retry == 0) {
1252 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1253 td, td->td_flags);
1254 retry = 10000000;
1255 }
a86ce0cd
MD
1256#ifdef _KERNEL_VIRTUAL
1257 pthread_yield();
1258#endif
df910c23 1259 }
cfaeae2a 1260 DEBUG_POP_INFO();
562273ea 1261 cpu_mfence();
37af14fe 1262 td->td_gd = mygd;
52eedfb5
MD
1263 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1264 td->td_flags &= ~TDF_MIGRATING;
1265 crit_exit_gd(mygd);
1266 } else {
1267 crit_enter_gd(mygd);
1268 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1269 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1270 crit_exit_gd(mygd);
a2a5ad0d
MD
1271 }
1272}
1273
f1d1c3fa
MD
1274/*
1275 * Generic deschedule. Descheduling threads other then your own should be
1276 * done only in carefully controlled circumstances. Descheduling is
1277 * asynchronous.
1278 *
1279 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1280 */
1281void
1282lwkt_deschedule(thread_t td)
1283{
f1d1c3fa
MD
1284 crit_enter();
1285 if (td == curthread) {
1286 _lwkt_dequeue(td);
1287 } else {
a72187e9 1288 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1289 _lwkt_dequeue(td);
1290 } else {
b8a98473 1291 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1292 }
1293 }
1294 crit_exit();
1295}
1296
4b5f931b
MD
1297/*
1298 * Set the target thread's priority. This routine does not automatically
1299 * switch to a higher priority thread, LWKT threads are not designed for
1300 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1301 */
1302void
1303lwkt_setpri(thread_t td, int pri)
1304{
f9235b6d
MD
1305 if (td->td_pri != pri) {
1306 KKASSERT(pri >= 0);
1307 crit_enter();
1308 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1309 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1310 _lwkt_dequeue(td);
1311 td->td_pri = pri;
1312 _lwkt_enqueue(td);
1313 } else {
1314 td->td_pri = pri;
1315 }
1316 crit_exit();
26a0694b 1317 }
26a0694b
MD
1318}
1319
03bd0a5e
MD
1320/*
1321 * Set the initial priority for a thread prior to it being scheduled for
1322 * the first time. The thread MUST NOT be scheduled before or during
1323 * this call. The thread may be assigned to a cpu other then the current
1324 * cpu.
1325 *
1326 * Typically used after a thread has been created with TDF_STOPPREQ,
1327 * and before the thread is initially scheduled.
1328 */
1329void
1330lwkt_setpri_initial(thread_t td, int pri)
1331{
1332 KKASSERT(pri >= 0);
1333 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1334 td->td_pri = pri;
03bd0a5e
MD
1335}
1336
26a0694b
MD
1337void
1338lwkt_setpri_self(int pri)
1339{
1340 thread_t td = curthread;
1341
4b5f931b
MD
1342 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1343 crit_enter();
1344 if (td->td_flags & TDF_RUNQ) {
1345 _lwkt_dequeue(td);
f9235b6d 1346 td->td_pri = pri;
4b5f931b
MD
1347 _lwkt_enqueue(td);
1348 } else {
f9235b6d 1349 td->td_pri = pri;
4b5f931b
MD
1350 }
1351 crit_exit();
1352}
1353
f9235b6d 1354/*
85946b6c 1355 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1356 */
1357void
85946b6c 1358lwkt_schedulerclock(thread_t td)
f9235b6d 1359{
85946b6c
MD
1360 globaldata_t gd = td->td_gd;
1361 thread_t xtd;
2a418930 1362
85946b6c
MD
1363 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1364 /*
1365 * If the current thread is at the head of the runq shift it to the
1366 * end of any equal-priority threads and request a LWKT reschedule
1367 * if it moved.
d992c377
MD
1368 *
1369 * Ignore upri in this situation. There will only be one user thread
1370 * in user mode, all others will be user threads running in kernel
1371 * mode and we have to make sure they get some cpu.
85946b6c
MD
1372 */
1373 xtd = TAILQ_NEXT(td, td_threadq);
1374 if (xtd && xtd->td_pri == td->td_pri) {
1375 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1376 while (xtd && xtd->td_pri == td->td_pri)
1377 xtd = TAILQ_NEXT(xtd, td_threadq);
1378 if (xtd)
1379 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1380 else
1381 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1382 need_lwkt_resched();
f9235b6d 1383 }
85946b6c
MD
1384 } else {
1385 /*
1386 * If we scheduled a thread other than the one at the head of the
1387 * queue always request a reschedule every tick.
