kernel - Major MPSAFE Infrastructure
[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);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
77912481
MD
154static int user_pri_sched = 0;
155SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, "");
05220613 156
4b5f931b
MD
157/*
158 * These helper procedures handle the runq, they can only be called from
159 * within a critical section.
75cdbe6c
MD
160 *
161 * WARNING! Prior to SMP being brought up it is possible to enqueue and
162 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
163 * instead of 'mycpu' when referencing the globaldata structure. Once
164 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 165 */
f1d1c3fa
MD
166static __inline
167void
168_lwkt_dequeue(thread_t td)
169{
170 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 171 struct globaldata *gd = td->td_gd;
4b5f931b 172
f1d1c3fa 173 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
174 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
175 gd->gd_fairq_total_pri -= td->td_pri;
176 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
177 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
178 }
179}
180
f9235b6d
MD
181/*
182 * Priority enqueue.
183 *
184 * NOTE: There are a limited number of lwkt threads runnable since user
185 * processes only schedule one at a time per cpu.
186 */
f1d1c3fa
MD
187static __inline
188void
189_lwkt_enqueue(thread_t td)
190{
f9235b6d
MD
191 thread_t xtd;
192
7f5d7ed7 193 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 194 struct globaldata *gd = td->td_gd;
4b5f931b 195
f1d1c3fa 196 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
197 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
198 if (xtd == NULL) {
199 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
200 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
201 } else {
202 while (xtd && xtd->td_pri > td->td_pri)
203 xtd = TAILQ_NEXT(xtd, td_threadq);
204 if (xtd)
205 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
206 else
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 }
209 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
210 }
211}
8ad65e08 212
40aaf5fc
NT
213static __boolean_t
214_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
215{
216 struct thread *td = (struct thread *)obj;
217
218 td->td_kstack = NULL;
219 td->td_kstack_size = 0;
220 td->td_flags = TDF_ALLOCATED_THREAD;
221 return (1);
222}
223
224static void
225_lwkt_thread_dtor(void *obj, void *privdata)
226{
227 struct thread *td = (struct thread *)obj;
228
229 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
230 ("_lwkt_thread_dtor: not allocated from objcache"));
231 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
232 td->td_kstack_size > 0,
233 ("_lwkt_thread_dtor: corrupted stack"));
234 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
235}
236
237/*
238 * Initialize the lwkt s/system.
239 */
240void
241lwkt_init(void)
242{
243 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
244 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
245 NULL, CACHE_NTHREADS/2,
246 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
247}
248
37af14fe
MD
249/*
250 * Schedule a thread to run. As the current thread we can always safely
251 * schedule ourselves, and a shortcut procedure is provided for that
252 * function.
253 *
254 * (non-blocking, self contained on a per cpu basis)
255 */
256void
257lwkt_schedule_self(thread_t td)
258{
259 crit_enter_quick(td);
f9235b6d
MD
260 KASSERT(td != &td->td_gd->gd_idlethread,
261 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 262 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 263 _lwkt_enqueue(td);
37af14fe
MD
264 crit_exit_quick(td);
265}
266
267/*
268 * Deschedule a thread.
269 *
270 * (non-blocking, self contained on a per cpu basis)
271 */
272void
273lwkt_deschedule_self(thread_t td)
274{
275 crit_enter_quick(td);
37af14fe
MD
276 _lwkt_dequeue(td);
277 crit_exit_quick(td);
278}
279
8ad65e08
MD
280/*
281 * LWKTs operate on a per-cpu basis
282 *
73e4f7b9 283 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
284 */
285void
286lwkt_gdinit(struct globaldata *gd)
287{
f9235b6d 288 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 289 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
290}
291
7d0bac62
MD
292/*
293 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
294 * or LWKT start address before it can be scheduled. If the target cpu is
295 * -1 the thread will be created on the current cpu.
0cfcada1
MD
296 *
297 * If you intend to create a thread without a process context this function
298 * does everything except load the startup and switcher function.
7d0bac62
MD
299 */
300thread_t
d3d32139 301lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 302{
c070746a 303 globaldata_t gd = mycpu;
99df837e 304 void *stack;
7d0bac62 305
c070746a
MD
306 /*
307 * If static thread storage is not supplied allocate a thread. Reuse
308 * a cached free thread if possible. gd_freetd is used to keep an exiting
309 * thread intact through the exit.
310 */
ef0fdad1 311 if (td == NULL) {
c070746a
MD
312 if ((td = gd->gd_freetd) != NULL)
313 gd->gd_freetd = NULL;
314 else
315 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
316 KASSERT((td->td_flags &
317 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
318 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
319 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 320 }
c070746a
MD
321
322 /*
323 * Try to reuse cached stack.
324 */
f470d0c8
MD
325 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
326 if (flags & TDF_ALLOCATED_STACK) {
e4846942 327 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
328 stack = NULL;
329 }
330 }
331 if (stack == NULL) {
e4846942 332 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 333 flags |= TDF_ALLOCATED_STACK;
99df837e 334 }
75cdbe6c 335 if (cpu < 0)
c070746a 336 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 337 else
f470d0c8 338 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 339 return(td);
7d0bac62
MD
340}
341
342/*
343 * Initialize a preexisting thread structure. This function is used by
344 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
345 *
f8c3996b
MD
346 * All threads start out in a critical section at a priority of
347 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
348 * appropriate. This function may send an IPI message when the
349 * requested cpu is not the current cpu and consequently gd_tdallq may
350 * not be initialized synchronously from the point of view of the originating
351 * cpu.
352 *
353 * NOTE! we have to be careful in regards to creating threads for other cpus
354 * if SMP has not yet been activated.
