hammer2(8): Oops, unbreak build.
[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
dae65060 38 * to use a critical section to avoid problems. Foreign thread
75cdbe6c 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>
dae65060 48#include <sys/malloc.h>
8ad65e08 49#include <sys/queue.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
5b49787b 56#include <sys/indefinite.h>
9d265729
MD
57
58#include <sys/thread2.h>
59#include <sys/spinlock2.h>
5b49787b 60#include <sys/indefinite2.h>
f1d1c3fa 61
8c72e3d5
AH
62#include <sys/dsched.h>
63
7d0bac62
MD
64#include <vm/vm.h>
65#include <vm/vm_param.h>
66#include <vm/vm_kern.h>
67#include <vm/vm_object.h>
68#include <vm/vm_page.h>
69#include <vm/vm_map.h>
70#include <vm/vm_pager.h>
71#include <vm/vm_extern.h>
7d0bac62 72
99df837e 73#include <machine/stdarg.h>
96728c05 74#include <machine/smp.h>
3a06728e 75#include <machine/clock.h>
99df837e 76
3a06728e
MD
77#define LOOPMASK
78
d850923c
AE
79#if !defined(KTR_CTXSW)
80#define KTR_CTXSW KTR_ALL
81#endif
82KTR_INFO_MASTER(ctxsw);
5bf48697
AE
83KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
84KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
85KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
86KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
1541028a 87
40aaf5fc
NT
88static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
89
0f7a3396
MD
90#ifdef INVARIANTS
91static int panic_on_cscount = 0;
92#endif
4b2467b7 93#ifdef DEBUG_LWKT_THREAD
e28c8ef4
SW
94static int64_t switch_count = 0;
95static int64_t preempt_hit = 0;
96static int64_t preempt_miss = 0;
97static int64_t preempt_weird = 0;
4b2467b7 98#endif
fb0f29c4 99static int lwkt_use_spin_port;
4b2467b7 100__read_mostly static struct objcache *thread_cache;
a46b4a23 101int cpu_mwait_spin = 0;
05220613 102
e381e77c 103static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 104static void lwkt_setcpu_remote(void *arg);
e381e77c 105
fb0f29c4
MD
106/*
107 * We can make all thread ports use the spin backend instead of the thread
108 * backend. This should only be set to debug the spin backend.
109 */
110TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
111
0f7a3396 112#ifdef INVARIANTS
0c52fa62
SG
113SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
114 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 115#endif
4b2467b7 116#ifdef DEBUG_LWKT_THREAD
0c52fa62
SG
117SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
118 "Number of switched threads");
dae65060 119SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 120 "Successful preemption events");
dae65060 121SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
122 "Failed preemption events");
123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
124 "Number of preempted threads.");
4b2467b7 125#endif
b12defdc
MD
126extern int lwkt_sched_debug;
127int lwkt_sched_debug = 0;
128SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
129 &lwkt_sched_debug, 0, "Scheduler debug");
4b2467b7 130__read_mostly static u_int lwkt_spin_loops = 10;
7fb451cb 131SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
b12defdc 132 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
4b2467b7 133__read_mostly static int preempt_enable = 1;
2a418930
MD
134SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
135 &preempt_enable, 0, "Enable preemption");
7b234d8c 136static int lwkt_cache_threads = 0;
765b1ae0
MD
137SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
138 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 139
4b5f931b
MD
140/*
141 * These helper procedures handle the runq, they can only be called from
142 * within a critical section.
75cdbe6c
MD
143 *
144 * WARNING! Prior to SMP being brought up it is possible to enqueue and
145 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
146 * instead of 'mycpu' when referencing the globaldata structure. Once
147 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 148 */
f1d1c3fa
MD
149static __inline
150void
151_lwkt_dequeue(thread_t td)
152{
153 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 154 struct globaldata *gd = td->td_gd;
4b5f931b 155
f1d1c3fa 156 td->td_flags &= ~TDF_RUNQ;
f9235b6d 157 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
de4d4cb0 158 --gd->gd_tdrunqcount;
f9235b6d 159 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 160 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
161 }
162}
163
f9235b6d
MD
164/*
165 * Priority enqueue.
166 *
d992c377
MD
167 * There are a limited number of lwkt threads runnable since user
168 * processes only schedule one at a time per cpu. However, there can
169 * be many user processes in kernel mode exiting from a tsleep() which
e3e6be1f 170 * become runnable.
d992c377 171 *
e6b81333
MD
172 * We scan the queue in both directions to help deal with degenerate
173 * situations when hundreds or thousands (or more) threads are runnable.
174 *
d992c377
MD
175 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
176 * will ignore user priority. This is to ensure that user threads in
177 * kernel mode get cpu at some point regardless of what the user
178 * scheduler thinks.
f9235b6d 179 */
f1d1c3fa
MD
180static __inline
181void
182_lwkt_enqueue(thread_t td)
183{
e6b81333
MD
184 thread_t xtd; /* forward scan */
185 thread_t rtd; /* reverse scan */
f9235b6d 186
7f5d7ed7 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 188 struct globaldata *gd = td->td_gd;
4b5f931b 189
f1d1c3fa 190 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
191 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
192 if (xtd == NULL) {
85946b6c
MD
193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d 195 } else {
e3e6be1f
MD
196 /*
197 * NOTE: td_upri - higher numbers more desireable, same sense
198 * as td_pri (typically reversed from lwp_upri).
199 *
200 * In the equal priority case we want the best selection
201 * at the beginning so the less desireable selections know
202 * that they have to setrunqueue/go-to-another-cpu, even
203 * though it means switching back to the 'best' selection.
204 * This also avoids degenerate situations when many threads
205 * are runnable or waking up at the same time.
206 *
207 * If upri matches exactly place at end/round-robin.
208 */
e6b81333
MD
209 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue);
210
d992c377 211 while (xtd &&
e6b81333 212 (xtd->td_pri > td->td_pri ||
d992c377 213 (xtd->td_pri == td->td_pri &&
e3e6be1f 214 xtd->td_upri >= td->td_upri))) {
85946b6c 215 xtd = TAILQ_NEXT(xtd, td_threadq);
e6b81333
MD
216
217 /*
218 * Doing a reverse scan at the same time is an optimization
219 * for the insert-closer-to-tail case that avoids having to
220 * scan the entire list. This situation can occur when
221 * thousands of threads are woken up at the same time.
222 */
223 if (rtd->td_pri > td->td_pri ||
224 (rtd->td_pri == td->td_pri &&
225 rtd->td_upri >= td->td_upri)) {
226 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq);
227 goto skip;
228 }
229 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq);
d992c377 230 }
85946b6c
MD
231 if (xtd)
232 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
233 else
234 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
f9235b6d 235 }
e6b81333 236skip:
de4d4cb0 237 ++gd->gd_tdrunqcount;
b12defdc
MD
238
239 /*
85946b6c 240 * Request a LWKT reschedule if we are now at the head of the queue.
b12defdc 241 */
85946b6c
MD
242 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
243 need_lwkt_resched();
f1d1c3fa
MD
244 }
245}
8ad65e08 246
e28c8ef4 247static boolean_t
40aaf5fc
NT
248_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
249{
250 struct thread *td = (struct thread *)obj;
251
252 td->td_kstack = NULL;
253 td->td_kstack_size = 0;
254 td->td_flags = TDF_ALLOCATED_THREAD;
4643740a 255 td->td_mpflags = 0;
40aaf5fc
NT
256 return (1);
257}
258
259static void
260_lwkt_thread_dtor(void *obj, void *privdata)
261{
262 struct thread *td = (struct thread *)obj;
263
264 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
265 ("_lwkt_thread_dtor: not allocated from objcache"));
266 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
267 td->td_kstack_size > 0,
268 ("_lwkt_thread_dtor: corrupted stack"));
269 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
7b234d8c
MD
270 td->td_kstack = NULL;
271 td->td_flags = 0;
40aaf5fc
NT
272}
273
274/*
275 * Initialize the lwkt s/system.
