kernel - rewrite the LWKT scheduler's priority mechanism
[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>
47#include <sys/queue.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f6bf3af1 52#include <sys/caps.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
684a93c4 58#include <sys/mplock2.h>
f1d1c3fa 59
8c72e3d5
AH
60#include <sys/dsched.h>
61
7d0bac62
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <vm/vm_kern.h>
65#include <vm/vm_object.h>
66#include <vm/vm_page.h>
67#include <vm/vm_map.h>
68#include <vm/vm_pager.h>
69#include <vm/vm_extern.h>
7d0bac62 70
99df837e 71#include <machine/stdarg.h>
96728c05 72#include <machine/smp.h>
99df837e 73
d850923c
AE
74#if !defined(KTR_CTXSW)
75#define KTR_CTXSW KTR_ALL
76#endif
77KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
78KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
79 sizeof(int) + sizeof(struct thread *));
80KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
81 sizeof(int) + sizeof(struct thread *));
82KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
83 sizeof (struct thread *) + sizeof(char *));
84KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 85
40aaf5fc
NT
86static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
87
0f7a3396
MD
88#ifdef INVARIANTS
89static int panic_on_cscount = 0;
90#endif
05220613
MD
91static __int64_t switch_count = 0;
92static __int64_t preempt_hit = 0;
93static __int64_t preempt_miss = 0;
94static __int64_t preempt_weird = 0;
f64b567c 95static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 96static int lwkt_use_spin_port;
40aaf5fc 97static struct objcache *thread_cache;
05220613 98
88ebb169 99#ifdef SMP
e381e77c 100static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 101#endif
f9235b6d 102static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 103
0855a2af
JG
104extern void cpu_heavy_restore(void);
105extern void cpu_lwkt_restore(void);
106extern void cpu_kthread_restore(void);
107extern void cpu_idle_restore(void);
108
b2b3ffcd 109#ifdef __x86_64__
85514115
MD
110
111static int
0855a2af
JG
112jg_tos_ok(struct thread *td)
113{
85514115
MD
114 void *tos;
115 int tos_ok;
116
0855a2af
JG
117 if (td == NULL) {
118 return 1;
119 }
120 KKASSERT(td->td_sp != NULL);
85514115
MD
121 tos = ((void **)td->td_sp)[0];
122 tos_ok = 0;
123 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
124 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
125 tos_ok = 1;
126 }
127 return tos_ok;
128}
129
85514115
MD
130#endif
131
fb0f29c4
MD
132/*
133 * We can make all thread ports use the spin backend instead of the thread
134 * backend. This should only be set to debug the spin backend.
135 */
136TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
137
0f7a3396
MD
138#ifdef INVARIANTS
139SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
140#endif
4b5f931b 141SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
142SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
143 "Successful preemption events");
144SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
145 "Failed preemption events");
26a0694b 146SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
147#ifdef INVARIANTS
148SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
149 &token_contention_count, 0, "spinning due to token contention");
38717797 150#endif
f9235b6d
MD
151static int fairq_enable = 1;
152SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
05220613 153
4b5f931b
MD
154/*
155 * These helper procedures handle the runq, they can only be called from
156 * within a critical section.
75cdbe6c
MD
157 *
158 * WARNING! Prior to SMP being brought up it is possible to enqueue and
159 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
160 * instead of 'mycpu' when referencing the globaldata structure. Once
161 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 162 */
f1d1c3fa
MD
163static __inline
164void
165_lwkt_dequeue(thread_t td)
166{
167 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 168 struct globaldata *gd = td->td_gd;
4b5f931b 169
f1d1c3fa 170 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
171 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
172 gd->gd_fairq_total_pri -= td->td_pri;
173 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
174 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
175 }
176}
177
f9235b6d
MD
178/*
179 * Priority enqueue.
180 *
181 * NOTE: There are a limited number of lwkt threads runnable since user
182 * processes only schedule one at a time per cpu.
183 */
f1d1c3fa
MD
184static __inline
185void
186_lwkt_enqueue(thread_t td)
187{
f9235b6d
MD
188 thread_t xtd;
189
7f5d7ed7 190 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 191 struct globaldata *gd = td->td_gd;
4b5f931b 192
f1d1c3fa 193 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
194 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
195 if (xtd == NULL) {
196 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
197 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
198 } else {
199 while (xtd && xtd->td_pri > td->td_pri)
200 xtd = TAILQ_NEXT(xtd, td_threadq);
201 if (xtd)
202 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
203 else
204 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
205 }
206 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
207 }
208}
8ad65e08 209
40aaf5fc
NT
210static __boolean_t
211_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
212{
213 struct thread *td = (struct thread *)obj;
214
215 td->td_kstack = NULL;
216 td->td_kstack_size = 0;
217 td->td_flags = TDF_ALLOCATED_THREAD;
218 return (1);
219}
220
221static void
222_lwkt_thread_dtor(void *obj, void *privdata)
223{
224 struct thread *td = (struct thread *)obj;
225
226 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
227 ("_lwkt_thread_dtor: not allocated from objcache"));
228 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
229 td->td_kstack_size > 0,
230 ("_lwkt_thread_dtor: corrupted stack"));
231 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
232}
233
234/*
235 * Initialize the lwkt s/system.
236 */
237void
238lwkt_init(void)
239{
240 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
241 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
242 NULL, CACHE_NTHREADS/2,
243 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
244}
245
37af14fe
MD
246/*
247 * Schedule a thread to run. As the current thread we can always safely
248 * schedule ourselves, and a shortcut procedure is provided for that
249 * function.
