syscons - Avoid potential blocking issue.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
cc9b6223 102static void lwkt_setcpu_remote(void *arg);
88ebb169 103#endif
f9235b6d 104static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 105
0855a2af
JG
106extern void cpu_heavy_restore(void);
107extern void cpu_lwkt_restore(void);
108extern void cpu_kthread_restore(void);
109extern void cpu_idle_restore(void);
110
fb0f29c4
MD
111/*
112 * We can make all thread ports use the spin backend instead of the thread
113 * backend. This should only be set to debug the spin backend.
114 */
115TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
116
0f7a3396 117#ifdef INVARIANTS
0c52fa62
SG
118SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
119 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 120#endif
0c52fa62
SG
121SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
122 "Number of switched threads");
9733f757 123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 124 "Successful preemption events");
9733f757 125SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
126 "Failed preemption events");
127SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
128 "Number of preempted threads.");
38717797
HP
129#ifdef INVARIANTS
130SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
131 &token_contention_count, 0, "spinning due to token contention");
38717797 132#endif
f9235b6d 133static int fairq_enable = 1;
2a418930
MD
134SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
135 &fairq_enable, 0, "Turn on fairq priority accumulators");
136static int lwkt_spin_loops = 10;
137SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
138 &lwkt_spin_loops, 0, "");
139static int lwkt_spin_delay = 1;
140SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
141 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
142static int lwkt_spin_method = 1;
143SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
144 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
d5b2d319
MD
145static int lwkt_spin_fatal = 0; /* disabled */
146SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
147 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 148static int preempt_enable = 1;
2a418930
MD
149SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
150 &preempt_enable, 0, "Enable preemption");
765b1ae0
MD
151static int lwkt_cache_threads = 32;
152SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
153 &lwkt_cache_threads, 0, "thread+kstack cache");
fbc024e4 154
2a418930
MD
155static __cachealign int lwkt_cseq_rindex;
156static __cachealign int lwkt_cseq_windex;
05220613 157
4b5f931b
MD
158/*
159 * These helper procedures handle the runq, they can only be called from
160 * within a critical section.
75cdbe6c
MD
161 *
162 * WARNING! Prior to SMP being brought up it is possible to enqueue and
163 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
164 * instead of 'mycpu' when referencing the globaldata structure. Once
165 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 166 */
f1d1c3fa
MD
167static __inline
168void
169_lwkt_dequeue(thread_t td)
170{
171 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 172 struct globaldata *gd = td->td_gd;
4b5f931b 173
f1d1c3fa 174 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
175 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
176 gd->gd_fairq_total_pri -= td->td_pri;
177 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 178 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
179 }
180}
181
f9235b6d
MD
182/*
183 * Priority enqueue.
184 *
185 * NOTE: There are a limited number of lwkt threads runnable since user
186 * processes only schedule one at a time per cpu.
187 */
f1d1c3fa
MD
188static __inline
189void
190_lwkt_enqueue(thread_t td)
191{
f9235b6d
MD
192 thread_t xtd;
193
7f5d7ed7 194 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 195 struct globaldata *gd = td->td_gd;
4b5f931b 196
f1d1c3fa 197 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
198 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
199 if (xtd == NULL) {
200 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
2a418930 201 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d
MD
202 } else {
203 while (xtd && xtd->td_pri > td->td_pri)
204 xtd = TAILQ_NEXT(xtd, td_threadq);
205 if (xtd)
206 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
207 else
208 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
209 }
210 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
211 }
212}
8ad65e08 213
40aaf5fc
NT
214static __boolean_t
215_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
216{
217 struct thread *td = (struct thread *)obj;
218
219 td->td_kstack = NULL;
220 td->td_kstack_size = 0;
221 td->td_flags = TDF_ALLOCATED_THREAD;
222 return (1);
223}
224
225static void
226_lwkt_thread_dtor(void *obj, void *privdata)
227{
228 struct thread *td = (struct thread *)obj;
229
230 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
231 ("_lwkt_thread_dtor: not allocated from objcache"));
232 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
233 td->td_kstack_size > 0,
234 ("_lwkt_thread_dtor: corrupted stack"));
235 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
236}
237
238/*
239 * Initialize the lwkt s/system.
765b1ae0
MD
240 *
241 * Nominally cache up to 32 thread + kstack structures.
40aaf5fc
NT
242 */
243void
244lwkt_init(void)
245{
765b1ae0
MD
246 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
247 thread_cache = objcache_create_mbacked(
248 M_THREAD, sizeof(struct thread),
249 NULL, lwkt_cache_threads,
250 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
251}
252
37af14fe
MD
253/*
254 * Schedule a thread to run. As the current thread we can always safely
255 * schedule ourselves, and a shortcut procedure is provided for that
256 * function.
257 *
258 * (non-blocking, self contained on a per cpu basis)
259 */
260void
261lwkt_schedule_self(thread_t td)
262{
cfaeae2a 263 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 264 crit_enter_quick(td);
f9235b6d
MD
265 KASSERT(td != &td->td_gd->gd_idlethread,
266 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 267 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 268 _lwkt_enqueue(td);
37af14fe
MD
269 crit_exit_quick(td);
270}
271
272/*
273 * Deschedule a thread.
274 *
275 * (non-blocking, self contained on a per cpu basis)
276 */
277void
278lwkt_deschedule_self(thread_t td)
279{
280 crit_enter_quick(td);
37af14fe
MD
281 _lwkt_dequeue(td);
282 crit_exit_quick(td);
283}
284
8ad65e08
MD
285/*
286 * LWKTs operate on a per-cpu basis
287 *
73e4f7b9 288 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
289 */
290void
291lwkt_gdinit(struct globaldata *gd)
292{
f9235b6d 293 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 294 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
295}
296
7d0bac62
MD
297/*
298 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
299 * or LWKT start address before it can be scheduled. If the target cpu is
300 * -1 the thread will be created on the current cpu.
0cfcada1
MD
301 *
302 * If you intend to create a thread without a process context this function
303 * does everything except load the startup and switcher function.
7d0bac62
MD
304 */
305thread_t
d3d32139 306lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 307{
c070746a 308 globaldata_t gd = mycpu;
99df837e 309 void *stack;
7d0bac62 310
c070746a
MD
311 /*
312 * If static thread storage is not supplied allocate a thread. Reuse
313 * a cached free thread if possible. gd_freetd is used to keep an exiting
314 * thread intact through the exit.
315 */
ef0fdad1 316 if (td == NULL) {
cf709dd2
MD
317 crit_enter_gd(gd);
318 if ((td = gd->gd_freetd) != NULL) {
319 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
320 TDF_RUNQ)) == 0);
c070746a 321 gd->gd_freetd = NULL;
cf709dd2 322 } else {
c070746a 323 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
324 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
325 TDF_RUNQ)) == 0);
326 }
327 crit_exit_gd(gd);
40aaf5fc
NT
328 KASSERT((td->td_flags &
329 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
330 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
331 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 332 }
c070746a
MD
333
334 /*
335 * Try to reuse cached stack.