7d0bac62 355 */
41a01a4d
MD
356#ifdef SMP
357
75cdbe6c
MD
358static void
359lwkt_init_thread_remote(void *arg)
360{
361 thread_t td = arg;
362
52eedfb5
MD
363 /*
364 * Protected by critical section held by IPI dispatch
365 */
75cdbe6c
MD
366 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
367}
368
41a01a4d
MD
369#endif
370
7d0bac62 371void
f470d0c8
MD
372lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
373 struct globaldata *gd)
7d0bac62 374{
37af14fe
MD
375 globaldata_t mygd = mycpu;
376
99df837e
MD
377 bzero(td, sizeof(struct thread));
378 td->td_kstack = stack;
f470d0c8 379 td->td_kstack_size = stksize;
d3d32139 380 td->td_flags = flags;
26a0694b 381 td->td_gd = gd;
f9235b6d
MD
382 td->td_pri = TDPRI_KERN_DAEMON;
383 td->td_critcount = 1;
3b998fa9 384 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
385#ifdef SMP
386 if ((flags & TDF_MPSAFE) == 0)
387 td->td_mpcount = 1;
388#endif
fb0f29c4
MD
389 if (lwkt_use_spin_port)
390 lwkt_initport_spin(&td->td_msgport);
391 else
392 lwkt_initport_thread(&td->td_msgport, td);
99df837e 393 pmap_init_thread(td);
0f7a3396 394#ifdef SMP
5d21b981
MD
395 /*
396 * Normally initializing a thread for a remote cpu requires sending an
397 * IPI. However, the idlethread is setup before the other cpus are
398 * activated so we have to treat it as a special case. XXX manipulation
399 * of gd_tdallq requires the BGL.
400 */
401 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 402 crit_enter_gd(mygd);
75cdbe6c 403 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 404 crit_exit_gd(mygd);
75cdbe6c 405 } else {
2db3b277 406 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 407 }
0f7a3396 408#else
37af14fe 409 crit_enter_gd(mygd);
0f7a3396 410 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 411 crit_exit_gd(mygd);
0f7a3396 412#endif
8c72e3d5
AH
413
414 dsched_new_thread(td);
73e4f7b9
MD
415}
416
417void
418lwkt_set_comm(thread_t td, const char *ctl, ...)
419{
e2565a42 420 __va_list va;
73e4f7b9 421
e2565a42 422 __va_start(va, ctl);
379210cb 423 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 424 __va_end(va);
e7c0dbba 425 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
426}
427
99df837e 428void
73e4f7b9 429lwkt_hold(thread_t td)
99df837e 430{
73e4f7b9
MD
431 ++td->td_refs;
432}
433
434void
435lwkt_rele(thread_t td)
436{
437 KKASSERT(td->td_refs > 0);
438 --td->td_refs;
439}
440
441void
442lwkt_wait_free(thread_t td)
443{
444 while (td->td_refs)
377d4740 445 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
446}
447
448void
449lwkt_free_thread(thread_t td)
450{
d9eea1a5 451 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
452 ("lwkt_free_thread: did not exit! %p", td));
453
40aaf5fc
NT
454 if (td->td_flags & TDF_ALLOCATED_THREAD) {
455 objcache_put(thread_cache, td);
456 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
457 /* client-allocated struct with internally allocated stack */
458 KASSERT(td->td_kstack && td->td_kstack_size > 0,
459 ("lwkt_free_thread: corrupted stack"));
460 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
461 td->td_kstack = NULL;
462 td->td_kstack_size = 0;
99df837e 463 }
e7c0dbba 464 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
465}
466
467
8ad65e08
MD
468/*
469 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
470 * switch to the idlethread. Switching must occur within a critical
471 * section to avoid races with the scheduling queue.
472 *
473 * We always have full control over our cpu's run queue. Other cpus
474 * that wish to manipulate our queue must use the cpu_*msg() calls to
475 * talk to our cpu, so a critical section is all that is needed and
476 * the result is very, very fast thread switching.
477 *
96728c05
MD
478 * The LWKT scheduler uses a fixed priority model and round-robins at
479 * each priority level. User process scheduling is a totally
480 * different beast and LWKT priorities should not be confused with
481 * user process priorities.
f1d1c3fa 482 *
96728c05
MD
483 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
484 * cleans it up. Note that the td_switch() function cannot do anything that
485 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
486 * the target thread (not the current thread). It's nice to have a scheduler
487 * that does not need the MP lock to work because it allows us to do some
488 * really cool high-performance MP lock optimizations.
69d78e99
MD
489 *
490 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
491 * is not called by the current thread in the preemption case, only when
492 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
493 */
494void
495lwkt_switch(void)
496{
37af14fe
MD
497 globaldata_t gd = mycpu;
498 thread_t td = gd->gd_curthread;
8ad65e08 499 thread_t ntd;
f9235b6d
MD
500 thread_t xtd;
501 thread_t nlast;
f9235b6d 502 int nquserok;
6f207a2c 503#ifdef SMP
8a8d5d85
MD
504 int mpheld;
505#endif
f9235b6d 506 int didaccumulate;
b37f18d6
MD
507 const char *lmsg; /* diagnostic - 'systat -pv 1' */
508 const void *laddr;
8ad65e08 509
46a3f46d 510 /*
27e88a6e
MD
511 * Switching from within a 'fast' (non thread switched) interrupt or IPI
512 * is illegal. However, we may have to do it anyway if we hit a fatal
513 * kernel trap or we have paniced.
514 *
515 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 516 */
27e88a6e
MD
517 if (gd->gd_intr_nesting_level) {
518 int savegdnest;
519 int savegdtrap;
520
521 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
522 panic("lwkt_switch: cannot switch from within "
523 "a fast interrupt, yet, td %p\n", td);
524 } else {
525 savegdnest = gd->gd_intr_nesting_level;
526 savegdtrap = gd->gd_trap_nesting_level;
527 gd->gd_intr_nesting_level = 0;
528 gd->gd_trap_nesting_level = 0;
a7422615
MD
529 if ((td->td_flags & TDF_PANICWARN) == 0) {
530 td->td_flags |= TDF_PANICWARN;
6ea70f76 531 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 532 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 533 print_backtrace(-1);
a7422615 534 }
27e88a6e
MD
535 lwkt_switch();
536 gd->gd_intr_nesting_level = savegdnest;
537 gd->gd_trap_nesting_level = savegdtrap;
538 return;
539 }
96728c05 540 }
ef0fdad1 541
cb973d15
MD
542 /*
543 * Passive release (used to transition from user to kernel mode
544 * when we block or switch rather then when we enter the kernel).
545 * This function is NOT called if we are switching into a preemption
546 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
547 * our current process designation (if we have one) and become a true
548 * LWKT thread, and may also hand the current process designation to
549 * another process and schedule thread.
cb973d15
MD
550 */
551 if (td->td_release)
552 td->td_release(td);
553
37af14fe 554 crit_enter_gd(gd);
3b998fa9 555 if (TD_TOKS_HELD(td))
9d265729
MD
556 lwkt_relalltokens(td);
557
558 /*
b02926de
MD
559 * We had better not be holding any spin locks, but don't get into an
560 * endless panic loop.