250 *
251 * (non-blocking, self contained on a per cpu basis)
252 */
253void
254lwkt_schedule_self(thread_t td)
255{
256 crit_enter_quick(td);
f9235b6d
MD
257 KASSERT(td != &td->td_gd->gd_idlethread,
258 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 259 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 260 _lwkt_enqueue(td);
37af14fe
MD
261 crit_exit_quick(td);
262}
263
264/*
265 * Deschedule a thread.
266 *
267 * (non-blocking, self contained on a per cpu basis)
268 */
269void
270lwkt_deschedule_self(thread_t td)
271{
272 crit_enter_quick(td);
37af14fe
MD
273 _lwkt_dequeue(td);
274 crit_exit_quick(td);
275}
276
8ad65e08
MD
277/*
278 * LWKTs operate on a per-cpu basis
279 *
73e4f7b9 280 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
281 */
282void
283lwkt_gdinit(struct globaldata *gd)
284{
f9235b6d 285 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 286 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
287}
288
7d0bac62
MD
289/*
290 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
291 * or LWKT start address before it can be scheduled. If the target cpu is
292 * -1 the thread will be created on the current cpu.
0cfcada1
MD
293 *
294 * If you intend to create a thread without a process context this function
295 * does everything except load the startup and switcher function.
7d0bac62
MD
296 */
297thread_t
d3d32139 298lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 299{
c070746a 300 globaldata_t gd = mycpu;
99df837e 301 void *stack;
7d0bac62 302
c070746a
MD
303 /*
304 * If static thread storage is not supplied allocate a thread. Reuse
305 * a cached free thread if possible. gd_freetd is used to keep an exiting
306 * thread intact through the exit.
307 */
ef0fdad1 308 if (td == NULL) {
c070746a
MD
309 if ((td = gd->gd_freetd) != NULL)
310 gd->gd_freetd = NULL;
311 else
312 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
313 KASSERT((td->td_flags &
314 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
315 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
316 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 317 }
c070746a
MD
318
319 /*
320 * Try to reuse cached stack.
321 */
f470d0c8
MD
322 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
323 if (flags & TDF_ALLOCATED_STACK) {
e4846942 324 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
325 stack = NULL;
326 }
327 }
328 if (stack == NULL) {
e4846942 329 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 330 flags |= TDF_ALLOCATED_STACK;
99df837e 331 }
75cdbe6c 332 if (cpu < 0)
c070746a 333 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 334 else
f470d0c8 335 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 336 return(td);
7d0bac62
MD
337}
338
339/*
340 * Initialize a preexisting thread structure. This function is used by
341 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
342 *
f8c3996b
MD
343 * All threads start out in a critical section at a priority of
344 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
345 * appropriate. This function may send an IPI message when the
346 * requested cpu is not the current cpu and consequently gd_tdallq may
347 * not be initialized synchronously from the point of view of the originating
348 * cpu.
349 *
350 * NOTE! we have to be careful in regards to creating threads for other cpus
351 * if SMP has not yet been activated.
7d0bac62 352 */
41a01a4d
MD
353#ifdef SMP
354
75cdbe6c
MD
355static void
356lwkt_init_thread_remote(void *arg)
357{
358 thread_t td = arg;
359
52eedfb5
MD
360 /*
361 * Protected by critical section held by IPI dispatch
362 */
75cdbe6c
MD
363 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
364}
365
41a01a4d
MD
366#endif
367
7d0bac62 368void
f470d0c8
MD
369lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
370 struct globaldata *gd)
7d0bac62 371{
37af14fe
MD
372 globaldata_t mygd = mycpu;
373
99df837e
MD
374 bzero(td, sizeof(struct thread));
375 td->td_kstack = stack;
f470d0c8 376 td->td_kstack_size = stksize;
d3d32139 377 td->td_flags = flags;
26a0694b 378 td->td_gd = gd;
f9235b6d
MD
379 td->td_pri = TDPRI_KERN_DAEMON;
380 td->td_critcount = 1;
3b998fa9 381 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
382#ifdef SMP
383 if ((flags & TDF_MPSAFE) == 0)
384 td->td_mpcount = 1;
385#endif
fb0f29c4
MD
386 if (lwkt_use_spin_port)
387 lwkt_initport_spin(&td->td_msgport);
388 else
389 lwkt_initport_thread(&td->td_msgport, td);
99df837e 390 pmap_init_thread(td);
0f7a3396 391#ifdef SMP
5d21b981
MD
392 /*
393 * Normally initializing a thread for a remote cpu requires sending an
394 * IPI. However, the idlethread is setup before the other cpus are
395 * activated so we have to treat it as a special case. XXX manipulation
396 * of gd_tdallq requires the BGL.
397 */
398 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 399 crit_enter_gd(mygd);
75cdbe6c 400 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 401 crit_exit_gd(mygd);
75cdbe6c 402 } else {
2db3b277 403 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 404 }
0f7a3396 405#else
37af14fe 406 crit_enter_gd(mygd);
0f7a3396 407 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 408 crit_exit_gd(mygd);
0f7a3396 409#endif
8c72e3d5
AH
410
411 dsched_new_thread(td);
73e4f7b9
MD
412}
413
414void
415lwkt_set_comm(thread_t td, const char *ctl, ...)