336 */
f470d0c8
MD
337 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
338 if (flags & TDF_ALLOCATED_STACK) {
e4846942 339 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
340 stack = NULL;
341 }
342 }
343 if (stack == NULL) {
e40cfbd7 344 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 345 flags |= TDF_ALLOCATED_STACK;
99df837e 346 }
75cdbe6c 347 if (cpu < 0)
c070746a 348 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 349 else
f470d0c8 350 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 351 return(td);
7d0bac62
MD
352}
353
354/*
355 * Initialize a preexisting thread structure. This function is used by
356 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
357 *
f8c3996b
MD
358 * All threads start out in a critical section at a priority of
359 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
360 * appropriate. This function may send an IPI message when the
361 * requested cpu is not the current cpu and consequently gd_tdallq may
362 * not be initialized synchronously from the point of view of the originating
363 * cpu.
364 *
365 * NOTE! we have to be careful in regards to creating threads for other cpus
366 * if SMP has not yet been activated.
7d0bac62 367 */
41a01a4d
MD
368#ifdef SMP
369
75cdbe6c
MD
370static void
371lwkt_init_thread_remote(void *arg)
372{
373 thread_t td = arg;
374
52eedfb5
MD
375 /*
376 * Protected by critical section held by IPI dispatch
377 */
75cdbe6c
MD
378 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
379}
380
41a01a4d
MD
381#endif
382
fdce8919
MD
383/*
384 * lwkt core thread structural initialization.
385 *
386 * NOTE: All threads are initialized as mpsafe threads.
387 */
7d0bac62 388void
f470d0c8
MD
389lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
390 struct globaldata *gd)
7d0bac62 391{
37af14fe
MD
392 globaldata_t mygd = mycpu;
393
99df837e
MD
394 bzero(td, sizeof(struct thread));
395 td->td_kstack = stack;
f470d0c8 396 td->td_kstack_size = stksize;
d3d32139 397 td->td_flags = flags;
26a0694b 398 td->td_gd = gd;
f9235b6d
MD
399 td->td_pri = TDPRI_KERN_DAEMON;
400 td->td_critcount = 1;
3b998fa9 401 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
402 if (lwkt_use_spin_port)
403 lwkt_initport_spin(&td->td_msgport);
404 else
405 lwkt_initport_thread(&td->td_msgport, td);
99df837e 406 pmap_init_thread(td);
0f7a3396 407#ifdef SMP
5d21b981
MD
408 /*
409 * Normally initializing a thread for a remote cpu requires sending an
410 * IPI. However, the idlethread is setup before the other cpus are
411 * activated so we have to treat it as a special case. XXX manipulation
412 * of gd_tdallq requires the BGL.
413 */
414 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 415 crit_enter_gd(mygd);
75cdbe6c 416 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 417 crit_exit_gd(mygd);
75cdbe6c 418 } else {
2db3b277 419 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 420 }
0f7a3396 421#else
37af14fe 422 crit_enter_gd(mygd);
0f7a3396 423 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 424 crit_exit_gd(mygd);
0f7a3396 425#endif
8c72e3d5
AH
426
427 dsched_new_thread(td);
73e4f7b9
MD
428}
429
430void
431lwkt_set_comm(thread_t td, const char *ctl, ...)
432{
e2565a42 433 __va_list va;
73e4f7b9 434
e2565a42 435 __va_start(va, ctl);
379210cb 436 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 437 __va_end(va);
e7c0dbba 438 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
439}
440
99df837e 441void
73e4f7b9 442lwkt_hold(thread_t td)
99df837e 443{
74c9628e 444 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
445}
446
447void
448lwkt_rele(thread_t td)
449{
450 KKASSERT(td->td_refs > 0);
74c9628e 451 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
452}
453
454void
455lwkt_wait_free(thread_t td)
456{
457 while (td->td_refs)
377d4740 458 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
459}
460
461void
462lwkt_free_thread(thread_t td)
463{
74c9628e 464 KKASSERT(td->td_refs == 0);
cf709dd2 465 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
466 if (td->td_flags & TDF_ALLOCATED_THREAD) {
467 objcache_put(thread_cache, td);
468 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
469 /* client-allocated struct with internally allocated stack */
470 KASSERT(td->td_kstack && td->td_kstack_size > 0,
471 ("lwkt_free_thread: corrupted stack"));
472 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
473 td->td_kstack = NULL;
474 td->td_kstack_size = 0;
99df837e 475 }
e7c0dbba 476 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
477}
478
479
8ad65e08
MD
480/*
481 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
482 * switch to the idlethread. Switching must occur within a critical
483 * section to avoid races with the scheduling queue.
484 *
485 * We always have full control over our cpu's run queue. Other cpus
486 * that wish to manipulate our queue must use the cpu_*msg() calls to
487 * talk to our cpu, so a critical section is all that is needed and
488 * the result is very, very fast thread switching.
489 *
96728c05
MD
490 * The LWKT scheduler uses a fixed priority model and round-robins at
491 * each priority level. User process scheduling is a totally
492 * different beast and LWKT priorities should not be confused with
493 * user process priorities.
f1d1c3fa 494 *
69d78e99
MD
495 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
496 * is not called by the current thread in the preemption case, only when
497 * the preempting thread blocks (in order to return to the original thread).
cfaeae2a
MD
498 *
499 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
500 * migration and tsleep deschedule the current lwkt thread and call
501 * lwkt_switch(). In particular, the target cpu of the migration fully
502 * expects the thread to become non-runnable and can deadlock against
503 * cpusync operations if we run any IPIs prior to switching the thread out.
504 *
505 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
95858b91 506 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
8ad65e08
MD
507 */
508void
509lwkt_switch(void)
510{
37af14fe
MD
511 globaldata_t gd = mycpu;
512 thread_t td = gd->gd_curthread;
8ad65e08 513 thread_t ntd;
f9235b6d 514 thread_t xtd;
2a418930
MD
515 int spinning = lwkt_spin_loops; /* loops before HLTing */
516 int reqflags;
517 int cseq;
0f0466c0 518 int oseq;
d5b2d319 519 int fatal_count;
8ad65e08 520
da0b0e8b
MD
521 KKASSERT(gd->gd_processing_ipiq == 0);
522
46a3f46d 523 /*
27e88a6e
MD
524 * Switching from within a 'fast' (non thread switched) interrupt or IPI
525 * is illegal. However, we may have to do it anyway if we hit a fatal
526 * kernel trap or we have paniced.
527 *
528 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 529 */
27e88a6e
MD
530 if (gd->gd_intr_nesting_level) {
531 int savegdnest;
532 int savegdtrap;
533
5fddbda2 534 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
535 panic("lwkt_switch: Attempt to switch from a "
536 "a fast interrupt, ipi, or hard code section, "
537 "td %p\n",
538 td);
27e88a6e
MD
539 } else {
540 savegdnest = gd->gd_intr_nesting_level;
541 savegdtrap = gd->gd_trap_nesting_level;
542 gd->gd_intr_nesting_level = 0;
543 gd->gd_trap_nesting_level = 0;
a7422615
MD
544 if ((td->td_flags & TDF_PANICWARN) == 0) {
545 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
546 kprintf("Warning: thread switch from interrupt, IPI, "
547 "or hard code section.\n"
a7422615 548 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 549 print_backtrace(-1);
a7422615 550 }
27e88a6e
MD
551 lwkt_switch();
552 gd->gd_intr_nesting_level = savegdnest;
553 gd->gd_trap_nesting_level = savegdtrap;
554 return;
555 }
96728c05 556 }
ef0fdad1 557
cb973d15
MD
558 /*
559 * Passive release (used to transition from user to kernel mode
560 * when we block or switch rather then when we enter the kernel).