9d265729 561 */
bbb31c5d
MD
562 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
563 ("lwkt_switch: still holding a shared spinlock %p!",
564 gd->gd_spinlock_rd));
d666840a
MD
565 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
566 ("lwkt_switch: still holding %d exclusive spinlocks!",
567 gd->gd_spinlocks_wr));
9d265729 568
8a8d5d85
MD
569
570#ifdef SMP
571 /*
572 * td_mpcount cannot be used to determine if we currently hold the
573 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
574 * to get the lock, and switch out if it can't. Our ownership of
575 * the actual lock will remain stable while we are in a critical section
576 * (but, of course, another cpu may own or release the lock so the
577 * actual value of mp_lock is not stable).
8a8d5d85
MD
578 */
579 mpheld = MP_LOCK_HELD();
0f7a3396
MD
580#ifdef INVARIANTS
581 if (td->td_cscount) {
6ea70f76 582 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
583 td);
584 if (panic_on_cscount)
585 panic("switching while mastering cpusync");
586 }
587#endif
8a8d5d85 588#endif
f9235b6d
MD
589
590 /*
591 * If we had preempted another thread on this cpu, resume the preempted
592 * thread. This occurs transparently, whether the preempted thread
593 * was scheduled or not (it may have been preempted after descheduling
594 * itself).
595 *
596 * We have to setup the MP lock for the original thread after backing
597 * out the adjustment that was made to curthread when the original
598 * was preempted.
599 */
99df837e 600 if ((ntd = td->td_preempted) != NULL) {
26a0694b 601 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 602#ifdef SMP
96728c05 603 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 604 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
605 td, ntd, td->td_mpcount, ntd->td_mpcount);
606 }
8a8d5d85
MD
607 if (ntd->td_mpcount) {
608 td->td_mpcount -= ntd->td_mpcount;
609 KKASSERT(td->td_mpcount >= 0);
610 }
611#endif
26a0694b 612 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
613
614 /*
b9eb1c19
MD
615 * The interrupt may have woken a thread up, we need to properly
616 * set the reschedule flag if the originally interrupted thread is
617 * at a lower priority.
8ec60c3f 618 */
f9235b6d
MD
619 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
620 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 621 need_lwkt_resched();
f9235b6d 622 }
8a8d5d85 623 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
624 goto havethread_preempted;
625 }
626
627 /*
628 * Implement round-robin fairq with priority insertion. The priority
629 * insertion is handled by _lwkt_enqueue()
630 *
631 * We have to adjust the MP lock for the target thread. If we
632 * need the MP lock and cannot obtain it we try to locate a
633 * thread that does not need the MP lock. If we cannot, we spin
634 * instead of HLT.
635 *
636 * A similar issue exists for the tokens held by the target thread.
637 * If we cannot obtain ownership of the tokens we cannot immediately
638 * schedule the thread.
639 */
640 for (;;) {
641 clear_lwkt_resched();
642 didaccumulate = 0;
643 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
644
4b5f931b 645 /*
f9235b6d 646 * Hotpath if we can get all necessary resources.
41a01a4d 647 *
f9235b6d 648 * If nothing is runnable switch to the idle thread
41a01a4d 649 */
f9235b6d
MD
650 if (ntd == NULL) {
651 ntd = &gd->gd_idlethread;
652 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
653 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 654#ifdef SMP
f9235b6d
MD
655 if (ntd->td_mpcount) {
656 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
657 panic("Idle thread %p was holding the BGL!", ntd);
658 if (mpheld == 0) {
659 cpu_pause();
660 continue;
661 }
662 }
6f207a2c 663#endif
b37f18d6
MD
664 cpu_time.cp_msg[0] = 0;
665 cpu_time.cp_stallpc = 0;
f9235b6d
MD
666 goto haveidle;
667 }
41a01a4d 668
8ec60c3f 669 /*
f9235b6d 670 * Hotpath schedule
6f207a2c
MD
671 *
672 * NOTE: For UP there is no mplock and lwkt_getalltokens()
673 * always succeeds.
8ec60c3f 674 */
f9235b6d
MD
675 if (ntd->td_fairq_accum >= 0 &&
676#ifdef SMP
677 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
678#endif
b37f18d6 679 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 680 ) {
8a8d5d85 681#ifdef SMP
f9235b6d
MD
682 clr_mplock_contention_mask(gd);
683#endif
684 goto havethread;
685 }
686
b37f18d6
MD
687 lmsg = NULL;
688 laddr = NULL;
689
f9235b6d
MD
690#ifdef SMP
691 /* Reload mpheld (it become stale after mplock/token ops) */
692 mpheld = MP_LOCK_HELD();
b37f18d6
MD
693 if (ntd->td_mpcount && mpheld == 0) {
694 lmsg = "mplock";
695 laddr = ntd->td_mplock_stallpc;
696 }
f9235b6d
MD
697#endif
698
699 /*
700 * Coldpath - unable to schedule ntd, continue looking for threads
701 * to schedule. This is only allowed of the (presumably) kernel
702 * thread exhausted its fair share. A kernel thread stuck on
703 * resources does not currently allow a user thread to get in
704 * front of it.
705 */
706#ifdef SMP
707 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
708 (ntd->td_fairq_accum < 0));
6f207a2c
MD
709#else
710 nquserok = 1;
f9235b6d
MD
711#endif
712 nlast = NULL;
713
714 for (;;) {
41a01a4d 715 /*
f9235b6d
MD
716 * If the fair-share scheduler ran out ntd gets moved to the
717 * end and its accumulator will be bumped, if it didn't we
718 * maintain the same queue position.
df6b8ba0 719 *
f9235b6d 720 * nlast keeps track of the last element prior to any moves.
41a01a4d 721 */
f9235b6d
MD
722 if (ntd->td_fairq_accum < 0) {
723 xtd = TAILQ_NEXT(ntd, td_threadq);
724 lwkt_fairq_accumulate(gd, ntd);
725 didaccumulate = 1;
726 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
727 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
728 if (nlast == NULL) {
729 nlast = ntd;
730 if (xtd == NULL)
731 xtd = ntd;
732 }
733 ntd = xtd;
734 } else {
735 ntd = TAILQ_NEXT(ntd, td_threadq);
736 }
a453459d 737
f9235b6d
MD
738 /*
739 * If we exhausted the run list switch to the idle thread.