416{
e2565a42 417 __va_list va;
73e4f7b9 418
e2565a42 419 __va_start(va, ctl);
379210cb 420 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 421 __va_end(va);
e7c0dbba 422 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
423}
424
99df837e 425void
73e4f7b9 426lwkt_hold(thread_t td)
99df837e 427{
73e4f7b9
MD
428 ++td->td_refs;
429}
430
431void
432lwkt_rele(thread_t td)
433{
434 KKASSERT(td->td_refs > 0);
435 --td->td_refs;
436}
437
438void
439lwkt_wait_free(thread_t td)
440{
441 while (td->td_refs)
377d4740 442 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
443}
444
445void
446lwkt_free_thread(thread_t td)
447{
d9eea1a5 448 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
449 ("lwkt_free_thread: did not exit! %p", td));
450
40aaf5fc
NT
451 if (td->td_flags & TDF_ALLOCATED_THREAD) {
452 objcache_put(thread_cache, td);
453 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
454 /* client-allocated struct with internally allocated stack */
455 KASSERT(td->td_kstack && td->td_kstack_size > 0,
456 ("lwkt_free_thread: corrupted stack"));
457 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
458 td->td_kstack = NULL;
459 td->td_kstack_size = 0;
99df837e 460 }
e7c0dbba 461 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
462}
463
464
8ad65e08
MD
465/*
466 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
467 * switch to the idlethread. Switching must occur within a critical
468 * section to avoid races with the scheduling queue.
469 *
470 * We always have full control over our cpu's run queue. Other cpus
471 * that wish to manipulate our queue must use the cpu_*msg() calls to
472 * talk to our cpu, so a critical section is all that is needed and
473 * the result is very, very fast thread switching.
474 *
96728c05
MD
475 * The LWKT scheduler uses a fixed priority model and round-robins at
476 * each priority level. User process scheduling is a totally
477 * different beast and LWKT priorities should not be confused with
478 * user process priorities.
f1d1c3fa 479 *
96728c05
MD
480 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
481 * cleans it up. Note that the td_switch() function cannot do anything that
482 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
483 * the target thread (not the current thread). It's nice to have a scheduler
484 * that does not need the MP lock to work because it allows us to do some
485 * really cool high-performance MP lock optimizations.
69d78e99
MD
486 *
487 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
488 * is not called by the current thread in the preemption case, only when
489 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
490 */
491void
492lwkt_switch(void)
493{
37af14fe
MD
494 globaldata_t gd = mycpu;
495 thread_t td = gd->gd_curthread;
8ad65e08 496 thread_t ntd;
f9235b6d
MD
497 thread_t xtd;
498 thread_t nlast;
8a8d5d85 499#ifdef SMP
f9235b6d 500 int nquserok;
8a8d5d85
MD
501 int mpheld;
502#endif
f9235b6d 503 int didaccumulate;
8ad65e08 504
46a3f46d 505 /*
27e88a6e
MD
506 * Switching from within a 'fast' (non thread switched) interrupt or IPI
507 * is illegal. However, we may have to do it anyway if we hit a fatal
508 * kernel trap or we have paniced.
509 *
510 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 511 */
27e88a6e
MD
512 if (gd->gd_intr_nesting_level) {
513 int savegdnest;
514 int savegdtrap;
515
516 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
517 panic("lwkt_switch: cannot switch from within "
518 "a fast interrupt, yet, td %p\n", td);
519 } else {
520 savegdnest = gd->gd_intr_nesting_level;
521 savegdtrap = gd->gd_trap_nesting_level;
522 gd->gd_intr_nesting_level = 0;
523 gd->gd_trap_nesting_level = 0;
a7422615
MD
524 if ((td->td_flags & TDF_PANICWARN) == 0) {
525 td->td_flags |= TDF_PANICWARN;
6ea70f76 526 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 527 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 528 print_backtrace(-1);
a7422615 529 }
27e88a6e
MD
530 lwkt_switch();
531 gd->gd_intr_nesting_level = savegdnest;
532 gd->gd_trap_nesting_level = savegdtrap;
533 return;
534 }
96728c05 535 }
ef0fdad1 536
cb973d15
MD
537 /*
538 * Passive release (used to transition from user to kernel mode
539 * when we block or switch rather then when we enter the kernel).
540 * This function is NOT called if we are switching into a preemption
541 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
542 * our current process designation (if we have one) and become a true
543 * LWKT thread, and may also hand the current process designation to
544 * another process and schedule thread.
cb973d15
MD
545 */
546 if (td->td_release)
547 td->td_release(td);
548
37af14fe 549 crit_enter_gd(gd);
3b998fa9 550 if (TD_TOKS_HELD(td))
9d265729
MD
551 lwkt_relalltokens(td);
552
553 /*
b02926de
MD
554 * We had better not be holding any spin locks, but don't get into an
555 * endless panic loop.
9d265729 556 */
bbb31c5d
MD
557 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
558 ("lwkt_switch: still holding a shared spinlock %p!",
559 gd->gd_spinlock_rd));
d666840a
MD
560 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
561 ("lwkt_switch: still holding %d exclusive spinlocks!",
562 gd->gd_spinlocks_wr));
9d265729 563
8a8d5d85
MD
564
565#ifdef SMP
566 /*
567 * td_mpcount cannot be used to determine if we currently hold the
568 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
569 * to get the lock, and switch out if it can't. Our ownership of
570 * the actual lock will remain stable while we are in a critical section
571 * (but, of course, another cpu may own or release the lock so the
572 * actual value of mp_lock is not stable).