765b1ae0 276 *
7b234d8c
MD
277 * Nominally cache up to 32 thread + kstack structures. Cache more on
278 * systems with a lot of cpu cores.
40aaf5fc 279 */
ced589cb 280static void
40aaf5fc
NT
281lwkt_init(void)
282{
765b1ae0 283 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
7b234d8c
MD
284 if (lwkt_cache_threads == 0) {
285 lwkt_cache_threads = ncpus * 4;
286 if (lwkt_cache_threads < 32)
287 lwkt_cache_threads = 32;
288 }
765b1ae0
MD
289 thread_cache = objcache_create_mbacked(
290 M_THREAD, sizeof(struct thread),
2fce2579 291 0, lwkt_cache_threads,
765b1ae0 292 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc 293}
ced589cb 294SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
40aaf5fc 295
37af14fe
MD
296/*
297 * Schedule a thread to run. As the current thread we can always safely
298 * schedule ourselves, and a shortcut procedure is provided for that
299 * function.
300 *
301 * (non-blocking, self contained on a per cpu basis)
302 */
303void
304lwkt_schedule_self(thread_t td)
305{
cfaeae2a 306 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 307 crit_enter_quick(td);
f9235b6d
MD
308 KASSERT(td != &td->td_gd->gd_idlethread,
309 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
4643740a
MD
310 KKASSERT(td->td_lwp == NULL ||
311 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
37af14fe 312 _lwkt_enqueue(td);
37af14fe
MD
313 crit_exit_quick(td);
314}
315
316/*
317 * Deschedule a thread.
318 *
319 * (non-blocking, self contained on a per cpu basis)
320 */
321void
322lwkt_deschedule_self(thread_t td)
323{
324 crit_enter_quick(td);
37af14fe
MD
325 _lwkt_dequeue(td);
326 crit_exit_quick(td);
327}
328
8ad65e08
MD
329/*
330 * LWKTs operate on a per-cpu basis
331 *
73e4f7b9 332 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
333 */
334void
335lwkt_gdinit(struct globaldata *gd)
336{
f9235b6d 337 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 338 TAILQ_INIT(&gd->gd_tdallq);
e8e93b5f 339 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE);
8ad65e08
MD
340}
341
7d0bac62
MD
342/*
343 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
344 * or LWKT start address before it can be scheduled. If the target cpu is
345 * -1 the thread will be created on the current cpu.
0cfcada1
MD
346 *
347 * If you intend to create a thread without a process context this function
348 * does everything except load the startup and switcher function.
7d0bac62
MD
349 */
350thread_t
d3d32139 351lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 352{
d2d8515b 353 static int cpu_rotator;
c070746a 354 globaldata_t gd = mycpu;
99df837e 355 void *stack;
7d0bac62 356
c070746a
MD
357 /*
358 * If static thread storage is not supplied allocate a thread. Reuse
359 * a cached free thread if possible. gd_freetd is used to keep an exiting
360 * thread intact through the exit.
361 */
ef0fdad1 362 if (td == NULL) {
cf709dd2
MD
363 crit_enter_gd(gd);
364 if ((td = gd->gd_freetd) != NULL) {
365 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
366 TDF_RUNQ)) == 0);
c070746a 367 gd->gd_freetd = NULL;
cf709dd2 368 } else {
c070746a 369 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
370 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
371 TDF_RUNQ)) == 0);
372 }
373 crit_exit_gd(gd);
40aaf5fc 374 KASSERT((td->td_flags &
2af9d75d
MD
375 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
376 TDF_ALLOCATED_THREAD,
40aaf5fc
NT
377 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
378 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 379 }
c070746a
MD
380
381 /*
382 * Try to reuse cached stack.
383 */
f470d0c8
MD
384 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
385 if (flags & TDF_ALLOCATED_STACK) {
e4846942 386 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
387 stack = NULL;
388 }
389 }
390 if (stack == NULL) {
070a58b3
MD
391 if (cpu < 0)
392 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0);
393 else
394 stack = (void *)kmem_alloc_stack(&kernel_map, stksize,
395 KM_CPU(cpu));
ef0fdad1 396 flags |= TDF_ALLOCATED_STACK;
99df837e 397 }
d2d8515b
MD
398 if (cpu < 0) {
399 cpu = ++cpu_rotator;
400 cpu_ccfence();
5dd85b08 401 cpu = (uint32_t)cpu % (uint32_t)ncpus;
d2d8515b
MD
402 }
403 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 404 return(td);
7d0bac62
MD
405}
406
407/*
408 * Initialize a preexisting thread structure. This function is used by
409 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
410 *
f8c3996b
MD
411 * All threads start out in a critical section at a priority of
412 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
dae65060 413 * appropriate. This function may send an IPI message when the
75cdbe6c
MD
414 * requested cpu is not the current cpu and consequently gd_tdallq may
415 * not be initialized synchronously from the point of view of the originating
416 * cpu.
417 *
418 * NOTE! we have to be careful in regards to creating threads for other cpus
419 * if SMP has not yet been activated.
7d0bac62 420 */
75cdbe6c
MD
421static void
422lwkt_init_thread_remote(void *arg)
423{
424 thread_t td = arg;
425
52eedfb5
MD
426 /*
427 * Protected by critical section held by IPI dispatch
428 */
75cdbe6c
MD
429 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
430}
431
fdce8919
MD
432/*
433 * lwkt core thread structural initialization.
434 *
435 * NOTE: All threads are initialized as mpsafe threads.
436 */
7d0bac62 437void
f470d0c8
MD
438lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
439 struct globaldata *gd)
7d0bac62 440{
37af14fe
MD
441 globaldata_t mygd = mycpu;
442
99df837e
MD
443 bzero(td, sizeof(struct thread));
444 td->td_kstack = stack;
f470d0c8 445 td->td_kstack_size = stksize;
d3d32139 446 td->td_flags = flags;
4643740a 447 td->td_mpflags = 0;
f256b6c0 448 td->td_type = TD_TYPE_GENERIC;
26a0694b 449 td->td_gd = gd;
f9235b6d
MD
450 td->td_pri = TDPRI_KERN_DAEMON;
451 td->td_critcount = 1;
54341a3b 452 td->td_toks_have = NULL;
3b998fa9 453 td->td_toks_stop = &td->td_toks_base;
c068fb59
SZ
454 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
455 lwkt_initport_spin(&td->td_msgport, td,
456 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
457 } else {
fb0f29c4 458 lwkt_initport_thread(&td->td_msgport, td);
c068fb59 459 }
99df837e 460 pmap_init_thread(td);
5d21b981
MD
461 /*
462 * Normally initializing a thread for a remote cpu requires sending an
463 * IPI. However, the idlethread is setup before the other cpus are
464 * activated so we have to treat it as a special case. XXX manipulation
465 * of gd_tdallq requires the BGL.
466 */
467 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 468 crit_enter_gd(mygd);
75cdbe6c 469 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 470 crit_exit_gd(mygd);
75cdbe6c 471 } else {
2db3b277 472 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 473 }
3573cf7b 474 dsched_enter_thread(td);
73e4f7b9
MD
475}
476
477void
478lwkt_set_comm(thread_t td, const char *ctl, ...)
479{
e2565a42 480 __va_list va;
73e4f7b9 481
e2565a42 482 __va_start(va, ctl);
379210cb 483 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 484 __va_end(va);
5bf48697 485 KTR_LOG(ctxsw_newtd, td, td->td_comm);
7d0bac62
MD
486}
487
eb2adbf5
MD
488/*
489 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
490 * this does not prevent the thread from migrating to another cpu so the
491 * gd_tdallq state is not protected by this.