561 * This function is NOT called if we are switching into a preemption
562 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
563 * our current process designation (if we have one) and become a true
564 * LWKT thread, and may also hand the current process designation to
565 * another process and schedule thread.
cb973d15
MD
566 */
567 if (td->td_release)
568 td->td_release(td);
569
37af14fe 570 crit_enter_gd(gd);
3b998fa9 571 if (TD_TOKS_HELD(td))
9d265729
MD
572 lwkt_relalltokens(td);
573
574 /*
b02926de
MD
575 * We had better not be holding any spin locks, but don't get into an
576 * endless panic loop.
9d265729 577 */
d666840a
MD
578 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
579 ("lwkt_switch: still holding %d exclusive spinlocks!",
580 gd->gd_spinlocks_wr));
9d265729 581
8a8d5d85
MD
582
583#ifdef SMP
0f7a3396
MD
584#ifdef INVARIANTS
585 if (td->td_cscount) {
6ea70f76 586 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
587 td);
588 if (panic_on_cscount)
589 panic("switching while mastering cpusync");
590 }
591#endif
8a8d5d85 592#endif
f9235b6d
MD
593
594 /*
595 * If we had preempted another thread on this cpu, resume the preempted
596 * thread. This occurs transparently, whether the preempted thread
597 * was scheduled or not (it may have been preempted after descheduling
598 * itself).
599 *
600 * We have to setup the MP lock for the original thread after backing
601 * out the adjustment that was made to curthread when the original
602 * was preempted.
603 */
99df837e 604 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
605 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
606 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
607
608 /*
b9eb1c19
MD
609 * The interrupt may have woken a thread up, we need to properly
610 * set the reschedule flag if the originally interrupted thread is
611 * at a lower priority.
8ec60c3f 612 */
f9235b6d
MD
613 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
614 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 615 need_lwkt_resched();
f9235b6d 616 }
8a8d5d85 617 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
618 goto havethread_preempted;
619 }
620
621 /*
622 * Implement round-robin fairq with priority insertion. The priority
623 * insertion is handled by _lwkt_enqueue()
624 *
f9235b6d 625 * If we cannot obtain ownership of the tokens we cannot immediately
cfaeae2a
MD
626 * schedule the target thread.
627 *
628 * Reminder: Again, we cannot afford to run any IPIs in this path if
629 * the current thread has been descheduled.
f9235b6d
MD
630 */
631 for (;;) {
2a418930
MD
632 /*
633 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
634 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
635 * received).
636 */
637 for (;;) {
638 reqflags = gd->gd_reqflags;
639 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
640 (reqflags & ~RQF_AST_LWKT_RESCHED) |
641 RQF_WAKEUP)) {
642 break;
643 }
644 }
f9235b6d 645
4b5f931b 646 /*
2a418930
MD
647 * Hotpath - pull the head of the run queue and attempt to schedule
648 * it. Fairq exhaustion moves the task to the end of the list. If
649 * no threads are runnable we switch to the idle thread.
41a01a4d 650 */
2a418930
MD
651 for (;;) {
652 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
653
654 if (ntd == NULL) {
655 /*
656 * Runq is empty, switch to idle and clear RQF_WAKEUP
657 * to allow it to halt.
658 */
659 ntd = &gd->gd_idlethread;
6f207a2c 660#ifdef SMP
2a418930 661 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 662 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 663#endif
2a418930
MD
664 cpu_time.cp_msg[0] = 0;
665 cpu_time.cp_stallpc = 0;
666 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
667 goto haveidle;
668 }
669
670 if (ntd->td_fairq_accum >= 0)
671 break;
672
cfaeae2a 673 /*splz_check(); cannot do this here, see above */
2a418930
MD
674 lwkt_fairq_accumulate(gd, ntd);
675 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
676 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
f9235b6d 677 }
41a01a4d 678
8ec60c3f 679 /*
2a418930
MD
680 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
681 * unwanted decontention IPIs.
6f207a2c
MD
682 *
683 * NOTE: For UP there is no mplock and lwkt_getalltokens()
684 * always succeeds.
8ec60c3f 685 */
2a418930 686 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 687 goto havethread;
f9235b6d 688
f9235b6d 689 /*
2a418930
MD
690 * Coldpath (SMP only since tokens always succeed on UP)
691 *
692 * We had some contention on the thread we wanted to schedule.
693 * What we do now is try to find a thread that we can schedule
694 * in its stead until decontention reschedules on our cpu.
695 *
696 * The coldpath scan does NOT rearrange threads in the run list
697 * and it also ignores the accumulator.
698 *
699 * We do not immediately schedule a user priority thread, instead
700 * we record it in xtd and continue looking for kernel threads.
701 * A cpu can only have one user priority thread (normally) so just
702 * record the first one.
703 *
704 * NOTE: This scan will also include threads whos fairq's were
705 * accumulated in the first loop.
f9235b6d 706 */
2a418930
MD
707 ++token_contention_count;
708 xtd = NULL;
709 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
41a01a4d 710 /*
2a418930
MD
711 * Try to switch to this thread. If the thread is running at
712 * user priority we clear WAKEUP to allow decontention IPIs
713 * (since this thread is simply running until the one we wanted
714 * decontends), and we make sure that LWKT_RESCHED is not set.
df6b8ba0 715 *
2a418930
MD
716 * Otherwise for kernel threads we leave WAKEUP set to avoid
717 * unnecessary decontention IPIs.
41a01a4d 718 */
2a418930
MD
719 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
720 if (xtd == NULL)
721 xtd = ntd;
722 continue;
f9235b6d 723 }
a453459d 724
f9235b6d 725 /*
2a418930
MD
726 * Do not let the fairq get too negative. Even though we are
727 * ignoring it atm once the scheduler decontends a very negative
728 * thread will get moved to the end of the queue.
f9235b6d 729 */
2a418930
MD
730 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
731 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
732 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
733 goto havethread;
8a8d5d85 734 }
f9235b6d 735
df6b8ba0 736 /*
2a418930 737 * Well fubar, this thread is contended as well, loop
df6b8ba0 738 */
2a418930
MD
739 /* */
740 }
741
742 /*
743 * We exhausted the run list but we may have recorded a user
744 * thread to try. We have three choices based on
745 * lwkt.decontention_method.
746 *
747 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
748 * IPIs (to interrupt the user process) and test
749 * RQF_AST_LWKT_RESCHED at the same time.
750 *
751 * This results in significant decontention IPI traffic but may
752 * be more responsive.
753 *
754 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
755 * An automatic LWKT reschedule will occur on the next hardclock
756 * (typically 100hz).
757 *
758 * This results in no decontention IPI traffic but may be less
759 * responsive. This is the default.
760 *
761 * (2) Refuse to schedule the user process at this time.
762 *
763 * This is highly experimental and should not be used under
764 * normal circumstances. This can cause a user process to
765 * get starved out in situations where kernel threads are
766 * fighting each other for tokens.