8a8d5d85
MD
573 */
574 mpheld = MP_LOCK_HELD();
0f7a3396
MD
575#ifdef INVARIANTS
576 if (td->td_cscount) {
6ea70f76 577 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
578 td);
579 if (panic_on_cscount)
580 panic("switching while mastering cpusync");
581 }
582#endif
8a8d5d85 583#endif
f9235b6d
MD
584
585 /*
586 * If we had preempted another thread on this cpu, resume the preempted
587 * thread. This occurs transparently, whether the preempted thread
588 * was scheduled or not (it may have been preempted after descheduling
589 * itself).
590 *
591 * We have to setup the MP lock for the original thread after backing
592 * out the adjustment that was made to curthread when the original
593 * was preempted.
594 */
99df837e 595 if ((ntd = td->td_preempted) != NULL) {
26a0694b 596 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 597#ifdef SMP
96728c05 598 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 599 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
600 td, ntd, td->td_mpcount, ntd->td_mpcount);
601 }
8a8d5d85
MD
602 if (ntd->td_mpcount) {
603 td->td_mpcount -= ntd->td_mpcount;
604 KKASSERT(td->td_mpcount >= 0);
605 }
606#endif
26a0694b 607 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
608
609 /*
b9eb1c19
MD
610 * The interrupt may have woken a thread up, we need to properly
611 * set the reschedule flag if the originally interrupted thread is
612 * at a lower priority.
8ec60c3f 613 */
f9235b6d
MD
614 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
615 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 616 need_lwkt_resched();
f9235b6d 617 }
8a8d5d85 618 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
619 goto havethread_preempted;
620 }
621
622 /*
623 * Implement round-robin fairq with priority insertion. The priority
624 * insertion is handled by _lwkt_enqueue()
625 *
626 * We have to adjust the MP lock for the target thread. If we
627 * need the MP lock and cannot obtain it we try to locate a
628 * thread that does not need the MP lock. If we cannot, we spin
629 * instead of HLT.
630 *
631 * A similar issue exists for the tokens held by the target thread.
632 * If we cannot obtain ownership of the tokens we cannot immediately
633 * schedule the thread.
634 */
635 for (;;) {
636 clear_lwkt_resched();
637 didaccumulate = 0;
638 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
639
4b5f931b 640 /*
f9235b6d 641 * Hotpath if we can get all necessary resources.
41a01a4d 642 *
f9235b6d 643 * If nothing is runnable switch to the idle thread
41a01a4d 644 */
f9235b6d
MD
645 if (ntd == NULL) {
646 ntd = &gd->gd_idlethread;
647 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
648 ntd->td_flags |= TDF_IDLE_NOHLT;
649 if (ntd->td_mpcount) {
650 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
651 panic("Idle thread %p was holding the BGL!", ntd);
652 if (mpheld == 0) {
653 cpu_pause();
654 continue;
655 }
656 }
657 goto haveidle;
658 }
41a01a4d 659
8ec60c3f 660 /*
f9235b6d 661 * Hotpath schedule
8ec60c3f 662 */
f9235b6d
MD
663 if (ntd->td_fairq_accum >= 0 &&
664#ifdef SMP
665 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
666#endif
667 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd))
668 ) {
8a8d5d85 669#ifdef SMP
f9235b6d
MD
670 clr_mplock_contention_mask(gd);
671#endif
672 goto havethread;
673 }
674
675#ifdef SMP
676 /* Reload mpheld (it become stale after mplock/token ops) */
677 mpheld = MP_LOCK_HELD();
678#endif
679
680 /*
681 * Coldpath - unable to schedule ntd, continue looking for threads
682 * to schedule. This is only allowed of the (presumably) kernel
683 * thread exhausted its fair share. A kernel thread stuck on
684 * resources does not currently allow a user thread to get in
685 * front of it.
686 */
687#ifdef SMP
688 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
689 (ntd->td_fairq_accum < 0));
690#endif
691 nlast = NULL;
692
693 for (;;) {
41a01a4d 694 /*
f9235b6d
MD
695 * If the fair-share scheduler ran out ntd gets moved to the
696 * end and its accumulator will be bumped, if it didn't we
697 * maintain the same queue position.
df6b8ba0 698 *
f9235b6d 699 * nlast keeps track of the last element prior to any moves.
41a01a4d 700 */
f9235b6d
MD
701 if (ntd->td_fairq_accum < 0) {
702 xtd = TAILQ_NEXT(ntd, td_threadq);
703 lwkt_fairq_accumulate(gd, ntd);
704 didaccumulate = 1;
705 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
706 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
707 if (nlast == NULL) {
708 nlast = ntd;
709 if (xtd == NULL)
710 xtd = ntd;
711 }
712 ntd = xtd;
713 } else {
714 ntd = TAILQ_NEXT(ntd, td_threadq);
715 }
a453459d 716
f9235b6d
MD
717 /*
718 * If we exhausted the run list switch to the idle thread.
719 * Since one or more threads had resource acquisition issues
720 * we do not allow the idle thread to halt.
721 *
722 * NOTE: nlast can be NULL.
723 */
724 if (ntd == nlast) {
e0a90d3b 725 cpu_pause();
f9235b6d
MD
726 ntd = &gd->gd_idlethread;
727 ntd->td_flags |= TDF_IDLE_NOHLT;
728 set_mplock_contention_mask(gd);
729 cpu_mplock_contested();
730 if (ntd->td_mpcount) {
731 mpheld = MP_LOCK_HELD();
732 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
733 panic("Idle thread %p was holding the BGL!", ntd);
734 if (mpheld == 0) {
735 cpu_pause();
736 break; /* try again from the top, almost */
b9eb1c19 737 }
8a8d5d85 738 }
684a93c4
MD
739
740 /*
f9235b6d
MD
741 * If fairq accumulations occured we do not schedule the
742 * idle thread. This will cause us to try again from
743 * the (almost) top.