492 */
99df837e 493void
73e4f7b9 494lwkt_hold(thread_t td)
99df837e 495{
74c9628e 496 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
497}
498
499void
500lwkt_rele(thread_t td)
501{
502 KKASSERT(td->td_refs > 0);
74c9628e 503 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
504}
505
73e4f7b9
MD
506void
507lwkt_free_thread(thread_t td)
508{
74c9628e 509 KKASSERT(td->td_refs == 0);
c17a6852
MD
510 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
511 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
40aaf5fc
NT
512 if (td->td_flags & TDF_ALLOCATED_THREAD) {
513 objcache_put(thread_cache, td);
514 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
515 /* client-allocated struct with internally allocated stack */
516 KASSERT(td->td_kstack && td->td_kstack_size > 0,
517 ("lwkt_free_thread: corrupted stack"));
518 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
519 td->td_kstack = NULL;
520 td->td_kstack_size = 0;
99df837e 521 }
a86ce0cd 522
e7c0dbba 523 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
524}
525
526
8ad65e08 527/*
dae65060 528 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
529 * switch to the idlethread. Switching must occur within a critical
530 * section to avoid races with the scheduling queue.
531 *
532 * We always have full control over our cpu's run queue. Other cpus
533 * that wish to manipulate our queue must use the cpu_*msg() calls to
534 * talk to our cpu, so a critical section is all that is needed and
535 * the result is very, very fast thread switching.
536 *
96728c05
MD
537 * The LWKT scheduler uses a fixed priority model and round-robins at
538 * each priority level. User process scheduling is a totally
539 * different beast and LWKT priorities should not be confused with
540 * user process priorities.
f1d1c3fa 541 *
69d78e99
MD
542 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
543 * is not called by the current thread in the preemption case, only when
544 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
545 *
546 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
547 * migration and tsleep deschedule the current lwkt thread and call
548 * lwkt_switch(). In particular, the target cpu of the migration fully
549 * expects the thread to become non-runnable and can deadlock against
550 * cpusync operations if we run any IPIs prior to switching the thread out.
551 *
552 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 553 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
554 */
555void
556lwkt_switch(void)
557{
37af14fe
MD
558 globaldata_t gd = mycpu;
559 thread_t td = gd->gd_curthread;
8ad65e08 560 thread_t ntd;
c6a766f4 561 thread_t xtd;
5411d8f1 562 int upri;
3a06728e
MD
563#ifdef LOOPMASK
564 uint64_t tsc_base = rdtsc();
565#endif
8ad65e08 566
da0b0e8b 567 KKASSERT(gd->gd_processing_ipiq == 0);
121f93bc 568 KKASSERT(td->td_flags & TDF_RUNNING);
da0b0e8b 569
46a3f46d 570 /*
27e88a6e
MD
571 * Switching from within a 'fast' (non thread switched) interrupt or IPI
572 * is illegal. However, we may have to do it anyway if we hit a fatal
573 * kernel trap or we have paniced.
574 *
575 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 576 */
27e88a6e
MD
577 if (gd->gd_intr_nesting_level) {
578 int savegdnest;
579 int savegdtrap;
580
5fddbda2 581 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22 582 panic("lwkt_switch: Attempt to switch from a "
5a8df152 583 "fast interrupt, ipi, or hard code section, "
4a28fe22
MD
584 "td %p\n",
585 td);
27e88a6e
MD
586 } else {
587 savegdnest = gd->gd_intr_nesting_level;
588 savegdtrap = gd->gd_trap_nesting_level;
589 gd->gd_intr_nesting_level = 0;
590 gd->gd_trap_nesting_level = 0;
a7422615
MD
591 if ((td->td_flags & TDF_PANICWARN) == 0) {
592 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
593 kprintf("Warning: thread switch from interrupt, IPI, "
594 "or hard code section.\n"
a7422615 595 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 596 print_backtrace(-1);
a7422615 597 }
27e88a6e
MD
598 lwkt_switch();
599 gd->gd_intr_nesting_level = savegdnest;
600 gd->gd_trap_nesting_level = savegdtrap;
601 return;
602 }
96728c05 603 }
ef0fdad1 604
cb973d15 605 /*
85946b6c
MD
606 * Release our current user process designation if we are blocking
607 * or if a user reschedule was requested.
608 *
609 * NOTE: This function is NOT called if we are switching into or
610 * returning from a preemption.
611 *
612 * NOTE: Releasing our current user process designation may cause
613 * it to be assigned to another thread, which in turn will
614 * cause us to block in the usched acquire code when we attempt
615 * to return to userland.
616 *
617 * NOTE: On SMP systems this can be very nasty when heavy token
618 * contention is present so we want to be careful not to
619 * release the designation gratuitously.
cb973d15 620 */
85946b6c
MD
621 if (td->td_release &&
622 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
cb973d15 623 td->td_release(td);
85946b6c 624 }
cb973d15 625
85946b6c 626 /*
5411d8f1
MD
627 * Release all tokens. Once we do this we must remain in the critical
628 * section and cannot run IPIs or other interrupts until we switch away
629 * because they may implode if they try to get a token using our thread
630 * context.
85946b6c 631 */
37af14fe 632 crit_enter_gd(gd);
3b998fa9 633 if (TD_TOKS_HELD(td))
9d265729
MD
634 lwkt_relalltokens(td);
635
636 /*
b02926de
MD
637 * We had better not be holding any spin locks, but don't get into an
638 * endless panic loop.
9d265729 639 */
0846e4ce 640 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
d666840a 641 ("lwkt_switch: still holding %d exclusive spinlocks!",
0846e4ce 642 gd->gd_spinlocks));
9d265729 643
0f7a3396
MD
644#ifdef INVARIANTS
645 if (td->td_cscount) {
6ea70f76 646 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
647 td);
648 if (panic_on_cscount)
649 panic("switching while mastering cpusync");
650 }
8a8d5d85 651#endif
f9235b6d
MD
652
653 /*
654 * If we had preempted another thread on this cpu, resume the preempted
655 * thread. This occurs transparently, whether the preempted thread
656 * was scheduled or not (it may have been preempted after descheduling
657 * itself).
658 *
659 * We have to setup the MP lock for the original thread after backing
660 * out the adjustment that was made to curthread when the original
661 * was preempted.
662 */
99df837e 663 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
664 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
665 ntd->td_flags |= TDF_PREEMPT_DONE;
7fb451cb 666 ntd->td_contended = 0; /* reset contended */
8ec60c3f
MD
667
668 /*
b9eb1c19
MD
669 * The interrupt may have woken a thread up, we need to properly
670 * set the reschedule flag if the originally interrupted thread is
671 * at a lower priority.
85946b6c 672 *
c6a766f4
MD
673 * NOTE: The interrupt may not have descheduled ntd.
674 *
675 * NOTE: We do not reschedule if there are no threads on the runq.
676 * (ntd could be the idlethread).
8ec60c3f 677 */
c6a766f4
MD
678 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
679 if (xtd && xtd != ntd)
8ec60c3f 680 need_lwkt_resched();
f9235b6d
MD
681 goto havethread_preempted;
682 }
683
b12defdc 684 /*
5411d8f1
MD
685 * Figure out switch target. If we cannot switch to our desired target
686 * look for a thread that we can switch to.
cfaeae2a 687 *
5411d8f1
MD
688 * NOTE! The limited spin loop and related parameters are extremely
689 * important for system performance, particularly for pipes and
690 * concurrent conflicting VM faults.
f9235b6d 691 */
5411d8f1
MD
692 clear_lwkt_resched();
693 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
694
695 if (ntd) {
696 do {
697 if (TD_TOKS_NOT_HELD(ntd) ||
698 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
699 {
700 goto havethread;
701 }
7fb451cb 702 ++ntd->td_contended; /* overflow ok */
b1793cc6
MD
703 if (gd->gd_indefinite.type == 0)
704 indefinite_init(&gd->gd_indefinite, NULL, 0, 't');
3a06728e
MD
705#ifdef LOOPMASK
706 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
707 kprintf("lwkt_switch: excessive contended %d "
708 "thread %p\n", ntd->td_contended, ntd);
709 tsc_base = rdtsc();
710 }
711#endif
5411d8f1
MD
712 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
713 upri = ntd->td_upri;
f9235b6d 714
f9235b6d 715 /*
5411d8f1
MD
716 * Bleh, the thread we wanted to switch to has a contended token.