740 * Since one or more threads had resource acquisition issues
741 * we do not allow the idle thread to halt.
742 *
743 * NOTE: nlast can be NULL.
744 */
745 if (ntd == nlast) {
e0a90d3b 746 cpu_pause();
f9235b6d
MD
747 ntd = &gd->gd_idlethread;
748 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 749#ifdef SMP
f9235b6d
MD
750 set_mplock_contention_mask(gd);
751 cpu_mplock_contested();
752 if (ntd->td_mpcount) {
753 mpheld = MP_LOCK_HELD();
754 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
755 panic("Idle thread %p was holding the BGL!", ntd);
756 if (mpheld == 0) {
757 cpu_pause();
758 break; /* try again from the top, almost */
b9eb1c19 759 }
8a8d5d85 760 }
6f207a2c 761#endif
684a93c4
MD
762
763 /*
f9235b6d
MD
764 * If fairq accumulations occured we do not schedule the
765 * idle thread. This will cause us to try again from
766 * the (almost) top.
684a93c4 767 */
f9235b6d 768 if (didaccumulate)
b37f18d6
MD
769 break; /* try again from the top, almost */
770 if (lmsg)
771 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
772 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 773 goto haveidle;
8a8d5d85 774 }
f9235b6d 775
df6b8ba0 776 /*
f9235b6d 777 * Try to switch to this thread.
6f207a2c
MD
778 *
779 * NOTE: For UP there is no mplock and lwkt_getalltokens()
780 * always succeeds.
df6b8ba0 781 */
77912481
MD
782 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
783 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
f9235b6d
MD
784#ifdef SMP
785 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
8a8d5d85 786#endif
b37f18d6 787 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 788 ) {
a453459d 789#ifdef SMP
f9235b6d
MD
790 clr_mplock_contention_mask(gd);
791#endif
792 goto havethread;
df6b8ba0 793 }
f9235b6d
MD
794#ifdef SMP
795 /* Reload mpheld (it become stale after mplock/token ops) */
796 mpheld = MP_LOCK_HELD();
b37f18d6
MD
797 if (ntd->td_mpcount && mpheld == 0) {
798 lmsg = "mplock";
799 laddr = ntd->td_mplock_stallpc;
800 }
801
f9235b6d
MD
802 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
803 nquserok = 0;
a453459d 804#endif
4b5f931b 805 }
f1d1c3fa 806 }
8a8d5d85
MD
807
808 /*
f9235b6d
MD
809 * Do the actual switch. WARNING: mpheld is stale here.
810 *
811 * We must always decrement td_fairq_accum on non-idle threads just
812 * in case a thread never gets a tick due to being in a continuous
813 * critical section. The page-zeroing code does that.
814 *
815 * If the thread we came up with is a higher or equal priority verses
816 * the thread at the head of the queue we move our thread to the
817 * front. This way we can always check the front of the queue.
818 */
819havethread:
820 ++gd->gd_cnt.v_swtch;
821 --ntd->td_fairq_accum;
822 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
823 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
824 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
825 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
826 }
827havethread_preempted:
828 ;
829 /*
830 * If the new target does not need the MP lock and we are holding it,
831 * release the MP lock. If the new target requires the MP lock we have
832 * already acquired it for the target.
833 *
834 * WARNING: mpheld is stale here.
8a8d5d85 835 */
f9235b6d
MD
836haveidle:
837 KASSERT(ntd->td_critcount,
838 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
839#ifdef SMP
840 if (ntd->td_mpcount == 0 ) {
841 if (MP_LOCK_HELD())
842 cpu_rel_mplock();
843 } else {
a453459d 844 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
845 }
846#endif
94f6d86e
MD
847 if (td != ntd) {
848 ++switch_count;
b2b3ffcd 849#ifdef __x86_64__
f9235b6d
MD
850 {
851 int tos_ok __debugvar = jg_tos_ok(ntd);
852 KKASSERT(tos_ok);
853 }
85514115 854#endif
a1f0fb66 855 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 856 td->td_switch(ntd);
94f6d86e 857 }
37af14fe
MD
858 /* NOTE: current cpu may have changed after switch */
859 crit_exit_quick(td);
8ad65e08
MD
860}
861
b68b7282 862/*
96728c05
MD
863 * Request that the target thread preempt the current thread. Preemption
864 * only works under a specific set of conditions:
b68b7282 865 *
96728c05
MD
866 * - We are not preempting ourselves
867 * - The target thread is owned by the current cpu
868 * - We are not currently being preempted
869 * - The target is not currently being preempted
d3d1cbc8
MD
870 * - We are not holding any spin locks
871 * - The target thread is not holding any tokens
96728c05
MD
872 * - We are able to satisfy the target's MP lock requirements (if any).
873 *
874 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
875 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 876 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
877 * to determine whether preemption is possible (aka usually just the crit
878 * priority of lwkt_schedule() itself).
b68b7282 879 *
26a0694b
MD
880 * XXX at the moment we run the target thread in a critical section during
881 * the preemption in order to prevent the target from taking interrupts
882 * that *WE* can't. Preemption is strictly limited to interrupt threads
883 * and interrupt-like threads, outside of a critical section, and the
884 * preempted source thread will be resumed the instant the target blocks
885 * whether or not the source is scheduled (i.e. preemption is supposed to
886 * be as transparent as possible).
4b5f931b 887 *
8a8d5d85
MD
888 * The target thread inherits our MP count (added to its own) for the
889 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
890 * MP lock during the preemption. Therefore, any preempting targets must be
891 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
892 * out of sync with the physical mp_lock, but we do not have to preserve
893 * the original ownership of the lock if it was out of synch (that is, we
894 * can leave it synchronized on return).
b68b7282
MD
895 */
896void
f9235b6d 897lwkt_preempt(thread_t ntd, int critcount)
b68b7282 898{
46a3f46d 899 struct globaldata *gd = mycpu;
0a3f9b47 900 thread_t td;
8a8d5d85
MD
901#ifdef SMP
902 int mpheld;
57c254db 903 int savecnt;
8a8d5d85 904#endif
b68b7282 905
26a0694b 906 /*
96728c05
MD
907 * The caller has put us in a critical section. We can only preempt
908 * if the caller of the caller was not in a critical section (basically
f9235b6d 909 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 910 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 911 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
912 *
913 * YYY The target thread must be in a critical section (else it must
914 * inherit our critical section? I dunno yet).