767 */
768 if (xtd) {
769 ntd = xtd;
770
771 switch(lwkt_spin_method) {
772 case 0:
773 for (;;) {
774 reqflags = gd->gd_reqflags;
775 if (atomic_cmpset_int(&gd->gd_reqflags,
776 reqflags,
777 reqflags & ~RQF_WAKEUP)) {
778 break;
779 }
780 }
781 break;
782 case 1:
783 reqflags = gd->gd_reqflags;
784 break;
785 default:
786 goto skip;
787 break;
788 }
789 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
b5d16701 790 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 791 ) {
2a418930
MD
792 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
793 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
794 goto havethread;
df6b8ba0 795 }
9ac1ee6e 796
2a418930 797skip:
9ac1ee6e 798 /*
2a418930
MD
799 * Make sure RQF_WAKEUP is set if we failed to schedule the
800 * user thread to prevent the idle thread from halting.
9ac1ee6e 801 */
2a418930
MD
802 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
803 }
804
805 /*
806 * We exhausted the run list, meaning that all runnable threads
807 * are contended.
808 */
809 cpu_pause();
810 ntd = &gd->gd_idlethread;
811#ifdef SMP
812 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
813 ASSERT_NO_TOKENS_HELD(ntd);
814 /* contention case, do not clear contention mask */
815#endif
816
817 /*
818 * Ok, we might want to spin a few times as some tokens are held for
819 * very short periods of time and IPI overhead is 1uS or worse
820 * (meaning it is usually better to spin). Regardless we have to
821 * call splz_check() to be sure to service any interrupts blocked
822 * by our critical section, otherwise we could livelock e.g. IPIs.
823 *
824 * The IPI mechanic is really a last resort. In nearly all other
825 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
826 *
827 * When we decide not to spin we clear RQF_WAKEUP and switch to
828 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
829 * to halt and decontended tokens will issue an IPI to us. The
830 * idle thread will check for pending reschedules already set
831 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
832 * to here.
cfaeae2a
MD
833 *
834 * Also, if TDF_RUNQ is not set the current thread is trying to
835 * deschedule, possibly in an atomic fashion. We cannot afford to
836 * stay here.
2a418930 837 */
cfaeae2a 838 if (spinning <= 0 || (td->td_flags & TDF_RUNQ) == 0) {
2a418930
MD
839 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
840 goto haveidle;
4b5f931b 841 }
2a418930 842 --spinning;
c5724852
MD
843
844 /*
2a418930
MD
845 * When spinning a delay is required both to avoid livelocks from
846 * token order reversals (a thread may be trying to acquire multiple
847 * tokens), and also to reduce cpu cache management traffic.
848 *
849 * In order to scale to a large number of CPUs we use a time slot
850 * resequencer to force contending cpus into non-contending
851 * time-slots. The scheduler may still contend with the lock holder
852 * but will not (generally) contend with all the other cpus trying
853 * trying to get the same token.
854 *
855 * The resequencer uses a FIFO counter mechanic. The owner of the
856 * rindex at the head of the FIFO is allowed to pull itself off
857 * the FIFO and fetchadd is used to enter into the FIFO. This bit
858 * of code is VERY cache friendly and forces all spinning schedulers
859 * into their own time slots.
c5724852 860 *
2a418930
MD
861 * This code has been tested to 48-cpus and caps the cache
862 * contention load at ~1uS intervals regardless of the number of
863 * cpus. Scaling beyond 64 cpus might require additional smarts
864 * (such as separate FIFOs for specific token cases).
865 *
866 * WARNING! We can't call splz_check() or anything else here as
867 * it could cause a deadlock.
c5724852 868 */
bd52bedf 869#if defined(INVARIANTS) && defined(__amd64__)
d5b2d319
MD
870 if ((read_rflags() & PSL_I) == 0) {
871 cpu_enable_intr();
872 panic("lwkt_switch() called with interrupts disabled");
873 }
874#endif
2a418930 875 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
d5b2d319 876 fatal_count = lwkt_spin_fatal;
0f0466c0
MD
877 while ((oseq = lwkt_cseq_rindex) != cseq) {
878 cpu_ccfence();
8b402283 879#if !defined(_KERNEL_VIRTUAL)
0f0466c0
MD
880 if (cpu_mi_feature & CPU_MI_MONITOR) {
881 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
8b402283
SW
882 } else
883#endif
884 {
0f0466c0
MD
885 DELAY(1);
886 cpu_lfence();
887 }
d5b2d319
MD
888 if (fatal_count && --fatal_count == 0)
889 panic("lwkt_switch: fatal spin wait");
2a418930
MD
890 }
891 cseq = lwkt_spin_delay; /* don't trust the system operator */
892 cpu_ccfence();
893 if (cseq < 1)
894 cseq = 1;
895 if (cseq > 1000)
896 cseq = 1000;
897 DELAY(cseq);
898 atomic_add_int(&lwkt_cseq_rindex, 1);
cfaeae2a 899 splz_check(); /* ok, we already checked that td is still scheduled */
2a418930 900 /* highest level for(;;) loop */
f1d1c3fa 901 }
8a8d5d85 902
2a418930 903havethread:
8a8d5d85 904 /*
f9235b6d
MD
905 * We must always decrement td_fairq_accum on non-idle threads just
906 * in case a thread never gets a tick due to being in a continuous
2a418930 907 * critical section. The page-zeroing code does this, for example.
f9235b6d
MD
908 *
909 * If the thread we came up with is a higher or equal priority verses
910 * the thread at the head of the queue we move our thread to the
911 * front. This way we can always check the front of the queue.
be71787b
MD
912 *
913 * Clear gd_idle_repeat when doing a normal switch to a non-idle
914 * thread.
f9235b6d 915 */
f9235b6d
MD
916 ++gd->gd_cnt.v_swtch;
917 --ntd->td_fairq_accum;
9ac1ee6e 918 ntd->td_wmesg = NULL;
f9235b6d
MD
919 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
920 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
921 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
922 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
923 }
be71787b 924 gd->gd_idle_repeat = 0;
2a418930 925
f9235b6d 926havethread_preempted:
f9235b6d
MD
927 /*
928 * If the new target does not need the MP lock and we are holding it,
929 * release the MP lock. If the new target requires the MP lock we have
930 * already acquired it for the target.
8a8d5d85 931 */
2a418930 932 ;
f9235b6d
MD
933haveidle:
934 KASSERT(ntd->td_critcount,
b5d16701
MD
935 ("priority problem in lwkt_switch %d %d",
936 td->td_critcount, ntd->td_critcount));
937
94f6d86e 938 if (td != ntd) {
cc9b6223
MD
939 /*
940 * Execute the actual thread switch operation. This function
941 * returns to the current thread and returns the previous thread
942 * (which may be different from the thread we switched to).
943 *
944 * We are responsible for marking ntd as TDF_RUNNING.
945 */
94f6d86e 946 ++switch_count;
a1f0fb66 947 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
cc9b6223
MD
948 ntd->td_flags |= TDF_RUNNING;
949 lwkt_switch_return(td->td_switch(ntd));
950 /* ntd invalid, td_switch() can return a different thread_t */
94f6d86e 951 }
37af14fe
MD
952 /* NOTE: current cpu may have changed after switch */
953 crit_exit_quick(td);
8ad65e08
MD
954}
955
cc9b6223
MD
956/*
957 * Called by assembly in the td_switch (thread restore path) for thread
958 * bootstrap cases which do not 'return' to lwkt_switch().
959 */
960void
961lwkt_switch_return(thread_t otd)
962{
963#ifdef SMP
964 globaldata_t rgd;
965
966 /*
967 * Check if otd was migrating. Now that we are on ntd we can finish
968 * up the migration. This is a bit messy but it is the only place
969 * where td is known to be fully descheduled.