717 * See if we can switch to another thread.
2a418930 718 *
5411d8f1 719 * We generally don't want to do this because it represents a
5042f74d
MD
720 * priority inversion, but contending tokens on the same cpu can
721 * cause real problems if we don't now that we have an exclusive
722 * priority mechanism over shared for tokens.
723 *
724 * The solution is to allow threads with pending tokens to compete
725 * for them (a lower priority thread will get less cpu once it
726 * returns from the kernel anyway). If a thread does not have
727 * any contending tokens, we go by td_pri and upri.
f9235b6d 728 */
b12defdc 729 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
5042f74d
MD
730 if (TD_TOKS_NOT_HELD(ntd) &&
731 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) {
732 continue;
733 }
734 if (upri < ntd->td_upri)
735 upri = ntd->td_upri;
b12defdc 736
5411d8f1
MD
737 /*
738 * Try this one.
739 */
740 if (TD_TOKS_NOT_HELD(ntd) ||
741 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
742 goto havethread;
743 }
7fb451cb 744 ++ntd->td_contended; /* overflow ok */
b12defdc 745 }
b12defdc
MD
746
747 /*
5411d8f1
MD
748 * Fall through, switch to idle thread to get us out of the current
749 * context. Since we were contended, prevent HLT by flagging a
750 * LWKT reschedule.
b12defdc 751 */
5411d8f1 752 need_lwkt_resched();
f1d1c3fa 753 }
8a8d5d85 754
5411d8f1
MD
755 /*
756 * We either contended on ntd or the runq is empty. We must switch
757 * through the idle thread to get out of the current context.
758 */
759 ntd = &gd->gd_idlethread;
760 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
761 ASSERT_NO_TOKENS_HELD(ntd);
762 cpu_time.cp_msg[0] = 0;
5411d8f1
MD
763 goto haveidle;
764
2a418930 765havethread:
b12defdc 766 /*
be71787b
MD
767 * Clear gd_idle_repeat when doing a normal switch to a non-idle
768 * thread.
f9235b6d 769 */
9ac1ee6e 770 ntd->td_wmesg = NULL;
7fb451cb 771 ntd->td_contended = 0; /* reset once scheduled */
b12defdc 772 ++gd->gd_cnt.v_swtch;
be71787b 773 gd->gd_idle_repeat = 0;
2a418930 774
5b49787b
MD
775 /*
776 * If we were busy waiting record final disposition
777 */
b1793cc6
MD
778 if (gd->gd_indefinite.type)
779 indefinite_done(&gd->gd_indefinite);
5b49787b 780
f9235b6d 781havethread_preempted:
f9235b6d
MD
782 /*
783 * If the new target does not need the MP lock and we are holding it,
784 * release the MP lock. If the new target requires the MP lock we have
785 * already acquired it for the target.
8a8d5d85 786 */
2a418930 787 ;
f9235b6d
MD
788haveidle:
789 KASSERT(ntd->td_critcount,
b5d16701
MD
790 ("priority problem in lwkt_switch %d %d",
791 td->td_critcount, ntd->td_critcount));
792
94f6d86e 793 if (td != ntd) {
cc9b6223
MD
794 /*
795 * Execute the actual thread switch operation. This function
796 * returns to the current thread and returns the previous thread
797 * (which may be different from the thread we switched to).
798 *
799 * We are responsible for marking ntd as TDF_RUNNING.
800 */
121f93bc 801 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
4b2467b7 802#ifdef DEBUG_LWKT_THREAD
94f6d86e 803 ++switch_count;
4b2467b7 804#endif
a1f0fb66 805 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
806 ntd->td_flags |= TDF_RUNNING;
807 lwkt_switch_return(td->td_switch(ntd));
808 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 809 }
b12defdc 810
b12defdc 811 /*
54341a3b 812 * catch-all. XXX is this strictly needed?
b12defdc
MD
813 */
814 splz_check();
54341a3b 815
37af14fe
MD
816 /* NOTE: current cpu may have changed after switch */
817 crit_exit_quick(td);
8ad65e08
MD
818}
819
cc9b6223
MD
820/*
821 * Called by assembly in the td_switch (thread restore path) for thread
822 * bootstrap cases which do not 'return' to lwkt_switch().
823 */
824void
825lwkt_switch_return(thread_t otd)
826{
cc9b6223 827 globaldata_t rgd;
3a06728e
MD
828#ifdef LOOPMASK
829 uint64_t tsc_base = rdtsc();
830#endif
831 int exiting;
832
833 exiting = otd->td_flags & TDF_EXITING;
834 cpu_ccfence();
cc9b6223
MD
835
836 /*
837 * Check if otd was migrating. Now that we are on ntd we can finish
838 * up the migration. This is a bit messy but it is the only place
839 * where td is known to be fully descheduled.
840 *
841 * We can only activate the migration if otd was migrating but not
842 * held on the cpu due to a preemption chain. We still have to
843 * clear TDF_RUNNING on the old thread either way.
844 *
845 * We are responsible for clearing the previously running thread's
846 * TDF_RUNNING.
847 */
848 if ((rgd = otd->td_migrate_gd) != NULL &&
849 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
850 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
851 (TDF_MIGRATING | TDF_RUNNING));
852 otd->td_migrate_gd = NULL;
853 otd->td_flags &= ~TDF_RUNNING;
854 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
855 } else {
856 otd->td_flags &= ~TDF_RUNNING;
857 }
2b07d9aa
MD
858
859 /*
860 * Final exit validations (see lwp_wait()). Note that otd becomes
861 * invalid the *instant* we set TDF_MP_EXITSIG.
3a06728e
MD
862 *
863 * Use the EXITING status loaded from before we clear TDF_RUNNING,
864 * because if it is not set otd becomes invalid the instant we clear
865 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
866 * might 'steal' TDF_EXITING from another switch-return!).
2b07d9aa 867 */
3a06728e 868 while (exiting) {
2b07d9aa
MD
869 u_int mpflags;
870
871 mpflags = otd->td_mpflags;
872 cpu_ccfence();
873
874 if (mpflags & TDF_MP_EXITWAIT) {
875 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
876 mpflags | TDF_MP_EXITSIG)) {
877 wakeup(otd);
878 break;
879 }
880 } else {
881 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
882 mpflags | TDF_MP_EXITSIG)) {
883 wakeup(otd);
884 break;
885 }
886 }
3a06728e
MD
887
888#ifdef LOOPMASK
889 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
890 kprintf("lwkt_switch_return: excessive TDF_EXITING "
891 "thread %p\n", otd);
892 tsc_base = rdtsc();
893 }
894#endif
2b07d9aa 895 }
cc9b6223
MD
896}
897
b68b7282 898/*
96728c05 899 * Request that the target thread preempt the current thread. Preemption
203592a0
MD
900 * can only occur only:
901 *
902 * - If our critical section is the one that we were called with
903 * - The relative priority of the target thread is higher
904 * - The target is not excessively interrupt-nested via td_nest_count
905 * - The target thread holds no tokens.
906 * - The target thread is not already scheduled and belongs to the
907 * current cpu.
908 * - The current thread is not holding any spin-locks.
96728c05
MD
909 *
910 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
911 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 912 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
913 * to determine whether preemption is possible (aka usually just the crit
914 * priority of lwkt_schedule() itself).
b68b7282 915 *
54341a3b
MD
916 * Preemption is typically limited to interrupt threads.