684a93c4 744 */
f9235b6d
MD
745 if (didaccumulate)
746 break;
747 goto haveidle;
8a8d5d85 748 }
f9235b6d 749
df6b8ba0 750 /*
f9235b6d 751 * Try to switch to this thread.
df6b8ba0 752 */
f9235b6d
MD
753 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok) &&
754 ntd->td_fairq_accum >= 0 &&
755#ifdef SMP
756 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
8a8d5d85 757#endif
f9235b6d
MD
758 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd))
759 ) {
a453459d 760#ifdef SMP
f9235b6d
MD
761 clr_mplock_contention_mask(gd);
762#endif
763 goto havethread;
df6b8ba0 764 }
f9235b6d
MD
765#ifdef SMP
766 /* Reload mpheld (it become stale after mplock/token ops) */
767 mpheld = MP_LOCK_HELD();
768 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
769 nquserok = 0;
a453459d 770#endif
4b5f931b 771 }
f1d1c3fa 772 }
8a8d5d85
MD
773
774 /*
f9235b6d
MD
775 * Do the actual switch. WARNING: mpheld is stale here.
776 *
777 * We must always decrement td_fairq_accum on non-idle threads just
778 * in case a thread never gets a tick due to being in a continuous
779 * critical section. The page-zeroing code does that.
780 *
781 * If the thread we came up with is a higher or equal priority verses
782 * the thread at the head of the queue we move our thread to the
783 * front. This way we can always check the front of the queue.
784 */
785havethread:
786 ++gd->gd_cnt.v_swtch;
787 --ntd->td_fairq_accum;
788 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
789 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
790 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
791 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
792 }
793havethread_preempted:
794 ;
795 /*
796 * If the new target does not need the MP lock and we are holding it,
797 * release the MP lock. If the new target requires the MP lock we have
798 * already acquired it for the target.
799 *
800 * WARNING: mpheld is stale here.
8a8d5d85 801 */
f9235b6d
MD
802haveidle:
803 KASSERT(ntd->td_critcount,
804 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
805#ifdef SMP
806 if (ntd->td_mpcount == 0 ) {
807 if (MP_LOCK_HELD())
808 cpu_rel_mplock();
809 } else {
a453459d 810 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
811 }
812#endif
94f6d86e
MD
813 if (td != ntd) {
814 ++switch_count;
b2b3ffcd 815#ifdef __x86_64__
f9235b6d
MD
816 {
817 int tos_ok __debugvar = jg_tos_ok(ntd);
818 KKASSERT(tos_ok);
819 }
85514115 820#endif
a1f0fb66 821 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 822 td->td_switch(ntd);
94f6d86e 823 }
37af14fe
MD
824 /* NOTE: current cpu may have changed after switch */
825 crit_exit_quick(td);
8ad65e08
MD
826}
827
b68b7282 828/*
96728c05
MD
829 * Request that the target thread preempt the current thread. Preemption
830 * only works under a specific set of conditions:
b68b7282 831 *
96728c05
MD
832 * - We are not preempting ourselves
833 * - The target thread is owned by the current cpu
834 * - We are not currently being preempted
835 * - The target is not currently being preempted
d3d1cbc8
MD
836 * - We are not holding any spin locks
837 * - The target thread is not holding any tokens
96728c05
MD
838 * - We are able to satisfy the target's MP lock requirements (if any).
839 *
840 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
841 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 842 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
843 * to determine whether preemption is possible (aka usually just the crit
844 * priority of lwkt_schedule() itself).
b68b7282 845 *
26a0694b
MD
846 * XXX at the moment we run the target thread in a critical section during
847 * the preemption in order to prevent the target from taking interrupts
848 * that *WE* can't. Preemption is strictly limited to interrupt threads
849 * and interrupt-like threads, outside of a critical section, and the
850 * preempted source thread will be resumed the instant the target blocks
851 * whether or not the source is scheduled (i.e. preemption is supposed to
852 * be as transparent as possible).
4b5f931b 853 *
8a8d5d85
MD
854 * The target thread inherits our MP count (added to its own) for the
855 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
856 * MP lock during the preemption. Therefore, any preempting targets must be
857 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
858 * out of sync with the physical mp_lock, but we do not have to preserve
859 * the original ownership of the lock if it was out of synch (that is, we
860 * can leave it synchronized on return).
b68b7282
MD
861 */
862void
f9235b6d 863lwkt_preempt(thread_t ntd, int critcount)
b68b7282 864{
46a3f46d 865 struct globaldata *gd = mycpu;
0a3f9b47 866 thread_t td;
8a8d5d85
MD
867#ifdef SMP
868 int mpheld;
57c254db 869 int savecnt;
8a8d5d85 870#endif
b68b7282 871
26a0694b 872 /*
96728c05
MD
873 * The caller has put us in a critical section. We can only preempt
874 * if the caller of the caller was not in a critical section (basically
f9235b6d 875 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 876 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 877 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
878 *
879 * YYY The target thread must be in a critical section (else it must
880 * inherit our critical section? I dunno yet).