41a01a4d 915 *
0a3f9b47 916 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 917 */
f9235b6d 918 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 919
0a3f9b47 920 td = gd->gd_curthread;
f9235b6d 921 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
922 ++preempt_miss;
923 return;
924 }
f9235b6d 925 if (td->td_critcount > critcount) {
96728c05 926 ++preempt_miss;
8ec60c3f 927 need_lwkt_resched();
96728c05
MD
928 return;
929 }
930#ifdef SMP
46a3f46d 931 if (ntd->td_gd != gd) {
96728c05 932 ++preempt_miss;
8ec60c3f 933 need_lwkt_resched();
96728c05
MD
934 return;
935 }
936#endif
41a01a4d 937 /*
77912481
MD
938 * We don't have to check spinlocks here as they will also bump
939 * td_critcount.
d3d1cbc8
MD
940 *
941 * Do not try to preempt if the target thread is holding any tokens.
942 * We could try to acquire the tokens but this case is so rare there
943 * is no need to support it.
41a01a4d 944 */
77912481
MD
945 KKASSERT(gd->gd_spinlock_rd == NULL);
946 KKASSERT(gd->gd_spinlocks_wr == 0);
947
3b998fa9 948 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
949 ++preempt_miss;
950 need_lwkt_resched();
951 return;
952 }
26a0694b
MD
953 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
954 ++preempt_weird;
8ec60c3f 955 need_lwkt_resched();
26a0694b
MD
956 return;
957 }
958 if (ntd->td_preempted) {
4b5f931b 959 ++preempt_hit;
8ec60c3f 960 need_lwkt_resched();
26a0694b 961 return;
b68b7282 962 }
8a8d5d85 963#ifdef SMP
a2a5ad0d
MD
964 /*
965 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
966 * to or from the MP lock. In this case td_mpcount will be pre-disposed
967 * (non-zero) but not actually synchronized with the actual state of the
968 * lock. We can use it to imply an MP lock requirement for the
969 * preemption but we cannot use it to test whether we hold the MP lock
970 * or not.
a2a5ad0d 971 */
96728c05 972 savecnt = td->td_mpcount;
71ef2f5c 973 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
974 ntd->td_mpcount += td->td_mpcount;
975 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
976 ntd->td_mpcount -= td->td_mpcount;
977 ++preempt_miss;
8ec60c3f 978 need_lwkt_resched();
8a8d5d85
MD
979 return;
980 }
981#endif
26a0694b 982
8ec60c3f
MD
983 /*
984 * Since we are able to preempt the current thread, there is no need to
985 * call need_lwkt_resched().
986 */
26a0694b
MD
987 ++preempt_hit;
988 ntd->td_preempted = td;
989 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 990 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 991 td->td_switch(ntd);
b9eb1c19 992
26a0694b 993 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
994#ifdef SMP
995 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
996 mpheld = MP_LOCK_HELD();
997 if (mpheld && td->td_mpcount == 0)
96728c05 998 cpu_rel_mplock();
71ef2f5c 999 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
1000 panic("lwkt_preempt(): MP lock was not held through");
1001#endif
26a0694b
MD
1002 ntd->td_preempted = NULL;
1003 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1004}
1005
f1d1c3fa 1006/*
faaeffac 1007 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 1008 *
faaeffac
MD
1009 * td_nest_count prevents deep nesting via splz() or doreti() which
1010 * might otherwise blow out the kernel stack. Note that except for
1011 * this special case, we MUST call splz() here to handle any
1012 * pending ints, particularly after we switch, or we might accidently
1013 * halt the cpu with interrupts pending.
ef0fdad1 1014 *
f1d1c3fa
MD
1015 * (self contained on a per cpu basis)
1016 */
1017void
faaeffac 1018splz_check(void)
f1d1c3fa 1019{
7966cb69
MD
1020 globaldata_t gd = mycpu;
1021 thread_t td = gd->gd_curthread;
ef0fdad1 1022
f9235b6d 1023 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
f1d1c3fa 1024 splz();
f1d1c3fa
MD
1025}
1026
8ad65e08 1027/*
f9235b6d
MD
1028 * This function is used to negotiate a passive release of the current
1029 * process/lwp designation with the user scheduler, allowing the user
1030 * scheduler to schedule another user thread. The related kernel thread
1031 * (curthread) continues running in the released state.
8ad65e08
MD
1032 */
1033void
f9235b6d 1034lwkt_passive_release(struct thread *td)
8ad65e08 1035{
f9235b6d
MD
1036 struct lwp *lp = td->td_lwp;
1037
1038 td->td_release = NULL;
1039 lwkt_setpri_self(TDPRI_KERN_USER);
1040 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1041}
1042
f9235b6d 1043
3824f392 1044/*
f9235b6d
MD
1045 * This implements a normal yield. This routine is virtually a nop if
1046 * there is nothing to yield to but it will always run any pending interrupts
1047 * if called from a critical section.
1048 *
1049 * This yield is designed for kernel threads without a user context.
1050 *
1051 * (self contained on a per cpu basis)
3824f392
MD
1052 */
1053void
f9235b6d 1054lwkt_yield(void)
3824f392 1055{
f9235b6d
MD
1056 globaldata_t gd = mycpu;
1057 thread_t td = gd->gd_curthread;
1058 thread_t xtd;
3824f392 1059
f9235b6d
MD
1060 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1061 splz();
1062 if (td->td_fairq_accum < 0) {
1063 lwkt_schedule_self(curthread);
1064 lwkt_switch();
1065 } else {
1066 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1067 if (xtd && xtd->td_pri > td->td_pri) {
1068 lwkt_schedule_self(curthread);
1069 lwkt_switch();
1070 }
1071 }
3824f392
MD
1072}
1073
1074/*
f9235b6d
MD
1075 * This yield is designed for kernel threads with a user context.
1076 *
1077 * The kernel acting on behalf of the user is potentially cpu-bound,
1078 * this function will efficiently allow other threads to run and also
1079 * switch to other processes by releasing.
3824f392
MD
1080 *
1081 * The lwkt_user_yield() function is designed to have very low overhead
1082 * if no yield is determined to be needed.