917 *
918 * Operation works in a fairly straight-forward manner. The normal
919 * scheduling code is bypassed and we switch directly to the target
920 * thread. When the target thread attempts to block or switch away
921 * code at the base of lwkt_switch() will switch directly back to our
922 * thread. Our thread is able to retain whatever tokens it holds and
923 * if the target needs one of them the target will switch back to us
924 * and reschedule itself normally.
b68b7282
MD
925 */
926void
f9235b6d 927lwkt_preempt(thread_t ntd, int critcount)
b68b7282 928{
46a3f46d 929 struct globaldata *gd = mycpu;
cc9b6223 930 thread_t xtd;
0a3f9b47 931 thread_t td;
2d910aaf 932 int save_gd_intr_nesting_level;
b68b7282 933
26a0694b 934 /*
96728c05
MD
935 * The caller has put us in a critical section. We can only preempt
936 * if the caller of the caller was not in a critical section (basically
f9235b6d 937 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 938 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 939 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
940 *
941 * YYY The target thread must be in a critical section (else it must
942 * inherit our critical section? I dunno yet).
26a0694b 943 */
f9235b6d 944 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 945
b12defdc 946 td = gd->gd_curthread;
fbc024e4 947 if (preempt_enable == 0) {
4b2467b7 948#ifdef DEBUG_LWKT_THREAD
fbc024e4 949 ++preempt_miss;
4b2467b7 950#endif
fbc024e4
MD
951 return;
952 }
f9235b6d 953 if (ntd->td_pri <= td->td_pri) {
4b2467b7 954#ifdef DEBUG_LWKT_THREAD
57c254db 955 ++preempt_miss;
4b2467b7 956#endif
57c254db
MD
957 return;
958 }
f9235b6d 959 if (td->td_critcount > critcount) {
4b2467b7 960#ifdef DEBUG_LWKT_THREAD
96728c05 961 ++preempt_miss;
4b2467b7 962#endif
96728c05
MD
963 return;
964 }
203592a0 965 if (td->td_nest_count >= 2) {
4b2467b7 966#ifdef DEBUG_LWKT_THREAD
203592a0 967 ++preempt_miss;
4b2467b7 968#endif
203592a0
MD
969 return;
970 }
121f93bc 971 if (td->td_cscount) {
4b2467b7 972#ifdef DEBUG_LWKT_THREAD
121f93bc 973 ++preempt_miss;
4b2467b7 974#endif
121f93bc
MD
975 return;
976 }
46a3f46d 977 if (ntd->td_gd != gd) {
4b2467b7 978#ifdef DEBUG_LWKT_THREAD
96728c05 979 ++preempt_miss;
4b2467b7 980#endif
96728c05
MD
981 return;
982 }
ee89e80b 983
41a01a4d 984 /*
77912481
MD
985 * We don't have to check spinlocks here as they will also bump
986 * td_critcount.
d3d1cbc8
MD
987 *
988 * Do not try to preempt if the target thread is holding any tokens.
989 * We could try to acquire the tokens but this case is so rare there
990 * is no need to support it.
41a01a4d 991 */
0846e4ce 992 KKASSERT(gd->gd_spinlocks == 0);
77912481 993
3b998fa9 994 if (TD_TOKS_HELD(ntd)) {
4b2467b7 995#ifdef DEBUG_LWKT_THREAD
d3d1cbc8 996 ++preempt_miss;
4b2467b7 997#endif
d3d1cbc8
MD
998 return;
999 }
26a0694b 1000 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
4b2467b7 1001#ifdef DEBUG_LWKT_THREAD
26a0694b 1002 ++preempt_weird;
4b2467b7 1003#endif
26a0694b
MD
1004 return;
1005 }
1006 if (ntd->td_preempted) {
4b2467b7 1007#ifdef DEBUG_LWKT_THREAD
4b5f931b 1008 ++preempt_hit;
4b2467b7 1009#endif
26a0694b 1010 return;
b68b7282 1011 }
da0b0e8b 1012 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1013
8ec60c3f
MD
1014 /*
1015 * Since we are able to preempt the current thread, there is no need to
1016 * call need_lwkt_resched().
2d910aaf
MD
1017 *
1018 * We must temporarily clear gd_intr_nesting_level around the switch
1019 * since switchouts from the target thread are allowed (they will just
1020 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1021 *
1022 * A preemption must switch back to the original thread, assert the
1023 * case.
8ec60c3f 1024 */
4b2467b7 1025#ifdef DEBUG_LWKT_THREAD
26a0694b 1026 ++preempt_hit;
4b2467b7 1027#endif
26a0694b
MD
1028 ntd->td_preempted = td;
1029 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1030 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1031 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1032 gd->gd_intr_nesting_level = 0;
121f93bc
MD
1033
1034 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
cc9b6223
MD
1035 ntd->td_flags |= TDF_RUNNING;
1036 xtd = td->td_switch(ntd);
1037 KKASSERT(xtd == ntd);
1038 lwkt_switch_return(xtd);
2d910aaf 1039 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1040
26a0694b
MD
1041 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1042 ntd->td_preempted = NULL;
1043 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1044}
1045
f1d1c3fa 1046/*
faaeffac 1047 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1048 * This will work inside a critical section but not inside a hard code
1049 * section.
ef0fdad1 1050 *
f1d1c3fa
MD
1051 * (self contained on a per cpu basis)
1052 */
1053void
faaeffac 1054splz_check(void)
f1d1c3fa 1055{
7966cb69
MD
1056 globaldata_t gd = mycpu;
1057 thread_t td = gd->gd_curthread;
ef0fdad1 1058
4a28fe22
MD
1059 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1060 gd->gd_intr_nesting_level == 0 &&
1061 td->td_nest_count < 2)
1062 {
f1d1c3fa 1063 splz();
4a28fe22
MD
1064 }
1065}
1066
1067/*
1068 * This version is integrated into crit_exit, reqflags has already
1069 * been tested but td_critcount has not.
1070 *
1071 * We only want to execute the splz() on the 1->0 transition of
1072 * critcount and not in a hard code section or if too deeply nested.
925040f2 1073 *
0846e4ce 1074 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
4a28fe22
MD
1075 */
1076void
1077lwkt_maybe_splz(thread_t td)
1078{
1079 globaldata_t gd = td->td_gd;
1080
1081 if (td->td_critcount == 0 &&
1082 gd->gd_intr_nesting_level == 0 &&
1083 td->td_nest_count < 2)
1084 {
1085 splz();
1086 }
f1d1c3fa
MD
1087}
1088
e6546af9
MD
1089/*
1090 * Drivers which set up processing co-threads can call this function to
1091 * run the co-thread at a higher priority and to allow it to preempt
1092 * normal threads.
1093 */
1094void
1095lwkt_set_interrupt_support_thread(void)
1096{
1097 thread_t td = curthread;
1098
1099 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1100 td->td_flags |= TDF_INTTHREAD;
1101 td->td_preemptable = lwkt_preempt;
1102}
1103
1104
8ad65e08 1105/*
f9235b6d
MD
1106 * This function is used to negotiate a passive release of the current
1107 * process/lwp designation with the user scheduler, allowing the user
1108 * scheduler to schedule another user thread. The related kernel thread
1109 * (curthread) continues running in the released state.
8ad65e08
MD
1110 */
1111void
f9235b6d 1112lwkt_passive_release(struct thread *td)
8ad65e08 1113{
f9235b6d
MD
1114 struct lwp *lp = td->td_lwp;
1115
1116 td->td_release = NULL;
1117 lwkt_setpri_self(TDPRI_KERN_USER);
d992c377 1118
f9235b6d 1119 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1120}
1121
f9235b6d 1122
3824f392 1123/*
d2d8515b
MD
1124 * This implements a LWKT yield, allowing a kernel thread to yield to other
1125 * kernel threads at the same or higher priority. This function can be
1126 * called in a tight loop and will typically only yield once per tick.
f9235b6d 1127 *
d2d8515b
MD
1128 * Most kernel threads run at the same priority in order to allow equal
1129 * sharing.
f9235b6d
MD
1130 *
1131 * (self contained on a per cpu basis)
3824f392
MD
1132 */
1133void
f9235b6d 1134lwkt_yield(void)
3824f392 1135{
f9235b6d
MD
1136 globaldata_t gd = mycpu;
1137 thread_t td = gd->gd_curthread;
3824f392 1138
fbe96076
MD
1139 /*
1140 * Should never be called with spinlocks held but there is a path
1141 * via ACPI where it might happen.