970 *
971 * We can only activate the migration if otd was migrating but not
972 * held on the cpu due to a preemption chain. We still have to
973 * clear TDF_RUNNING on the old thread either way.
974 *
975 * We are responsible for clearing the previously running thread's
976 * TDF_RUNNING.
977 */
978 if ((rgd = otd->td_migrate_gd) != NULL &&
979 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
980 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
981 (TDF_MIGRATING | TDF_RUNNING));
982 otd->td_migrate_gd = NULL;
983 otd->td_flags &= ~TDF_RUNNING;
984 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
985 } else {
986 otd->td_flags &= ~TDF_RUNNING;
987 }
988#else
989 otd->td_flags &= ~TDF_RUNNING;
990#endif
991}
992
b68b7282 993/*
96728c05
MD
994 * Request that the target thread preempt the current thread. Preemption
995 * only works under a specific set of conditions:
b68b7282 996 *
96728c05
MD
997 * - We are not preempting ourselves
998 * - The target thread is owned by the current cpu
999 * - We are not currently being preempted
1000 * - The target is not currently being preempted
d3d1cbc8
MD
1001 * - We are not holding any spin locks
1002 * - The target thread is not holding any tokens
96728c05
MD
1003 * - We are able to satisfy the target's MP lock requirements (if any).
1004 *
1005 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
1006 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 1007 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
1008 * to determine whether preemption is possible (aka usually just the crit
1009 * priority of lwkt_schedule() itself).
b68b7282 1010 *
26a0694b
MD
1011 * XXX at the moment we run the target thread in a critical section during
1012 * the preemption in order to prevent the target from taking interrupts
1013 * that *WE* can't. Preemption is strictly limited to interrupt threads
1014 * and interrupt-like threads, outside of a critical section, and the
1015 * preempted source thread will be resumed the instant the target blocks
1016 * whether or not the source is scheduled (i.e. preemption is supposed to
1017 * be as transparent as possible).
b68b7282
MD
1018 */
1019void
f9235b6d 1020lwkt_preempt(thread_t ntd, int critcount)
b68b7282 1021{
46a3f46d 1022 struct globaldata *gd = mycpu;
cc9b6223 1023 thread_t xtd;
0a3f9b47 1024 thread_t td;
2d910aaf 1025 int save_gd_intr_nesting_level;
b68b7282 1026
26a0694b 1027 /*
96728c05
MD
1028 * The caller has put us in a critical section. We can only preempt
1029 * if the caller of the caller was not in a critical section (basically
f9235b6d 1030 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 1031 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 1032 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
1033 *
1034 * YYY The target thread must be in a critical section (else it must
1035 * inherit our critical section? I dunno yet).
41a01a4d 1036 *
0a3f9b47 1037 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 1038 */
f9235b6d 1039 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 1040
fbc024e4
MD
1041 if (preempt_enable == 0) {
1042 ++preempt_miss;
1043 return;
1044 }
1045
0a3f9b47 1046 td = gd->gd_curthread;
f9235b6d 1047 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
1048 ++preempt_miss;
1049 return;
1050 }
f9235b6d 1051 if (td->td_critcount > critcount) {
96728c05 1052 ++preempt_miss;
8ec60c3f 1053 need_lwkt_resched();
96728c05
MD
1054 return;
1055 }
1056#ifdef SMP
46a3f46d 1057 if (ntd->td_gd != gd) {
96728c05 1058 ++preempt_miss;
8ec60c3f 1059 need_lwkt_resched();
96728c05
MD
1060 return;
1061 }
1062#endif
41a01a4d 1063 /*
77912481
MD
1064 * We don't have to check spinlocks here as they will also bump
1065 * td_critcount.
d3d1cbc8
MD
1066 *
1067 * Do not try to preempt if the target thread is holding any tokens.
1068 * We could try to acquire the tokens but this case is so rare there
1069 * is no need to support it.
41a01a4d 1070 */
77912481
MD
1071 KKASSERT(gd->gd_spinlocks_wr == 0);
1072
3b998fa9 1073 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
1074 ++preempt_miss;
1075 need_lwkt_resched();
1076 return;
1077 }
26a0694b
MD
1078 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1079 ++preempt_weird;
8ec60c3f 1080 need_lwkt_resched();
26a0694b
MD
1081 return;
1082 }
1083 if (ntd->td_preempted) {
4b5f931b 1084 ++preempt_hit;
8ec60c3f 1085 need_lwkt_resched();
26a0694b 1086 return;
b68b7282 1087 }
da0b0e8b 1088 KKASSERT(gd->gd_processing_ipiq == 0);
26a0694b 1089
8ec60c3f
MD
1090 /*
1091 * Since we are able to preempt the current thread, there is no need to
1092 * call need_lwkt_resched().
2d910aaf
MD
1093 *
1094 * We must temporarily clear gd_intr_nesting_level around the switch
1095 * since switchouts from the target thread are allowed (they will just
1096 * return to our thread), and since the target thread has its own stack.
cc9b6223
MD
1097 *
1098 * A preemption must switch back to the original thread, assert the
1099 * case.
8ec60c3f 1100 */
26a0694b
MD
1101 ++preempt_hit;
1102 ntd->td_preempted = td;
1103 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1104 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1105 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1106 gd->gd_intr_nesting_level = 0;
cc9b6223
MD
1107 ntd->td_flags |= TDF_RUNNING;
1108 xtd = td->td_switch(ntd);
1109 KKASSERT(xtd == ntd);
1110 lwkt_switch_return(xtd);
2d910aaf 1111 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1112
26a0694b
MD
1113 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1114 ntd->td_preempted = NULL;
1115 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1116}
1117
f1d1c3fa 1118/*
faaeffac 1119 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1120 * This will work inside a critical section but not inside a hard code
1121 * section.
ef0fdad1 1122 *
f1d1c3fa
MD
1123 * (self contained on a per cpu basis)
1124 */
1125void
faaeffac 1126splz_check(void)
f1d1c3fa 1127{
7966cb69
MD
1128 globaldata_t gd = mycpu;
1129 thread_t td = gd->gd_curthread;
ef0fdad1 1130
4a28fe22
MD
1131 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1132 gd->gd_intr_nesting_level == 0 &&
1133 td->td_nest_count < 2)
1134 {
f1d1c3fa 1135 splz();
4a28fe22
MD
1136 }
1137}
1138
1139/*
1140 * This version is integrated into crit_exit, reqflags has already
1141 * been tested but td_critcount has not.
1142 *
1143 * We only want to execute the splz() on the 1->0 transition of
1144 * critcount and not in a hard code section or if too deeply nested.
1145 */
1146void
1147lwkt_maybe_splz(thread_t td)
1148{
1149 globaldata_t gd = td->td_gd;
1150
1151 if (td->td_critcount == 0 &&
1152 gd->gd_intr_nesting_level == 0 &&
1153 td->td_nest_count < 2)
1154 {
1155 splz();
1156 }
f1d1c3fa
MD
1157}
1158
8ad65e08 1159/*
f9235b6d
MD
1160 * This function is used to negotiate a passive release of the current
1161 * process/lwp designation with the user scheduler, allowing the user
1162 * scheduler to schedule another user thread. The related kernel thread
1163 * (curthread) continues running in the released state.