1083 */
1084void
1085lwkt_user_yield(void)
1086{
f9235b6d
MD
1087 globaldata_t gd = mycpu;
1088 thread_t td = gd->gd_curthread;
1089
1090 /*
1091 * Always run any pending interrupts in case we are in a critical
1092 * section.
1093 */
1094 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1095 splz();
3824f392
MD
1096
1097#ifdef SMP
1098 /*
1099 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1100 * kernel can prevent other cpus from servicing interrupt threads
1101 * which still require the MP lock (which is a lot of them). This
1102 * has a chaining effect since if the interrupt is blocked, so is
1103 * the event, so normal scheduling will not pick up on the problem.
1104 */
684a93c4
MD
1105 if (mp_lock_contention_mask && td->td_mpcount) {
1106 yield_mplock(td);
3824f392
MD
1107 }
1108#endif
1109
1110 /*
f9235b6d
MD
1111 * Switch (which forces a release) if another kernel thread needs
1112 * the cpu, if userland wants us to resched, or if our kernel
1113 * quantum has run out.
3824f392 1114 */
f9235b6d
MD
1115 if (lwkt_resched_wanted() ||
1116 user_resched_wanted() ||
1117 td->td_fairq_accum < 0)
1118 {
3824f392 1119 lwkt_switch();
3824f392
MD
1120 }
1121
f9235b6d 1122#if 0
3824f392 1123 /*
f9235b6d
MD
1124 * Reacquire the current process if we are released.
1125 *
1126 * XXX not implemented atm. The kernel may be holding locks and such,
1127 * so we want the thread to continue to receive cpu.
3824f392 1128 */
f9235b6d
MD
1129 if (td->td_release == NULL && lp) {
1130 lp->lwp_proc->p_usched->acquire_curproc(lp);
1131 td->td_release = lwkt_passive_release;
1132 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1133 }
f9235b6d 1134#endif
b9eb1c19
MD
1135}
1136
8ad65e08 1137/*
f1d1c3fa
MD
1138 * Generic schedule. Possibly schedule threads belonging to other cpus and
1139 * deal with threads that might be blocked on a wait queue.
1140 *
0a3f9b47
MD
1141 * We have a little helper inline function which does additional work after
1142 * the thread has been enqueued, including dealing with preemption and
1143 * setting need_lwkt_resched() (which prevents the kernel from returning
1144 * to userland until it has processed higher priority threads).
6330a558
MD
1145 *
1146 * It is possible for this routine to be called after a failed _enqueue
1147 * (due to the target thread migrating, sleeping, or otherwise blocked).
1148 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1149 *
1150 * reschedok is an optimized constant propagated from lwkt_schedule() or
1151 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1152 * reschedule to be requested if the target thread has a higher priority.
1153 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1154 * be 0, prevented undesired reschedules.
8ad65e08 1155 */
0a3f9b47
MD
1156static __inline
1157void
f9235b6d 1158_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1159{
b9eb1c19 1160 thread_t otd;
c730be20 1161
6330a558 1162 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1163 if (ntd->td_preemptable && reschedok) {
f9235b6d 1164 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1165 } else if (reschedok) {
b9eb1c19 1166 otd = curthread;
f9235b6d 1167 if (ntd->td_pri > otd->td_pri)
c730be20 1168 need_lwkt_resched();
6330a558 1169 }
f9235b6d
MD
1170
1171 /*
1172 * Give the thread a little fair share scheduler bump if it
1173 * has been asleep for a while. This is primarily to avoid
1174 * a degenerate case for interrupt threads where accumulator
1175 * crosses into negative territory unnecessarily.
1176 */
1177 if (ntd->td_fairq_lticks != ticks) {
1178 ntd->td_fairq_lticks = ticks;
1179 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1180 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1181 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1182 }
0a3f9b47
MD
1183 }
1184}
1185
361d01dd 1186static __inline
8ad65e08 1187void
361d01dd 1188_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1189{
37af14fe
MD
1190 globaldata_t mygd = mycpu;
1191
41a01a4d 1192 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1193 crit_enter_gd(mygd);
9388413d 1194 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1195 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1196 _lwkt_enqueue(td);
1197 } else {
f1d1c3fa 1198 /*
7cd8d145
MD
1199 * If we own the thread, there is no race (since we are in a
1200 * critical section). If we do not own the thread there might
1201 * be a race but the target cpu will deal with it.
f1d1c3fa 1202 */
0f7a3396 1203#ifdef SMP
7cd8d145 1204 if (td->td_gd == mygd) {
9d265729 1205 _lwkt_enqueue(td);
f9235b6d 1206 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1207 } else {
e381e77c 1208 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1209 }
0f7a3396 1210#else
7cd8d145 1211 _lwkt_enqueue(td);
f9235b6d 1212 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1213#endif
8ad65e08 1214 }
37af14fe 1215 crit_exit_gd(mygd);
8ad65e08
MD
1216}
1217
361d01dd
MD
1218void
1219lwkt_schedule(thread_t td)
1220{
1221 _lwkt_schedule(td, 1);
1222}
1223
1224void
1225lwkt_schedule_noresched(thread_t td)
1226{
1227 _lwkt_schedule(td, 0);
1228}
1229
88ebb169
SW
1230#ifdef SMP
1231
e381e77c
MD
1232/*
1233 * When scheduled remotely if frame != NULL the IPIQ is being
1234 * run via doreti or an interrupt then preemption can be allowed.
1235 *
1236 * To allow preemption we have to drop the critical section so only
1237 * one is present in _lwkt_schedule_post.
1238 */
1239static void
1240lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1241{
1242 thread_t td = curthread;
1243 thread_t ntd = arg;
1244
1245 if (frame && ntd->td_preemptable) {
1246 crit_exit_noyield(td);
1247 _lwkt_schedule(ntd, 1);
1248 crit_enter_quick(td);
1249 } else {
1250 _lwkt_schedule(ntd, 1);
1251 }
1252}
1253
d9eea1a5 1254/*
52eedfb5
MD
1255 * Thread migration using a 'Pull' method. The thread may or may not be
1256 * the current thread. It MUST be descheduled and in a stable state.
1257 * lwkt_giveaway() must be called on the cpu owning the thread.
1258 *
1259 * At any point after lwkt_giveaway() is called, the target cpu may
1260 * 'pull' the thread by calling lwkt_acquire().