1142 */
1143 if (gd->gd_spinlocks)
1144 return;
1145
1146 /*
1147 * Safe to call splz if we are not too-heavily nested.
1148 */
f9235b6d
MD
1149 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1150 splz();
fbe96076
MD
1151
1152 /*
1153 * Caller allows switching
1154 */
85946b6c 1155 if (lwkt_resched_wanted()) {
dd8be70a
MD
1156 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
1157 lwkt_schedule_self(td);
f9235b6d 1158 lwkt_switch();
f9235b6d 1159 }
3824f392
MD
1160}
1161
40504122
MD
1162/*
1163 * The quick version processes pending interrupts and higher-priority
1164 * LWKT threads but will not round-robin same-priority LWKT threads.
de4d4cb0
MD
1165 *
1166 * When called while attempting to return to userland the only same-pri
1167 * threads are the ones which have already tried to become the current
1168 * user process.
40504122
MD
1169 */
1170void
1171lwkt_yield_quick(void)
1172{
1173 globaldata_t gd = mycpu;
1174 thread_t td = gd->gd_curthread;
1175
1176 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1177 splz();
1178 if (lwkt_resched_wanted()) {
9c99cb33 1179 crit_enter();
40504122
MD
1180 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1181 clear_lwkt_resched();
1182 } else {
dd8be70a 1183 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
40504122
MD
1184 lwkt_schedule_self(curthread);
1185 lwkt_switch();
1186 }
9c99cb33 1187 crit_exit();
40504122
MD
1188 }
1189}
1190
3824f392 1191/*
f9235b6d
MD
1192 * This yield is designed for kernel threads with a user context.
1193 *
1194 * The kernel acting on behalf of the user is potentially cpu-bound,
1195 * this function will efficiently allow other threads to run and also
1196 * switch to other processes by releasing.
3824f392
MD
1197 *
1198 * The lwkt_user_yield() function is designed to have very low overhead
1199 * if no yield is determined to be needed.
1200 */
1201void
1202lwkt_user_yield(void)
1203{
f9235b6d
MD
1204 globaldata_t gd = mycpu;
1205 thread_t td = gd->gd_curthread;
1206
fbe96076
MD
1207 /*
1208 * Should never be called with spinlocks held but there is a path
1209 * via ACPI where it might happen.
1210 */
1211 if (gd->gd_spinlocks)
1212 return;
1213
f9235b6d
MD
1214 /*
1215 * Always run any pending interrupts in case we are in a critical
1216 * section.
1217 */
1218 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1219 splz();
3824f392 1220
3824f392 1221 /*
f9235b6d
MD
1222 * Switch (which forces a release) if another kernel thread needs
1223 * the cpu, if userland wants us to resched, or if our kernel
1224 * quantum has run out.
3824f392 1225 */
f9235b6d 1226 if (lwkt_resched_wanted() ||
85946b6c 1227 user_resched_wanted())
f9235b6d 1228 {
3824f392 1229 lwkt_switch();
3824f392
MD
1230 }
1231
f9235b6d 1232#if 0
3824f392 1233 /*
f9235b6d
MD
1234 * Reacquire the current process if we are released.
1235 *
1236 * XXX not implemented atm. The kernel may be holding locks and such,
1237 * so we want the thread to continue to receive cpu.
3824f392 1238 */
f9235b6d
MD
1239 if (td->td_release == NULL && lp) {
1240 lp->lwp_proc->p_usched->acquire_curproc(lp);
1241 td->td_release = lwkt_passive_release;
1242 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1243 }
f9235b6d 1244#endif
b9eb1c19
MD
1245}
1246
8ad65e08 1247/*
f1d1c3fa
MD
1248 * Generic schedule. Possibly schedule threads belonging to other cpus and
1249 * deal with threads that might be blocked on a wait queue.
1250 *
0a3f9b47
MD
1251 * We have a little helper inline function which does additional work after
1252 * the thread has been enqueued, including dealing with preemption and
1253 * setting need_lwkt_resched() (which prevents the kernel from returning
1254 * to userland until it has processed higher priority threads).
6330a558
MD
1255 *
1256 * It is possible for this routine to be called after a failed _enqueue
1257 * (due to the target thread migrating, sleeping, or otherwise blocked).
1258 * We have to check that the thread is actually on the run queue!
8ad65e08 1259 */
0a3f9b47
MD
1260static __inline
1261void
85946b6c 1262_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
0a3f9b47 1263{
6330a558 1264 if (ntd->td_flags & TDF_RUNQ) {
85946b6c 1265 if (ntd->td_preemptable) {
f9235b6d 1266 ntd->td_preemptable(ntd, ccount); /* YYY +token */
6330a558 1267 }
0a3f9b47
MD
1268 }
1269}
1270
361d01dd 1271static __inline
8ad65e08 1272void
85946b6c 1273_lwkt_schedule(thread_t td)
8ad65e08 1274{
37af14fe
MD
1275 globaldata_t mygd = mycpu;
1276
cf709dd2
MD
1277 KASSERT(td != &td->td_gd->gd_idlethread,
1278 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1279 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1280 crit_enter_gd(mygd);
4643740a
MD
1281 KKASSERT(td->td_lwp == NULL ||
1282 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1283
37af14fe 1284 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1285 _lwkt_enqueue(td);
1286 } else {
f1d1c3fa 1287 /*
7cd8d145
MD
1288 * If we own the thread, there is no race (since we are in a
1289 * critical section). If we do not own the thread there might
1290 * be a race but the target cpu will deal with it.
f1d1c3fa 1291 */
7cd8d145 1292 if (td->td_gd == mygd) {
9d265729 1293 _lwkt_enqueue(td);
85946b6c 1294 _lwkt_schedule_post(mygd, td, 1);
f1d1c3fa 1295 } else {
e381e77c 1296 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1297 }
8ad65e08 1298 }
37af14fe 1299 crit_exit_gd(mygd);
8ad65e08
MD
1300}
1301
361d01dd
MD
1302void
1303lwkt_schedule(thread_t td)
1304{
85946b6c 1305 _lwkt_schedule(td);
361d01dd
MD
1306}
1307
1308void
85946b6c 1309lwkt_schedule_noresched(thread_t td) /* XXX not impl */
361d01dd 1310{
85946b6c 1311 _lwkt_schedule(td);
361d01dd
MD
1312}
1313
e381e77c
MD
1314/*
1315 * When scheduled remotely if frame != NULL the IPIQ is being
1316 * run via doreti or an interrupt then preemption can be allowed.
1317 *
1318 * To allow preemption we have to drop the critical section so only
1319 * one is present in _lwkt_schedule_post.
1320 */
1321static void
1322lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1323{
1324 thread_t td = curthread;
1325 thread_t ntd = arg;
1326
1327 if (frame && ntd->td_preemptable) {
1328 crit_exit_noyield(td);
85946b6c 1329 _lwkt_schedule(ntd);
e381e77c
MD
1330 crit_enter_quick(td);
1331 } else {
85946b6c 1332 _lwkt_schedule(ntd);
e381e77c
MD
1333 }
1334}
1335
d9eea1a5 1336/*
52eedfb5
MD
1337 * Thread migration using a 'Pull' method. The thread may or may not be
1338 * the current thread. It MUST be descheduled and in a stable state.
1339 * lwkt_giveaway() must be called on the cpu owning the thread.