1388 */
1389 need_lwkt_resched();
f9235b6d
MD
1390 }
1391}
1392
5d21b981 1393/*
52eedfb5
MD
1394 * Migrate the current thread to the specified cpu.
1395 *
cc9b6223
MD
1396 * This is accomplished by descheduling ourselves from the current cpu
1397 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1398 * 'old' thread wants to migrate after it has been completely switched out
1399 * and will complete the migration.
1400 *
1401 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1402 *
1403 * We must be sure to release our current process designation (if a user
1404 * process) before clearing out any tsleepq we are on because the release
1405 * code may re-add us.
ae8e83e6
MD
1406 *
1407 * We must be sure to remove ourselves from the current cpu's tsleepq
1408 * before potentially moving to another queue. The thread can be on
1409 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1410 */
5d21b981
MD
1411
1412void
1413lwkt_setcpu_self(globaldata_t rgd)
1414{
5d21b981
MD
1415 thread_t td = curthread;
1416
1417 if (td->td_gd != rgd) {
1418 crit_enter_quick(td);
cc9b6223 1419
95858b91
MD
1420 if (td->td_release)
1421 td->td_release(td);
ae8e83e6 1422 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1423 tsleep_remove(td);
cc9b6223
MD
1424
1425 /*
1426 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1427 * trying to deschedule ourselves and switch away, then deschedule
1428 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1429 * call lwkt_switch() to complete the operation.
1430 */
5d21b981
MD
1431 td->td_flags |= TDF_MIGRATING;
1432 lwkt_deschedule_self(td);
52eedfb5 1433 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1434 td->td_migrate_gd = rgd;
5d21b981 1435 lwkt_switch();
cc9b6223
MD
1436
1437 /*
1438 * We are now on the target cpu
1439 */
1440 KKASSERT(rgd == mycpu);
52eedfb5 1441 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1442 crit_exit_quick(td);
1443 }
5d21b981
MD
1444}
1445
ecdefdda
MD
1446void
1447lwkt_migratecpu(int cpuid)
1448{
ecdefdda
MD
1449 globaldata_t rgd;
1450
1451 rgd = globaldata_find(cpuid);
1452 lwkt_setcpu_self(rgd);
ecdefdda
MD
1453}
1454
5d21b981
MD
1455/*
1456 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1457 * do not have to enter another one).
1458 *
1459 * The thread (td) has already been completely descheduled from the
1460 * originating cpu and we can simply assert the case. The thread is
1461 * assigned to the new cpu and enqueued.
5d21b981 1462 *
cc9b6223 1463 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1464 */
1465static void
1466lwkt_setcpu_remote(void *arg)
1467{
1468 thread_t td = arg;
1469 globaldata_t gd = mycpu;
1470
cc9b6223 1471 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1472 td->td_gd = gd;
562273ea 1473 cpu_mfence();
5d21b981 1474 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1475 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1476 KKASSERT(td->td_lwp == NULL ||
1477 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1478 _lwkt_enqueue(td);
1479}
1480
553ea3c8 1481struct lwp *
4b5f931b
MD
1482lwkt_preempted_proc(void)
1483{
73e4f7b9 1484 thread_t td = curthread;
4b5f931b
MD
1485 while (td->td_preempted)
1486 td = td->td_preempted;
553ea3c8 1487 return(td->td_lwp);
4b5f931b
MD
1488}
1489
99df837e
MD
1490/*
1491 * Create a kernel process/thread/whatever. It shares it's address space
1492 * with proc0 - ie: kernel only.
1493 *
d2d8515b
MD
1494 * If the cpu is not specified one will be selected. In the future
1495 * specifying a cpu of -1 will enable kernel thread migration between
1496 * cpus.
99df837e
MD
1497 */
1498int
c9e9fb21
MD
1499lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1500 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1501{
73e4f7b9 1502 thread_t td;
e2565a42 1503 __va_list ap;
99df837e 1504
d3d32139 1505 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1506 tdflags);
a2a5ad0d
MD
1507 if (tdp)
1508 *tdp = td;
709799ea 1509 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1510
1511 /*
1512 * Set up arg0 for 'ps' etc
1513 */
e2565a42 1514 __va_start(ap, fmt);
379210cb 1515 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1516 __va_end(ap);
99df837e
MD
1517
1518 /*
1519 * Schedule the thread to run
1520 */
4643740a
MD
1521 if (td->td_flags & TDF_NOSTART)
1522 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1523 else
4643740a 1524 lwkt_schedule(td);
99df837e
MD
1525 return 0;
1526}
1527
1528/*
1529 * Destroy an LWKT thread. Warning! This function is not called when
1530 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1531 * uses a different reaping mechanism.