1261 *
ae8e83e6
MD
1262 * We have to make sure the thread is not sitting on a per-cpu tsleep
1263 * queue or it will blow up when it moves to another cpu.
1264 *
52eedfb5 1265 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1266 */
52eedfb5
MD
1267void
1268lwkt_giveaway(thread_t td)
1269{
3b4192fb 1270 globaldata_t gd = mycpu;
52eedfb5 1271
3b4192fb
MD
1272 crit_enter_gd(gd);
1273 if (td->td_flags & TDF_TSLEEPQ)
1274 tsleep_remove(td);
1275 KKASSERT(td->td_gd == gd);
1276 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1277 td->td_flags |= TDF_MIGRATING;
1278 crit_exit_gd(gd);
52eedfb5
MD
1279}
1280
a2a5ad0d
MD
1281void
1282lwkt_acquire(thread_t td)
1283{
37af14fe
MD
1284 globaldata_t gd;
1285 globaldata_t mygd;
a2a5ad0d 1286
52eedfb5 1287 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1288 gd = td->td_gd;
37af14fe 1289 mygd = mycpu;
52eedfb5 1290 if (gd != mycpu) {
35238fa5 1291 cpu_lfence();
52eedfb5 1292 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1293 crit_enter_gd(mygd);
df910c23
MD
1294 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1295#ifdef SMP
1296 lwkt_process_ipiq();
1297#endif
52eedfb5 1298 cpu_lfence();
df910c23 1299 }
37af14fe 1300 td->td_gd = mygd;
52eedfb5
MD
1301 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1302 td->td_flags &= ~TDF_MIGRATING;
1303 crit_exit_gd(mygd);
1304 } else {
1305 crit_enter_gd(mygd);
1306 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1307 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1308 crit_exit_gd(mygd);
a2a5ad0d
MD
1309 }
1310}
1311
52eedfb5
MD
1312#endif
1313
f1d1c3fa
MD
1314/*
1315 * Generic deschedule. Descheduling threads other then your own should be
1316 * done only in carefully controlled circumstances. Descheduling is
1317 * asynchronous.
1318 *
1319 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1320 */
1321void
1322lwkt_deschedule(thread_t td)
1323{
f1d1c3fa 1324 crit_enter();
b8a98473 1325#ifdef SMP
f1d1c3fa
MD
1326 if (td == curthread) {
1327 _lwkt_dequeue(td);
1328 } else {
a72187e9 1329 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1330 _lwkt_dequeue(td);
1331 } else {
b8a98473 1332 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1333 }
1334 }
b8a98473
MD
1335#else
1336 _lwkt_dequeue(td);
1337#endif
f1d1c3fa
MD
1338 crit_exit();
1339}
1340
4b5f931b
MD
1341/*
1342 * Set the target thread's priority. This routine does not automatically
1343 * switch to a higher priority thread, LWKT threads are not designed for
1344 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1345 */
1346void
1347lwkt_setpri(thread_t td, int pri)
1348{
a72187e9 1349 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1350 if (td->td_pri != pri) {
1351 KKASSERT(pri >= 0);
1352 crit_enter();
1353 if (td->td_flags & TDF_RUNQ) {
1354 _lwkt_dequeue(td);
1355 td->td_pri = pri;
1356 _lwkt_enqueue(td);
1357 } else {
1358 td->td_pri = pri;
1359 }
1360 crit_exit();
26a0694b 1361 }
26a0694b
MD
1362}
1363
03bd0a5e
MD
1364/*
1365 * Set the initial priority for a thread prior to it being scheduled for
1366 * the first time. The thread MUST NOT be scheduled before or during
1367 * this call. The thread may be assigned to a cpu other then the current
1368 * cpu.
1369 *
1370 * Typically used after a thread has been created with TDF_STOPPREQ,
1371 * and before the thread is initially scheduled.
1372 */
1373void
1374lwkt_setpri_initial(thread_t td, int pri)
1375{
1376 KKASSERT(pri >= 0);
1377 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1378 td->td_pri = pri;
03bd0a5e
MD
1379}
1380
26a0694b
MD
1381void
1382lwkt_setpri_self(int pri)
1383{
1384 thread_t td = curthread;
1385
4b5f931b
MD
1386 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1387 crit_enter();
1388 if (td->td_flags & TDF_RUNQ) {
1389 _lwkt_dequeue(td);
f9235b6d 1390 td->td_pri = pri;
4b5f931b
MD
1391 _lwkt_enqueue(td);
1392 } else {
f9235b6d 1393 td->td_pri = pri;
4b5f931b
MD
1394 }
1395 crit_exit();
1396}
1397
f9235b6d
MD
1398/*
1399 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1400 *
1401 * Example: two competing threads, same priority N. decrement by (2*N)
1402 * increment by N*8, each thread will get 4 ticks.
1403 */
1404void
1405lwkt_fairq_schedulerclock(thread_t td)
1406{
1407 if (fairq_enable) {
1408 while (td) {
1409 if (td != &td->td_gd->gd_idlethread) {
1410 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1411 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1412 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1413 if (td->td_fairq_accum < 0)
1414 need_lwkt_resched();
1415 td->td_fairq_lticks = ticks;
1416 }
1417 td = td->td_preempted;
1418 }
1419 }
1420}
1421
1422static void
1423lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1424{
1425 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1426 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1427 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1428}
1429
5d21b981 1430/*
52eedfb5
MD
1431 * Migrate the current thread to the specified cpu.