41a01a4d 881 *
0a3f9b47 882 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 883 */
f9235b6d 884 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 885
0a3f9b47 886 td = gd->gd_curthread;
f9235b6d 887 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
888 ++preempt_miss;
889 return;
890 }
f9235b6d 891 if (td->td_critcount > critcount) {
96728c05 892 ++preempt_miss;
8ec60c3f 893 need_lwkt_resched();
96728c05
MD
894 return;
895 }
896#ifdef SMP
46a3f46d 897 if (ntd->td_gd != gd) {
96728c05 898 ++preempt_miss;
8ec60c3f 899 need_lwkt_resched();
96728c05
MD
900 return;
901 }
902#endif
41a01a4d 903 /*
d3d1cbc8 904 * Take the easy way out and do not preempt if we are holding
d666840a 905 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
906 * preempted interlock against the target thread's tokens and whether
907 * we can get all the target thread's tokens, but this situation
908 * should not occur very often so its easier to simply not preempt.
d666840a
MD
909 * Also, plain spinlocks are impossible to figure out at this point so
910 * just don't preempt.
d3d1cbc8
MD
911 *
912 * Do not try to preempt if the target thread is holding any tokens.
913 * We could try to acquire the tokens but this case is so rare there
914 * is no need to support it.
41a01a4d 915 */
bbb31c5d 916 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 917 ++preempt_miss;
8ec60c3f 918 need_lwkt_resched();
41a01a4d
MD
919 return;
920 }
3b998fa9 921 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
922 ++preempt_miss;
923 need_lwkt_resched();
924 return;
925 }
26a0694b
MD
926 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
927 ++preempt_weird;
8ec60c3f 928 need_lwkt_resched();
26a0694b
MD
929 return;
930 }
931 if (ntd->td_preempted) {
4b5f931b 932 ++preempt_hit;
8ec60c3f 933 need_lwkt_resched();
26a0694b 934 return;
b68b7282 935 }
8a8d5d85 936#ifdef SMP
a2a5ad0d
MD
937 /*
938 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
939 * to or from the MP lock. In this case td_mpcount will be pre-disposed
940 * (non-zero) but not actually synchronized with the actual state of the
941 * lock. We can use it to imply an MP lock requirement for the
942 * preemption but we cannot use it to test whether we hold the MP lock
943 * or not.
a2a5ad0d 944 */
96728c05 945 savecnt = td->td_mpcount;
71ef2f5c 946 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
947 ntd->td_mpcount += td->td_mpcount;
948 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
949 ntd->td_mpcount -= td->td_mpcount;
950 ++preempt_miss;
8ec60c3f 951 need_lwkt_resched();
8a8d5d85
MD
952 return;
953 }
954#endif
26a0694b 955
8ec60c3f
MD
956 /*
957 * Since we are able to preempt the current thread, there is no need to
958 * call need_lwkt_resched().
959 */
26a0694b
MD
960 ++preempt_hit;
961 ntd->td_preempted = td;
962 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 963 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 964 td->td_switch(ntd);
b9eb1c19 965
26a0694b 966 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
967#ifdef SMP
968 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
969 mpheld = MP_LOCK_HELD();
970 if (mpheld && td->td_mpcount == 0)
96728c05 971 cpu_rel_mplock();
71ef2f5c 972 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
973 panic("lwkt_preempt(): MP lock was not held through");
974#endif
26a0694b
MD
975 ntd->td_preempted = NULL;
976 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
977}
978
f1d1c3fa 979/*
faaeffac 980 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 981 *
faaeffac
MD
982 * td_nest_count prevents deep nesting via splz() or doreti() which
983 * might otherwise blow out the kernel stack. Note that except for
984 * this special case, we MUST call splz() here to handle any
985 * pending ints, particularly after we switch, or we might accidently
986 * halt the cpu with interrupts pending.
ef0fdad1 987 *
f1d1c3fa
MD
988 * (self contained on a per cpu basis)
989 */
990void
faaeffac 991splz_check(void)
f1d1c3fa 992{
7966cb69
MD
993 globaldata_t gd = mycpu;
994 thread_t td = gd->gd_curthread;
ef0fdad1 995
f9235b6d 996 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
f1d1c3fa 997 splz();
f1d1c3fa
MD
998}
999
8ad65e08 1000/*
f9235b6d
MD
1001 * This function is used to negotiate a passive release of the current
1002 * process/lwp designation with the user scheduler, allowing the user
1003 * scheduler to schedule another user thread. The related kernel thread
1004 * (curthread) continues running in the released state.
8ad65e08
MD
1005 */
1006void
f9235b6d 1007lwkt_passive_release(struct thread *td)
8ad65e08 1008{
f9235b6d
MD
1009 struct lwp *lp = td->td_lwp;
1010
1011 td->td_release = NULL;
1012 lwkt_setpri_self(TDPRI_KERN_USER);
1013 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1014}
1015
f9235b6d 1016
3824f392 1017/*
f9235b6d
MD
1018 * This implements a normal yield. This routine is virtually a nop if
1019 * there is nothing to yield to but it will always run any pending interrupts
1020 * if called from a critical section.
1021 *
1022 * This yield is designed for kernel threads without a user context.
1023 *
1024 * (self contained on a per cpu basis)
3824f392
MD
1025 */
1026void
f9235b6d 1027lwkt_yield(void)
3824f392 1028{
f9235b6d
MD
1029 globaldata_t gd = mycpu;
1030 thread_t td = gd->gd_curthread;
1031 thread_t xtd;
3824f392 1032
f9235b6d
MD
1033 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1034 splz();
1035 if (td->td_fairq_accum < 0) {
1036 lwkt_schedule_self(curthread);
1037 lwkt_switch();
1038 } else {
1039 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1040 if (xtd && xtd->td_pri > td->td_pri) {
1041 lwkt_schedule_self(curthread);
1042 lwkt_switch();
1043 }
1044 }
3824f392
MD
1045}
1046
1047/*
f9235b6d
MD
1048 * This yield is designed for kernel threads with a user context.