1532 */
1533void
1534lwkt_exit(void)
1535{
1536 thread_t td = curthread;
c070746a 1537 thread_t std;
8826f33a 1538 globaldata_t gd;
99df837e 1539
2883d2d8
MD
1540 /*
1541 * Do any cleanup that might block here
1542 */
99df837e 1543 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1544 kprintf("kthread %p %s has exited\n", td, td->td_comm);
2883d2d8
MD
1545 biosched_done(td);
1546 dsched_exit_thread(td);
c070746a
MD
1547
1548 /*
1549 * Get us into a critical section to interlock gd_freetd and loop
1550 * until we can get it freed.
1551 *
1552 * We have to cache the current td in gd_freetd because objcache_put()ing
1553 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1554 *
1555 * We are the current thread so of course our own TDF_RUNNING bit will
1556 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1557 */
1558 gd = mycpu;
37af14fe 1559 crit_enter_quick(td);
2af9d75d
MD
1560 for (;;) {
1561 if (td->td_refs) {
1562 tsleep(td, 0, "tdreap", 1);
1563 continue;
1564 }
1565 if ((std = gd->gd_freetd) != NULL) {
1566 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1567 gd->gd_freetd = NULL;
1568 objcache_put(thread_cache, std);
1569 continue;
1570 }
1571 break;
c070746a 1572 }
3b4192fb
MD
1573
1574 /*
1575 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1576 * the last time. We cannot block after this point or we may end
1577 * up with a stale td on the tsleepq.
eb2adbf5
MD
1578 *
1579 * None of this may block, the critical section is the only thing
1580 * protecting tdallq and the only thing preventing new lwkt_hold()
1581 * thread refs now.
3b4192fb
MD
1582 */
1583 if (td->td_flags & TDF_TSLEEPQ)
1584 tsleep_remove(td);
37af14fe 1585 lwkt_deschedule_self(td);
e56e4dea 1586 lwkt_remove_tdallq(td);
74c9628e 1587 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1588
1589 /*
1590 * Final cleanup
1591 */
1592 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1593 if (td->td_flags & TDF_ALLOCATED_THREAD)
1594 gd->gd_freetd = td;
99df837e
MD
1595 cpu_thread_exit();
1596}
1597
e56e4dea
MD
1598void
1599lwkt_remove_tdallq(thread_t td)
1600{
1601 KKASSERT(td->td_gd == mycpu);
1602 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1603}
1604
9cf43f91
MD
1605/*
1606 * Code reduction and branch prediction improvements. Call/return
1607 * overhead on modern cpus often degenerates into 0 cycles due to
1608 * the cpu's branch prediction hardware and return pc cache. We
1609 * can take advantage of this by not inlining medium-complexity
1610 * functions and we can also reduce the branch prediction impact
1611 * by collapsing perfectly predictable branches into a single
1612 * procedure instead of duplicating it.
1613 *
1614 * Is any of this noticeable? Probably not, so I'll take the
1615 * smaller code size.
1616 */
1617void
b6468f56 1618crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1619{
b6468f56 1620 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1621}
1622
2d93b37a
MD
1623void
1624crit_panic(void)
1625{
1626 thread_t td = curthread;
850634cc 1627 int lcrit = td->td_critcount;
2d93b37a 1628
850634cc
AH
1629 td->td_critcount = 0;
1630 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1631 /* NOT REACHED */
2d93b37a
MD
1632}
1633
bd8015ca
MD
1634/*
1635 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1636 * process the IPIQ while stopped.
bd8015ca
MD
1637 *
1638 * If we are dumping also try to process any pending interrupts. This may
1639 * or may not work depending on the state of the cpu at the point it was
1640 * stopped.
1641 */
1642void
1643lwkt_smp_stopped(void)
1644{
1645 globaldata_t gd = mycpu;
1646
bd8015ca
MD
1647 if (dumping) {
1648 lwkt_process_ipiq();
b19f40a4 1649 --gd->gd_intr_nesting_level;
bd8015ca 1650 splz();
b19f40a4 1651 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1652 } else {
1653 lwkt_process_ipiq();
1654 }
bd8015ca 1655}