8ad65e08
MD
1164 */
1165void
f9235b6d 1166lwkt_passive_release(struct thread *td)
8ad65e08 1167{
f9235b6d
MD
1168 struct lwp *lp = td->td_lwp;
1169
1170 td->td_release = NULL;
1171 lwkt_setpri_self(TDPRI_KERN_USER);
1172 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1173}
1174
f9235b6d 1175
3824f392 1176/*
f9235b6d
MD
1177 * This implements a normal yield. This routine is virtually a nop if
1178 * there is nothing to yield to but it will always run any pending interrupts
1179 * if called from a critical section.
1180 *
1181 * This yield is designed for kernel threads without a user context.
1182 *
1183 * (self contained on a per cpu basis)
3824f392
MD
1184 */
1185void
f9235b6d 1186lwkt_yield(void)
3824f392 1187{
f9235b6d
MD
1188 globaldata_t gd = mycpu;
1189 thread_t td = gd->gd_curthread;
1190 thread_t xtd;
3824f392 1191
f9235b6d
MD
1192 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1193 splz();
1194 if (td->td_fairq_accum < 0) {
1195 lwkt_schedule_self(curthread);
1196 lwkt_switch();
1197 } else {
1198 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1199 if (xtd && xtd->td_pri > td->td_pri) {
1200 lwkt_schedule_self(curthread);
1201 lwkt_switch();
1202 }
1203 }
3824f392
MD
1204}
1205
1206/*
f9235b6d
MD
1207 * This yield is designed for kernel threads with a user context.
1208 *
1209 * The kernel acting on behalf of the user is potentially cpu-bound,
1210 * this function will efficiently allow other threads to run and also
1211 * switch to other processes by releasing.
3824f392
MD
1212 *
1213 * The lwkt_user_yield() function is designed to have very low overhead
1214 * if no yield is determined to be needed.
1215 */
1216void
1217lwkt_user_yield(void)
1218{
f9235b6d
MD
1219 globaldata_t gd = mycpu;
1220 thread_t td = gd->gd_curthread;
1221
1222 /*
1223 * Always run any pending interrupts in case we are in a critical
1224 * section.
1225 */
1226 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1227 splz();
3824f392 1228
3824f392 1229 /*
f9235b6d
MD
1230 * Switch (which forces a release) if another kernel thread needs
1231 * the cpu, if userland wants us to resched, or if our kernel
1232 * quantum has run out.
3824f392 1233 */
f9235b6d
MD
1234 if (lwkt_resched_wanted() ||
1235 user_resched_wanted() ||
1236 td->td_fairq_accum < 0)
1237 {
3824f392 1238 lwkt_switch();
3824f392
MD
1239 }
1240
f9235b6d 1241#if 0
3824f392 1242 /*
f9235b6d
MD
1243 * Reacquire the current process if we are released.
1244 *
1245 * XXX not implemented atm. The kernel may be holding locks and such,
1246 * so we want the thread to continue to receive cpu.
3824f392 1247 */
f9235b6d
MD
1248 if (td->td_release == NULL && lp) {
1249 lp->lwp_proc->p_usched->acquire_curproc(lp);
1250 td->td_release = lwkt_passive_release;
1251 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1252 }
f9235b6d 1253#endif
b9eb1c19
MD
1254}
1255
8ad65e08 1256/*
f1d1c3fa
MD
1257 * Generic schedule. Possibly schedule threads belonging to other cpus and
1258 * deal with threads that might be blocked on a wait queue.
1259 *
0a3f9b47
MD
1260 * We have a little helper inline function which does additional work after
1261 * the thread has been enqueued, including dealing with preemption and
1262 * setting need_lwkt_resched() (which prevents the kernel from returning
1263 * to userland until it has processed higher priority threads).
6330a558
MD
1264 *
1265 * It is possible for this routine to be called after a failed _enqueue
1266 * (due to the target thread migrating, sleeping, or otherwise blocked).
1267 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1268 *
1269 * reschedok is an optimized constant propagated from lwkt_schedule() or
1270 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1271 * reschedule to be requested if the target thread has a higher priority.
1272 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1273 * be 0, prevented undesired reschedules.
8ad65e08 1274 */
0a3f9b47
MD
1275static __inline
1276void
f9235b6d 1277_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1278{
b9eb1c19 1279 thread_t otd;
c730be20 1280
6330a558 1281 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1282 if (ntd->td_preemptable && reschedok) {
f9235b6d 1283 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1284 } else if (reschedok) {
b9eb1c19 1285 otd = curthread;
f9235b6d 1286 if (ntd->td_pri > otd->td_pri)
c730be20 1287 need_lwkt_resched();
6330a558 1288 }
f9235b6d
MD
1289
1290 /*
1291 * Give the thread a little fair share scheduler bump if it
1292 * has been asleep for a while. This is primarily to avoid
1293 * a degenerate case for interrupt threads where accumulator
1294 * crosses into negative territory unnecessarily.
1295 */
1296 if (ntd->td_fairq_lticks != ticks) {
1297 ntd->td_fairq_lticks = ticks;
1298 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1299 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1300 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1301 }
0a3f9b47
MD
1302 }
1303}
1304
361d01dd 1305static __inline
8ad65e08 1306void
361d01dd 1307_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1308{
37af14fe
MD
1309 globaldata_t mygd = mycpu;
1310
cf709dd2
MD
1311 KASSERT(td != &td->td_gd->gd_idlethread,
1312 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
cfaeae2a 1313 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
37af14fe 1314 crit_enter_gd(mygd);
9388413d 1315 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1316 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1317 _lwkt_enqueue(td);
1318 } else {
f1d1c3fa 1319 /*
7cd8d145
MD
1320 * If we own the thread, there is no race (since we are in a
1321 * critical section). If we do not own the thread there might
1322 * be a race but the target cpu will deal with it.
f1d1c3fa 1323 */
0f7a3396 1324#ifdef SMP
7cd8d145 1325 if (td->td_gd == mygd) {
9d265729 1326 _lwkt_enqueue(td);
f9235b6d 1327 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1328 } else {
e381e77c 1329 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1330 }
0f7a3396 1331#else
7cd8d145 1332 _lwkt_enqueue(td);
f9235b6d 1333 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1334#endif
8ad65e08 1335 }
37af14fe 1336 crit_exit_gd(mygd);
8ad65e08
MD
1337}
1338
361d01dd
MD
1339void
1340lwkt_schedule(thread_t td)
1341{
1342 _lwkt_schedule(td, 1);
1343}
1344
1345void
1346lwkt_schedule_noresched(thread_t td)
1347{
1348 _lwkt_schedule(td, 0);
1349}
1350
88ebb169
SW
1351#ifdef SMP
1352
e381e77c
MD
1353/*
1354 * When scheduled remotely if frame != NULL the IPIQ is being
1355 * run via doreti or an interrupt then preemption can be allowed.
1356 *
1357 * To allow preemption we have to drop the critical section so only
1358 * one is present in _lwkt_schedule_post.
1359 */
1360static void
1361lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1362{
1363 thread_t td = curthread;
1364 thread_t ntd = arg;
1365
1366 if (frame && ntd->td_preemptable) {
1367 crit_exit_noyield(td);
1368 _lwkt_schedule(ntd, 1);
1369 crit_enter_quick(td);
1370 } else {
1371 _lwkt_schedule(ntd, 1);
1372 }
1373}
1374
d9eea1a5 1375/*
52eedfb5
MD
1376 * Thread migration using a 'Pull' method. The thread may or may not be
1377 * the current thread. It MUST be descheduled and in a stable state.