1432 *
1433 * This is accomplished by descheduling ourselves from the current cpu,
1434 * moving our thread to the tdallq of the target cpu, IPI messaging the
1435 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1436 * races while the thread is being migrated.
ae8e83e6
MD
1437 *
1438 * We must be sure to remove ourselves from the current cpu's tsleepq
1439 * before potentially moving to another queue. The thread can be on
1440 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1441 */
3d28ff59 1442#ifdef SMP
5d21b981 1443static void lwkt_setcpu_remote(void *arg);
3d28ff59 1444#endif
5d21b981
MD
1445
1446void
1447lwkt_setcpu_self(globaldata_t rgd)
1448{
1449#ifdef SMP
1450 thread_t td = curthread;
1451
1452 if (td->td_gd != rgd) {
1453 crit_enter_quick(td);
ae8e83e6 1454 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1455 tsleep_remove(td);
5d21b981
MD
1456 td->td_flags |= TDF_MIGRATING;
1457 lwkt_deschedule_self(td);
52eedfb5 1458 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1459 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1460 lwkt_switch();
1461 /* we are now on the target cpu */
52eedfb5 1462 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1463 crit_exit_quick(td);
1464 }
1465#endif
1466}
1467
ecdefdda
MD
1468void
1469lwkt_migratecpu(int cpuid)
1470{
1471#ifdef SMP
1472 globaldata_t rgd;
1473
1474 rgd = globaldata_find(cpuid);
1475 lwkt_setcpu_self(rgd);
1476#endif
1477}
1478
5d21b981
MD
1479/*
1480 * Remote IPI for cpu migration (called while in a critical section so we
1481 * do not have to enter another one). The thread has already been moved to
1482 * our cpu's allq, but we must wait for the thread to be completely switched
1483 * out on the originating cpu before we schedule it on ours or the stack
1484 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1485 * change to main memory.
1486 *
1487 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1488 * against wakeups. It is best if this interface is used only when there
1489 * are no pending events that might try to schedule the thread.
1490 */
3d28ff59 1491#ifdef SMP
5d21b981
MD
1492static void
1493lwkt_setcpu_remote(void *arg)
1494{
1495 thread_t td = arg;
1496 globaldata_t gd = mycpu;
1497
df910c23
MD
1498 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1499#ifdef SMP
1500 lwkt_process_ipiq();
1501#endif
35238fa5 1502 cpu_lfence();
df910c23 1503 }
5d21b981 1504 td->td_gd = gd;
35238fa5 1505 cpu_sfence();
5d21b981 1506 td->td_flags &= ~TDF_MIGRATING;
9388413d 1507 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1508 _lwkt_enqueue(td);
1509}
3d28ff59 1510#endif
5d21b981 1511
553ea3c8 1512struct lwp *
4b5f931b
MD
1513lwkt_preempted_proc(void)
1514{
73e4f7b9 1515 thread_t td = curthread;
4b5f931b
MD
1516 while (td->td_preempted)
1517 td = td->td_preempted;
553ea3c8 1518 return(td->td_lwp);
4b5f931b
MD
1519}
1520
99df837e
MD
1521/*
1522 * Create a kernel process/thread/whatever. It shares it's address space
1523 * with proc0 - ie: kernel only.
1524 *
365fa13f
MD
1525 * NOTE! By default new threads are created with the MP lock held. A
1526 * thread which does not require the MP lock should release it by calling
1527 * rel_mplock() at the start of the new thread.
99df837e
MD
1528 */
1529int
1530lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1531 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1532 const char *fmt, ...)
99df837e 1533{
73e4f7b9 1534 thread_t td;
e2565a42 1535 __va_list ap;
99df837e 1536
d3d32139 1537 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1538 tdflags);
a2a5ad0d
MD
1539 if (tdp)
1540 *tdp = td;
709799ea 1541 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1542
1543 /*
1544 * Set up arg0 for 'ps' etc
1545 */
e2565a42 1546 __va_start(ap, fmt);
379210cb 1547 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1548 __va_end(ap);
99df837e
MD
1549
1550 /*
1551 * Schedule the thread to run
1552 */
ef0fdad1
MD
1553 if ((td->td_flags & TDF_STOPREQ) == 0)
1554 lwkt_schedule(td);
1555 else
1556 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1557 return 0;
1558}
1559
1560/*
1561 * Destroy an LWKT thread. Warning! This function is not called when
1562 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1563 * uses a different reaping mechanism.
1564 */
1565void
1566lwkt_exit(void)
1567{
1568 thread_t td = curthread;
c070746a 1569 thread_t std;
8826f33a 1570 globaldata_t gd;
99df837e
MD
1571
1572 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1573 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1574 caps_exit(td);
c070746a
MD
1575
1576 /*
1577 * Get us into a critical section to interlock gd_freetd and loop
1578 * until we can get it freed.
1579 *
1580 * We have to cache the current td in gd_freetd because objcache_put()ing
1581 * it would rip it out from under us while our thread is still active.
1582 */
1583 gd = mycpu;
37af14fe 1584 crit_enter_quick(td);
c070746a
MD
1585 while ((std = gd->gd_freetd) != NULL) {
1586 gd->gd_freetd = NULL;
1587 objcache_put(thread_cache, std);
1588 }
3b4192fb
MD
1589
1590 /*
1591 * Remove thread resources from kernel lists and deschedule us for
1592 * the last time.
1593 */
1594 if (td->td_flags & TDF_TSLEEPQ)
1595 tsleep_remove(td);
79eae878 1596 biosched_done(td);
f8abf63c 1597 dsched_exit_thread(td);
37af14fe 1598 lwkt_deschedule_self(td);
e56e4dea 1599 lwkt_remove_tdallq(td);
c070746a
MD
1600 if (td->td_flags & TDF_ALLOCATED_THREAD)
1601 gd->gd_freetd = td;
99df837e
MD
1602 cpu_thread_exit();
1603}
1604
e56e4dea
MD
1605void
1606lwkt_remove_tdallq(thread_t td)
1607{
1608 KKASSERT(td->td_gd == mycpu);
1609 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1610}
1611
2d93b37a
MD
1612void
1613crit_panic(void)
1614{
1615 thread_t td = curthread;
850634cc 1616 int lcrit = td->td_critcount;
2d93b37a 1617
850634cc
AH
1618 td->td_critcount = 0;
1619 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
2d93b37a
MD
1620}
1621
d165e668
MD
1622#ifdef SMP
1623
bd8015ca
MD
1624/*
1625 * Called from debugger/panic on cpus which have been stopped. We must still
1626 * process the IPIQ while stopped, even if we were stopped while in a critical
1627 * section (XXX).
1628 *
1629 * If we are dumping also try to process any pending interrupts. This may
1630 * or may not work depending on the state of the cpu at the point it was
1631 * stopped.
1632 */
1633void
1634lwkt_smp_stopped(void)
1635{
1636 globaldata_t gd = mycpu;
1637
1638 crit_enter_gd(gd);
1639 if (dumping) {
1640 lwkt_process_ipiq();
1641 splz();
1642 } else {
1643 lwkt_process_ipiq();
1644 }
1645 crit_exit_gd(gd);
1646}
1647
d165e668 1648#endif