1049 *
1050 * The kernel acting on behalf of the user is potentially cpu-bound,
1051 * this function will efficiently allow other threads to run and also
1052 * switch to other processes by releasing.
3824f392
MD
1053 *
1054 * The lwkt_user_yield() function is designed to have very low overhead
1055 * if no yield is determined to be needed.
1056 */
1057void
1058lwkt_user_yield(void)
1059{
f9235b6d
MD
1060 globaldata_t gd = mycpu;
1061 thread_t td = gd->gd_curthread;
1062
1063 /*
1064 * Always run any pending interrupts in case we are in a critical
1065 * section.
1066 */
1067 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1068 splz();
3824f392
MD
1069
1070#ifdef SMP
1071 /*
1072 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1073 * kernel can prevent other cpus from servicing interrupt threads
1074 * which still require the MP lock (which is a lot of them). This
1075 * has a chaining effect since if the interrupt is blocked, so is
1076 * the event, so normal scheduling will not pick up on the problem.
1077 */
684a93c4
MD
1078 if (mp_lock_contention_mask && td->td_mpcount) {
1079 yield_mplock(td);
3824f392
MD
1080 }
1081#endif
1082
1083 /*
f9235b6d
MD
1084 * Switch (which forces a release) if another kernel thread needs
1085 * the cpu, if userland wants us to resched, or if our kernel
1086 * quantum has run out.
3824f392 1087 */
f9235b6d
MD
1088 if (lwkt_resched_wanted() ||
1089 user_resched_wanted() ||
1090 td->td_fairq_accum < 0)
1091 {
3824f392 1092 lwkt_switch();
3824f392
MD
1093 }
1094
f9235b6d 1095#if 0
3824f392 1096 /*
f9235b6d
MD
1097 * Reacquire the current process if we are released.
1098 *
1099 * XXX not implemented atm. The kernel may be holding locks and such,
1100 * so we want the thread to continue to receive cpu.
3824f392 1101 */
f9235b6d
MD
1102 if (td->td_release == NULL && lp) {
1103 lp->lwp_proc->p_usched->acquire_curproc(lp);
1104 td->td_release = lwkt_passive_release;
1105 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1106 }
f9235b6d 1107#endif
b9eb1c19
MD
1108}
1109
8ad65e08 1110/*
f1d1c3fa
MD
1111 * Generic schedule. Possibly schedule threads belonging to other cpus and
1112 * deal with threads that might be blocked on a wait queue.
1113 *
0a3f9b47
MD
1114 * We have a little helper inline function which does additional work after
1115 * the thread has been enqueued, including dealing with preemption and
1116 * setting need_lwkt_resched() (which prevents the kernel from returning
1117 * to userland until it has processed higher priority threads).
6330a558
MD
1118 *
1119 * It is possible for this routine to be called after a failed _enqueue
1120 * (due to the target thread migrating, sleeping, or otherwise blocked).
1121 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1122 *
1123 * reschedok is an optimized constant propagated from lwkt_schedule() or
1124 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1125 * reschedule to be requested if the target thread has a higher priority.
1126 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1127 * be 0, prevented undesired reschedules.
8ad65e08 1128 */
0a3f9b47
MD
1129static __inline
1130void
f9235b6d 1131_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1132{
b9eb1c19 1133 thread_t otd;
c730be20 1134
6330a558 1135 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1136 if (ntd->td_preemptable && reschedok) {
f9235b6d 1137 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1138 } else if (reschedok) {
b9eb1c19 1139 otd = curthread;
f9235b6d 1140 if (ntd->td_pri > otd->td_pri)
c730be20 1141 need_lwkt_resched();
6330a558 1142 }
f9235b6d
MD
1143
1144 /*
1145 * Give the thread a little fair share scheduler bump if it
1146 * has been asleep for a while. This is primarily to avoid
1147 * a degenerate case for interrupt threads where accumulator
1148 * crosses into negative territory unnecessarily.
1149 */
1150 if (ntd->td_fairq_lticks != ticks) {
1151 ntd->td_fairq_lticks = ticks;
1152 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1153 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1154 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1155 }
0a3f9b47
MD
1156 }
1157}
1158
361d01dd 1159static __inline
8ad65e08 1160void
361d01dd 1161_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1162{
37af14fe
MD
1163 globaldata_t mygd = mycpu;
1164
41a01a4d 1165 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1166 crit_enter_gd(mygd);
9388413d 1167 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1168 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1169 _lwkt_enqueue(td);
1170 } else {
f1d1c3fa 1171 /*
7cd8d145
MD
1172 * If we own the thread, there is no race (since we are in a
1173 * critical section). If we do not own the thread there might
1174 * be a race but the target cpu will deal with it.
f1d1c3fa 1175 */
0f7a3396 1176#ifdef SMP
7cd8d145 1177 if (td->td_gd == mygd) {
9d265729 1178 _lwkt_enqueue(td);
f9235b6d 1179 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1180 } else {
e381e77c 1181 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1182 }
0f7a3396 1183#else
7cd8d145 1184 _lwkt_enqueue(td);
f9235b6d 1185 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1186#endif
8ad65e08 1187 }
37af14fe 1188 crit_exit_gd(mygd);
8ad65e08
MD
1189}
1190
361d01dd
MD
1191void
1192lwkt_schedule(thread_t td)
1193{
1194 _lwkt_schedule(td, 1);
1195}
1196
1197void
1198lwkt_schedule_noresched(thread_t td)
1199{
1200 _lwkt_schedule(td, 0);
1201}
1202
88ebb169
SW
1203#ifdef SMP
1204
e381e77c
MD
1205/*
1206 * When scheduled remotely if frame != NULL the IPIQ is being
1207 * run via doreti or an interrupt then preemption can be allowed.