1378 * lwkt_giveaway() must be called on the cpu owning the thread.
1379 *
1380 * At any point after lwkt_giveaway() is called, the target cpu may
1381 * 'pull' the thread by calling lwkt_acquire().
1382 *
ae8e83e6
MD
1383 * We have to make sure the thread is not sitting on a per-cpu tsleep
1384 * queue or it will blow up when it moves to another cpu.
1385 *
52eedfb5 1386 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1387 */
52eedfb5
MD
1388void
1389lwkt_giveaway(thread_t td)
1390{
3b4192fb 1391 globaldata_t gd = mycpu;
52eedfb5 1392
3b4192fb
MD
1393 crit_enter_gd(gd);
1394 if (td->td_flags & TDF_TSLEEPQ)
1395 tsleep_remove(td);
1396 KKASSERT(td->td_gd == gd);
1397 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1398 td->td_flags |= TDF_MIGRATING;
1399 crit_exit_gd(gd);
52eedfb5
MD
1400}
1401
a2a5ad0d
MD
1402void
1403lwkt_acquire(thread_t td)
1404{
37af14fe
MD
1405 globaldata_t gd;
1406 globaldata_t mygd;
cc9b6223 1407 int retry = 10000000;
a2a5ad0d 1408
52eedfb5 1409 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1410 gd = td->td_gd;
37af14fe 1411 mygd = mycpu;
52eedfb5 1412 if (gd != mycpu) {
35238fa5 1413 cpu_lfence();
52eedfb5 1414 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1415 crit_enter_gd(mygd);
cfaeae2a 1416 DEBUG_PUSH_INFO("lwkt_acquire");
df910c23
MD
1417 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1418#ifdef SMP
1419 lwkt_process_ipiq();
1420#endif
52eedfb5 1421 cpu_lfence();
cc9b6223
MD
1422 if (--retry == 0) {
1423 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1424 td, td->td_flags);
1425 retry = 10000000;
1426 }
df910c23 1427 }
cfaeae2a 1428 DEBUG_POP_INFO();
562273ea 1429 cpu_mfence();
37af14fe 1430 td->td_gd = mygd;
52eedfb5
MD
1431 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1432 td->td_flags &= ~TDF_MIGRATING;
1433 crit_exit_gd(mygd);
1434 } else {
1435 crit_enter_gd(mygd);
1436 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1437 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1438 crit_exit_gd(mygd);
a2a5ad0d
MD
1439 }
1440}
1441
52eedfb5
MD
1442#endif
1443
f1d1c3fa
MD
1444/*
1445 * Generic deschedule. Descheduling threads other then your own should be
1446 * done only in carefully controlled circumstances. Descheduling is
1447 * asynchronous.
1448 *
1449 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1450 */
1451void
1452lwkt_deschedule(thread_t td)
1453{
f1d1c3fa 1454 crit_enter();
b8a98473 1455#ifdef SMP
f1d1c3fa
MD
1456 if (td == curthread) {
1457 _lwkt_dequeue(td);
1458 } else {
a72187e9 1459 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1460 _lwkt_dequeue(td);
1461 } else {
b8a98473 1462 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1463 }
1464 }
b8a98473
MD
1465#else
1466 _lwkt_dequeue(td);
1467#endif
f1d1c3fa
MD
1468 crit_exit();
1469}
1470
4b5f931b
MD
1471/*
1472 * Set the target thread's priority. This routine does not automatically
1473 * switch to a higher priority thread, LWKT threads are not designed for
1474 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1475 */
1476void
1477lwkt_setpri(thread_t td, int pri)
1478{
a72187e9 1479 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1480 if (td->td_pri != pri) {
1481 KKASSERT(pri >= 0);
1482 crit_enter();
1483 if (td->td_flags & TDF_RUNQ) {
1484 _lwkt_dequeue(td);
1485 td->td_pri = pri;
1486 _lwkt_enqueue(td);
1487 } else {
1488 td->td_pri = pri;
1489 }
1490 crit_exit();
26a0694b 1491 }
26a0694b
MD
1492}
1493
03bd0a5e
MD
1494/*
1495 * Set the initial priority for a thread prior to it being scheduled for
1496 * the first time. The thread MUST NOT be scheduled before or during
1497 * this call. The thread may be assigned to a cpu other then the current
1498 * cpu.
1499 *
1500 * Typically used after a thread has been created with TDF_STOPPREQ,
1501 * and before the thread is initially scheduled.
1502 */
1503void
1504lwkt_setpri_initial(thread_t td, int pri)
1505{
1506 KKASSERT(pri >= 0);
1507 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1508 td->td_pri = pri;
03bd0a5e
MD
1509}
1510
26a0694b
MD
1511void
1512lwkt_setpri_self(int pri)
1513{
1514 thread_t td = curthread;
1515
4b5f931b
MD
1516 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1517 crit_enter();
1518 if (td->td_flags & TDF_RUNQ) {
1519 _lwkt_dequeue(td);
f9235b6d 1520 td->td_pri = pri;
4b5f931b
MD
1521 _lwkt_enqueue(td);
1522 } else {
f9235b6d 1523 td->td_pri = pri;
4b5f931b
MD
1524 }
1525 crit_exit();
1526}
1527
f9235b6d
MD
1528/*
1529 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1530 *
1531 * Example: two competing threads, same priority N. decrement by (2*N)
1532 * increment by N*8, each thread will get 4 ticks.
1533 */
1534void
1535lwkt_fairq_schedulerclock(thread_t td)
1536{
2a418930
MD
1537 globaldata_t gd;
1538
f9235b6d
MD
1539 if (fairq_enable) {
1540 while (td) {
2a418930
MD
1541 gd = td->td_gd;
1542 if (td != &gd->gd_idlethread) {
1543 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1544 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1545 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
f9235b6d
MD
1546 if (td->td_fairq_accum < 0)
1547 need_lwkt_resched();
1548 td->td_fairq_lticks = ticks;
1549 }
1550 td = td->td_preempted;
1551 }
1552 }
1553}
1554
1555static void
1556lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1557{
1558 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1559 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1560 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1561}
1562
5d21b981 1563/*
52eedfb5
MD
1564 * Migrate the current thread to the specified cpu.
1565 *
cc9b6223
MD
1566 * This is accomplished by descheduling ourselves from the current cpu
1567 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1568 * 'old' thread wants to migrate after it has been completely switched out
1569 * and will complete the migration.
1570 *
1571 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1572 *
1573 * We must be sure to release our current process designation (if a user
1574 * process) before clearing out any tsleepq we are on because the release
1575 * code may re-add us.
ae8e83e6
MD
1576 *
1577 * We must be sure to remove ourselves from the current cpu's tsleepq
1578 * before potentially moving to another queue. The thread can be on
1579 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1580 */
5d21b981
MD
1581
1582void
1583lwkt_setcpu_self(globaldata_t rgd)
1584{
1585#ifdef SMP
1586 thread_t td = curthread;
1587
1588 if (td->td_gd != rgd) {
1589 crit_enter_quick(td);
cc9b6223 1590
95858b91
MD
1591 if (td->td_release)
1592 td->td_release(td);
ae8e83e6 1593 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1594 tsleep_remove(td);
cc9b6223
MD
1595
1596 /*
1597 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1598 * trying to deschedule ourselves and switch away, then deschedule
1599 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1600 * call lwkt_switch() to complete the operation.