1340 *
1341 * At any point after lwkt_giveaway() is called, the target cpu may
1342 * 'pull' the thread by calling lwkt_acquire().
1343 *
ae8e83e6
MD
1344 * We have to make sure the thread is not sitting on a per-cpu tsleep
1345 * queue or it will blow up when it moves to another cpu.
1346 *
52eedfb5 1347 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1348 */
52eedfb5
MD
1349void
1350lwkt_giveaway(thread_t td)
1351{
3b4192fb 1352 globaldata_t gd = mycpu;
52eedfb5 1353
3b4192fb
MD
1354 crit_enter_gd(gd);
1355 if (td->td_flags & TDF_TSLEEPQ)
1356 tsleep_remove(td);
1357 KKASSERT(td->td_gd == gd);
1358 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1359 td->td_flags |= TDF_MIGRATING;
1360 crit_exit_gd(gd);
52eedfb5
MD
1361}
1362
a2a5ad0d
MD
1363void
1364lwkt_acquire(thread_t td)
1365{
37af14fe
MD
1366 globaldata_t gd;
1367 globaldata_t mygd;
a2a5ad0d 1368
52eedfb5 1369 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1370 gd = td->td_gd;
37af14fe 1371 mygd = mycpu;
52eedfb5 1372 if (gd != mycpu) {
3a06728e
MD
1373#ifdef LOOPMASK
1374 uint64_t tsc_base = rdtsc();
1375#endif
35238fa5 1376 cpu_lfence();
52eedfb5 1377 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1378 crit_enter_gd(mygd);
cfaeae2a 1379 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23 1380 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
df910c23 1381 lwkt_process_ipiq();
52eedfb5 1382 cpu_lfence();
a86ce0cd 1383#ifdef _KERNEL_VIRTUAL
eb8c4738 1384 vkernel_yield();
3a06728e
MD
1385#endif
1386#ifdef LOOPMASK
1387 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1388 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1389 td, td->td_flags);
1390 tsc_base = rdtsc();
1391 }
a86ce0cd 1392#endif
df910c23 1393 }
cfaeae2a 1394 DEBUG_POP_INFO();
562273ea 1395 cpu_mfence();
37af14fe 1396 td->td_gd = mygd;
52eedfb5
MD
1397 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1398 td->td_flags &= ~TDF_MIGRATING;
1399 crit_exit_gd(mygd);
1400 } else {
1401 crit_enter_gd(mygd);
1402 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1403 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1404 crit_exit_gd(mygd);
a2a5ad0d
MD
1405 }
1406}
1407
f1d1c3fa
MD
1408/*
1409 * Generic deschedule. Descheduling threads other then your own should be
dae65060 1410 * done only in carefully controlled circumstances. Descheduling is
1411 * asynchronous.
f1d1c3fa
MD
1412 *
1413 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1414 */
1415void
1416lwkt_deschedule(thread_t td)
1417{
f1d1c3fa
MD
1418 crit_enter();
1419 if (td == curthread) {
1420 _lwkt_dequeue(td);
1421 } else {
a72187e9 1422 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1423 _lwkt_dequeue(td);
1424 } else {
b8a98473 1425 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1426 }
1427 }
1428 crit_exit();
1429}
1430
4b5f931b
MD
1431/*
1432 * Set the target thread's priority. This routine does not automatically
1433 * switch to a higher priority thread, LWKT threads are not designed for
1434 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1435 */
1436void
1437lwkt_setpri(thread_t td, int pri)
1438{
f9235b6d
MD
1439 if (td->td_pri != pri) {
1440 KKASSERT(pri >= 0);
1441 crit_enter();
1442 if (td->td_flags & TDF_RUNQ) {
d2d8515b 1443 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1444 _lwkt_dequeue(td);
1445 td->td_pri = pri;
1446 _lwkt_enqueue(td);
1447 } else {
1448 td->td_pri = pri;
1449 }
1450 crit_exit();
26a0694b 1451 }
26a0694b
MD
1452}
1453
03bd0a5e
MD
1454/*
1455 * Set the initial priority for a thread prior to it being scheduled for
1456 * the first time. The thread MUST NOT be scheduled before or during
1457 * this call. The thread may be assigned to a cpu other then the current
1458 * cpu.
1459 *
1460 * Typically used after a thread has been created with TDF_STOPPREQ,
1461 * and before the thread is initially scheduled.
1462 */
1463void
1464lwkt_setpri_initial(thread_t td, int pri)
1465{
1466 KKASSERT(pri >= 0);
1467 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1468 td->td_pri = pri;
03bd0a5e
MD
1469}
1470
26a0694b
MD
1471void
1472lwkt_setpri_self(int pri)
1473{
1474 thread_t td = curthread;
1475
4b5f931b
MD
1476 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1477 crit_enter();
1478 if (td->td_flags & TDF_RUNQ) {
1479 _lwkt_dequeue(td);
f9235b6d 1480 td->td_pri = pri;
4b5f931b
MD
1481 _lwkt_enqueue(td);
1482 } else {
f9235b6d 1483 td->td_pri = pri;
4b5f931b
MD
1484 }
1485 crit_exit();
1486}
1487
f9235b6d 1488/*
85946b6c 1489 * hz tick scheduler clock for LWKT threads
f9235b6d
MD
1490 */
1491void
85946b6c 1492lwkt_schedulerclock(thread_t td)
f9235b6d 1493{
85946b6c
MD
1494 globaldata_t gd = td->td_gd;
1495 thread_t xtd;
2a418930 1496
c6a766f4
MD
1497 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1498 if (xtd == td) {
85946b6c
MD
1499 /*
1500 * If the current thread is at the head of the runq shift it to the
1501 * end of any equal-priority threads and request a LWKT reschedule
1502 * if it moved.
d992c377
MD
1503 *
1504 * Ignore upri in this situation. There will only be one user thread
1505 * in user mode, all others will be user threads running in kernel
1506 * mode and we have to make sure they get some cpu.
85946b6c
MD
1507 */
1508 xtd = TAILQ_NEXT(td, td_threadq);
1509 if (xtd && xtd->td_pri == td->td_pri) {
1510 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1511 while (xtd && xtd->td_pri == td->td_pri)
1512 xtd = TAILQ_NEXT(xtd, td_threadq);
1513 if (xtd)
1514 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1515 else
1516 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1517 need_lwkt_resched();
f9235b6d 1518 }
c6a766f4 1519 } else if (xtd) {
85946b6c
MD
1520 /*
1521 * If we scheduled a thread other than the one at the head of the
1522 * queue always request a reschedule every tick.
1523 */
1524 need_lwkt_resched();
f9235b6d 1525 }
c6a766f4 1526 /* else curthread probably the idle thread, no need to reschedule */
f9235b6d
MD
1527}
1528
5d21b981 1529/*
dae65060 1530 * Migrate the current thread to the specified cpu.
52eedfb5 1531 *
cc9b6223
MD
1532 * This is accomplished by descheduling ourselves from the current cpu
1533 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1534 * 'old' thread wants to migrate after it has been completely switched out
1535 * and will complete the migration.
1536 *
1537 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1538 *
1539 * We must be sure to release our current process designation (if a user
1540 * process) before clearing out any tsleepq we are on because the release
1541 * code may re-add us.
ae8e83e6
MD
1542 *
1543 * We must be sure to remove ourselves from the current cpu's tsleepq
1544 * before potentially moving to another queue. The thread can be on
1545 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1546 */
5d21b981
MD
1547
1548void
1549lwkt_setcpu_self(globaldata_t rgd)
1550{
5d21b981
MD
1551 thread_t td = curthread;
1552
1553 if (td->td_gd != rgd) {
1554 crit_enter_quick(td);
cc9b6223 1555
95858b91
MD
1556 if (td->td_release)
1557 td->td_release(td);
ae8e83e6 1558 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1559 tsleep_remove(td);
cc9b6223
MD
1560
1561 /*
1562 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1563 * trying to deschedule ourselves and switch away, then deschedule
1564 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1565 * call lwkt_switch() to complete the operation.