1601 */
5d21b981
MD
1602 td->td_flags |= TDF_MIGRATING;
1603 lwkt_deschedule_self(td);
52eedfb5 1604 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
cc9b6223 1605 td->td_migrate_gd = rgd;
5d21b981 1606 lwkt_switch();
cc9b6223
MD
1607
1608 /*
1609 * We are now on the target cpu
1610 */
1611 KKASSERT(rgd == mycpu);
52eedfb5 1612 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1613 crit_exit_quick(td);
1614 }
1615#endif
1616}
1617
ecdefdda
MD
1618void
1619lwkt_migratecpu(int cpuid)
1620{
1621#ifdef SMP
1622 globaldata_t rgd;
1623
1624 rgd = globaldata_find(cpuid);
1625 lwkt_setcpu_self(rgd);
1626#endif
1627}
1628
cc9b6223 1629#ifdef SMP
5d21b981
MD
1630/*
1631 * Remote IPI for cpu migration (called while in a critical section so we
cc9b6223
MD
1632 * do not have to enter another one).
1633 *
1634 * The thread (td) has already been completely descheduled from the
1635 * originating cpu and we can simply assert the case. The thread is
1636 * assigned to the new cpu and enqueued.
5d21b981 1637 *
cc9b6223 1638 * The thread will re-add itself to tdallq when it resumes execution.
5d21b981
MD
1639 */
1640static void
1641lwkt_setcpu_remote(void *arg)
1642{
1643 thread_t td = arg;
1644 globaldata_t gd = mycpu;
1645
cc9b6223 1646 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
5d21b981 1647 td->td_gd = gd;
562273ea 1648 cpu_mfence();
5d21b981 1649 td->td_flags &= ~TDF_MIGRATING;
cc9b6223 1650 KKASSERT(td->td_migrate_gd == NULL);
9388413d 1651 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1652 _lwkt_enqueue(td);
1653}
3d28ff59 1654#endif
5d21b981 1655
553ea3c8 1656struct lwp *
4b5f931b
MD
1657lwkt_preempted_proc(void)
1658{
73e4f7b9 1659 thread_t td = curthread;
4b5f931b
MD
1660 while (td->td_preempted)
1661 td = td->td_preempted;
553ea3c8 1662 return(td->td_lwp);
4b5f931b
MD
1663}
1664
99df837e
MD
1665/*
1666 * Create a kernel process/thread/whatever. It shares it's address space
1667 * with proc0 - ie: kernel only.
1668 *
365fa13f
MD
1669 * NOTE! By default new threads are created with the MP lock held. A
1670 * thread which does not require the MP lock should release it by calling
1671 * rel_mplock() at the start of the new thread.
99df837e
MD
1672 */
1673int
c9e9fb21
MD
1674lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1675 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1676{
73e4f7b9 1677 thread_t td;
e2565a42 1678 __va_list ap;
99df837e 1679
d3d32139 1680 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1681 tdflags);
a2a5ad0d
MD
1682 if (tdp)
1683 *tdp = td;
709799ea 1684 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1685
1686 /*
1687 * Set up arg0 for 'ps' etc
1688 */
e2565a42 1689 __va_start(ap, fmt);
379210cb 1690 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1691 __va_end(ap);
99df837e
MD
1692
1693 /*
1694 * Schedule the thread to run
1695 */
ef0fdad1
MD
1696 if ((td->td_flags & TDF_STOPREQ) == 0)
1697 lwkt_schedule(td);
1698 else
1699 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1700 return 0;
1701}
1702
1703/*
1704 * Destroy an LWKT thread. Warning! This function is not called when
1705 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1706 * uses a different reaping mechanism.
1707 */
1708void
1709lwkt_exit(void)
1710{
1711 thread_t td = curthread;
c070746a 1712 thread_t std;
8826f33a 1713 globaldata_t gd;
99df837e 1714
2883d2d8
MD
1715 /*
1716 * Do any cleanup that might block here
1717 */
99df837e 1718 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1719 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1720 caps_exit(td);
2883d2d8
MD
1721 biosched_done(td);
1722 dsched_exit_thread(td);
c070746a
MD
1723
1724 /*
1725 * Get us into a critical section to interlock gd_freetd and loop
1726 * until we can get it freed.
1727 *
1728 * We have to cache the current td in gd_freetd because objcache_put()ing
1729 * it would rip it out from under us while our thread is still active.
1730 */
1731 gd = mycpu;
37af14fe 1732 crit_enter_quick(td);
c070746a 1733 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1734 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1735 gd->gd_freetd = NULL;
1736 objcache_put(thread_cache, std);
1737 }
3b4192fb
MD
1738
1739 /*
1740 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1741 * the last time. We cannot block after this point or we may end
1742 * up with a stale td on the tsleepq.
3b4192fb
MD
1743 */
1744 if (td->td_flags & TDF_TSLEEPQ)
1745 tsleep_remove(td);
37af14fe 1746 lwkt_deschedule_self(td);
e56e4dea 1747 lwkt_remove_tdallq(td);
74c9628e 1748 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1749
1750 /*
1751 * Final cleanup
1752 */
1753 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1754 if (td->td_flags & TDF_ALLOCATED_THREAD)
1755 gd->gd_freetd = td;
99df837e
MD
1756 cpu_thread_exit();
1757}
1758
e56e4dea
MD
1759void
1760lwkt_remove_tdallq(thread_t td)
1761{
1762 KKASSERT(td->td_gd == mycpu);
1763 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1764}
1765
9cf43f91
MD
1766/*
1767 * Code reduction and branch prediction improvements. Call/return
1768 * overhead on modern cpus often degenerates into 0 cycles due to
1769 * the cpu's branch prediction hardware and return pc cache. We
1770 * can take advantage of this by not inlining medium-complexity
1771 * functions and we can also reduce the branch prediction impact
1772 * by collapsing perfectly predictable branches into a single
1773 * procedure instead of duplicating it.
1774 *
1775 * Is any of this noticeable? Probably not, so I'll take the
1776 * smaller code size.
1777 */
1778void
b6468f56 1779crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1780{
b6468f56 1781 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1782}
1783
2d93b37a
MD
1784void
1785crit_panic(void)
1786{
1787 thread_t td = curthread;
850634cc 1788 int lcrit = td->td_critcount;
2d93b37a 1789
850634cc
AH
1790 td->td_critcount = 0;
1791 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1792 /* NOT REACHED */
2d93b37a
MD
1793}
1794
d165e668
MD
1795#ifdef SMP
1796
bd8015ca
MD
1797/*
1798 * Called from debugger/panic on cpus which have been stopped. We must still
1799 * process the IPIQ while stopped, even if we were stopped while in a critical
1800 * section (XXX).
1801 *
1802 * If we are dumping also try to process any pending interrupts. This may
1803 * or may not work depending on the state of the cpu at the point it was
1804 * stopped.
1805 */
1806void
1807lwkt_smp_stopped(void)
1808{
1809 globaldata_t gd = mycpu;
1810
1811 crit_enter_gd(gd);
1812 if (dumping) {
1813 lwkt_process_ipiq();
1814 splz();
1815 } else {
1816 lwkt_process_ipiq();
1817 }
1818 crit_exit_gd(gd);
1819}
1820
d165e668 1821#endif