1566 */
5d21b981
MD
1567 td->td_flags |= TDF_MIGRATING;
1568 lwkt_deschedule_self(td);
52eedfb5 1569 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1570 td->td_migrate_gd = rgd;
5d21b981 1571 lwkt_switch();
cc9b6223
MD
1572
1573 /*
1574 * We are now on the target cpu
1575 */
1576 KKASSERT(rgd == mycpu);
52eedfb5 1577 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1578 crit_exit_quick(td);
1579 }
5d21b981
MD
1580}
1581
ecdefdda
MD
1582void
1583lwkt_migratecpu(int cpuid)
1584{
ecdefdda
MD
1585 globaldata_t rgd;
1586
1587 rgd = globaldata_find(cpuid);
1588 lwkt_setcpu_self(rgd);
ecdefdda
MD
1589}
1590
5d21b981
MD
1591/*
1592 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1593 * do not have to enter another one).
1594 *
1595 * The thread (td) has already been completely descheduled from the
1596 * originating cpu and we can simply assert the case. The thread is
1597 * assigned to the new cpu and enqueued.
5d21b981 1598 *
cc9b6223 1599 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1600 */
1601static void
1602lwkt_setcpu_remote(void *arg)
1603{
1604 thread_t td = arg;
1605 globaldata_t gd = mycpu;
1606
cc9b6223 1607 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1608 td->td_gd = gd;
562273ea 1609 cpu_mfence();
5d21b981 1610 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1611 KKASSERT(td->td_migrate_gd == NULL);
4643740a
MD
1612 KKASSERT(td->td_lwp == NULL ||
1613 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
5d21b981
MD
1614 _lwkt_enqueue(td);
1615}
1616
553ea3c8 1617struct lwp *
4b5f931b
MD
1618lwkt_preempted_proc(void)
1619{
73e4f7b9 1620 thread_t td = curthread;
4b5f931b
MD
1621 while (td->td_preempted)
1622 td = td->td_preempted;
553ea3c8 1623 return(td->td_lwp);
4b5f931b
MD
1624}
1625
99df837e
MD
1626/*
1627 * Create a kernel process/thread/whatever. It shares it's address space
1628 * with proc0 - ie: kernel only.
1629 *
d2d8515b
MD
1630 * If the cpu is not specified one will be selected. In the future
1631 * specifying a cpu of -1 will enable kernel thread migration between
1632 * cpus.
99df837e
MD
1633 */
1634int
c9e9fb21
MD
1635lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1636 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1637{
73e4f7b9 1638 thread_t td;
e2565a42 1639 __va_list ap;
99df837e 1640
d3d32139 1641 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1642 tdflags);
a2a5ad0d
MD
1643 if (tdp)
1644 *tdp = td;
709799ea 1645 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1646
1647 /*
1648 * Set up arg0 for 'ps' etc
1649 */
e2565a42 1650 __va_start(ap, fmt);
379210cb 1651 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1652 __va_end(ap);
99df837e
MD
1653
1654 /*
1655 * Schedule the thread to run
1656 */
4643740a
MD
1657 if (td->td_flags & TDF_NOSTART)
1658 td->td_flags &= ~TDF_NOSTART;
ef0fdad1 1659 else
4643740a 1660 lwkt_schedule(td);
99df837e
MD
1661 return 0;
1662}
1663
1664/*
1665 * Destroy an LWKT thread. Warning! This function is not called when
1666 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1667 * uses a different reaping mechanism.
1668 */
1669void
1670lwkt_exit(void)
1671{
1672 thread_t td = curthread;
c070746a 1673 thread_t std;
8826f33a 1674 globaldata_t gd;
99df837e 1675
2883d2d8
MD
1676 /*
1677 * Do any cleanup that might block here
1678 */
2883d2d8
MD
1679 biosched_done(td);
1680 dsched_exit_thread(td);
c070746a
MD
1681
1682 /*
1683 * Get us into a critical section to interlock gd_freetd and loop
1684 * until we can get it freed.
1685 *
1686 * We have to cache the current td in gd_freetd because objcache_put()ing
1687 * it would rip it out from under us while our thread is still active.
2af9d75d
MD
1688 *
1689 * We are the current thread so of course our own TDF_RUNNING bit will
1690 * be set, so unlike the lwp reap code we don't wait for it to clear.
c070746a
MD
1691 */
1692 gd = mycpu;
37af14fe 1693 crit_enter_quick(td);
2af9d75d
MD
1694 for (;;) {
1695 if (td->td_refs) {
1696 tsleep(td, 0, "tdreap", 1);
1697 continue;
1698 }
1699 if ((std = gd->gd_freetd) != NULL) {
1700 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1701 gd->gd_freetd = NULL;
1702 objcache_put(thread_cache, std);
1703 continue;
1704 }
1705 break;
c070746a 1706 }
3b4192fb
MD
1707
1708 /*
1709 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1710 * the last time. We cannot block after this point or we may end
1711 * up with a stale td on the tsleepq.
eb2adbf5
MD
1712 *
1713 * None of this may block, the critical section is the only thing
1714 * protecting tdallq and the only thing preventing new lwkt_hold()
1715 * thread refs now.
3b4192fb
MD
1716 */
1717 if (td->td_flags & TDF_TSLEEPQ)
1718 tsleep_remove(td);
37af14fe 1719 lwkt_deschedule_self(td);
e56e4dea 1720 lwkt_remove_tdallq(td);
74c9628e 1721 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1722
1723 /*
1724 * Final cleanup
1725 */
1726 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1727 if (td->td_flags & TDF_ALLOCATED_THREAD)
1728 gd->gd_freetd = td;
99df837e
MD
1729 cpu_thread_exit();
1730}
1731
e56e4dea
MD
1732void
1733lwkt_remove_tdallq(thread_t td)
1734{
1735 KKASSERT(td->td_gd == mycpu);
1736 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1737}
1738
9cf43f91
MD
1739/*
1740 * Code reduction and branch prediction improvements. Call/return
1741 * overhead on modern cpus often degenerates into 0 cycles due to
1742 * the cpu's branch prediction hardware and return pc cache. We
1743 * can take advantage of this by not inlining medium-complexity
1744 * functions and we can also reduce the branch prediction impact
1745 * by collapsing perfectly predictable branches into a single
1746 * procedure instead of duplicating it.
1747 *
1748 * Is any of this noticeable? Probably not, so I'll take the
1749 * smaller code size.
1750 */
1751void
b6468f56 1752crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1753{
b6468f56 1754 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1755}
1756
2d93b37a
MD
1757void
1758crit_panic(void)
1759{
1760 thread_t td = curthread;
850634cc 1761 int lcrit = td->td_critcount;
2d93b37a 1762
850634cc 1763 td->td_critcount = 0;
a4d95680 1764 cpu_ccfence();
850634cc 1765 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1766 /* NOT REACHED */
2d93b37a
MD
1767}
1768
bd8015ca
MD
1769/*
1770 * Called from debugger/panic on cpus which have been stopped. We must still
b19f40a4 1771 * process the IPIQ while stopped.
bd8015ca
MD
1772 *
1773 * If we are dumping also try to process any pending interrupts. This may
1774 * or may not work depending on the state of the cpu at the point it was
1775 * stopped.
1776 */
1777void
1778lwkt_smp_stopped(void)
1779{
1780 globaldata_t gd = mycpu;
1781
bd8015ca
MD
1782 if (dumping) {
1783 lwkt_process_ipiq();
b19f40a4 1784 --gd->gd_intr_nesting_level;
bd8015ca 1785 splz();
b19f40a4 1786 ++gd->gd_intr_nesting_level;
bd8015ca
MD
1787 } else {
1788 lwkt_process_ipiq();
1789 }
63cff036 1790 cpu_smp_stopped();
bd8015ca 1791}