1208 *
1209 * To allow preemption we have to drop the critical section so only
1210 * one is present in _lwkt_schedule_post.
1211 */
1212static void
1213lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1214{
1215 thread_t td = curthread;
1216 thread_t ntd = arg;
1217
1218 if (frame && ntd->td_preemptable) {
1219 crit_exit_noyield(td);
1220 _lwkt_schedule(ntd, 1);
1221 crit_enter_quick(td);
1222 } else {
1223 _lwkt_schedule(ntd, 1);
1224 }
1225}
1226
d9eea1a5 1227/*
52eedfb5
MD
1228 * Thread migration using a 'Pull' method. The thread may or may not be
1229 * the current thread. It MUST be descheduled and in a stable state.
1230 * lwkt_giveaway() must be called on the cpu owning the thread.
1231 *
1232 * At any point after lwkt_giveaway() is called, the target cpu may
1233 * 'pull' the thread by calling lwkt_acquire().
1234 *
ae8e83e6
MD
1235 * We have to make sure the thread is not sitting on a per-cpu tsleep
1236 * queue or it will blow up when it moves to another cpu.
1237 *
52eedfb5 1238 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1239 */
52eedfb5
MD
1240void
1241lwkt_giveaway(thread_t td)
1242{
3b4192fb 1243 globaldata_t gd = mycpu;
52eedfb5 1244
3b4192fb
MD
1245 crit_enter_gd(gd);
1246 if (td->td_flags & TDF_TSLEEPQ)
1247 tsleep_remove(td);
1248 KKASSERT(td->td_gd == gd);
1249 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1250 td->td_flags |= TDF_MIGRATING;
1251 crit_exit_gd(gd);
52eedfb5
MD
1252}
1253
a2a5ad0d
MD
1254void
1255lwkt_acquire(thread_t td)
1256{
37af14fe
MD
1257 globaldata_t gd;
1258 globaldata_t mygd;
a2a5ad0d 1259
52eedfb5 1260 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1261 gd = td->td_gd;
37af14fe 1262 mygd = mycpu;
52eedfb5 1263 if (gd != mycpu) {
35238fa5 1264 cpu_lfence();
52eedfb5 1265 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1266 crit_enter_gd(mygd);
df910c23
MD
1267 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1268#ifdef SMP
1269 lwkt_process_ipiq();
1270#endif
52eedfb5 1271 cpu_lfence();
df910c23 1272 }
37af14fe 1273 td->td_gd = mygd;
52eedfb5
MD
1274 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1275 td->td_flags &= ~TDF_MIGRATING;
1276 crit_exit_gd(mygd);
1277 } else {
1278 crit_enter_gd(mygd);
1279 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1280 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1281 crit_exit_gd(mygd);
a2a5ad0d
MD
1282 }
1283}
1284
52eedfb5
MD
1285#endif
1286
f1d1c3fa
MD
1287/*
1288 * Generic deschedule. Descheduling threads other then your own should be
1289 * done only in carefully controlled circumstances. Descheduling is
1290 * asynchronous.
1291 *
1292 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1293 */
1294void
1295lwkt_deschedule(thread_t td)
1296{
f1d1c3fa 1297 crit_enter();
b8a98473 1298#ifdef SMP
f1d1c3fa
MD
1299 if (td == curthread) {
1300 _lwkt_dequeue(td);
1301 } else {
a72187e9 1302 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1303 _lwkt_dequeue(td);
1304 } else {
b8a98473 1305 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1306 }
1307 }
b8a98473
MD
1308#else
1309 _lwkt_dequeue(td);
1310#endif
f1d1c3fa
MD
1311 crit_exit();
1312}
1313
4b5f931b
MD
1314/*
1315 * Set the target thread's priority. This routine does not automatically
1316 * switch to a higher priority thread, LWKT threads are not designed for
1317 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1318 */
1319void
1320lwkt_setpri(thread_t td, int pri)
1321{
a72187e9 1322 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1323 if (td->td_pri != pri) {
1324 KKASSERT(pri >= 0);
1325 crit_enter();
1326 if (td->td_flags & TDF_RUNQ) {
1327 _lwkt_dequeue(td);
1328 td->td_pri = pri;
1329 _lwkt_enqueue(td);
1330 } else {
1331 td->td_pri = pri;
1332 }
1333 crit_exit();
26a0694b 1334 }
26a0694b
MD
1335}
1336
03bd0a5e
MD
1337/*
1338 * Set the initial priority for a thread prior to it being scheduled for
1339 * the first time. The thread MUST NOT be scheduled before or during
1340 * this call. The thread may be assigned to a cpu other then the current
1341 * cpu.
1342 *
1343 * Typically used after a thread has been created with TDF_STOPPREQ,
1344 * and before the thread is initially scheduled.
1345 */
1346void
1347lwkt_setpri_initial(thread_t td, int pri)
1348{
1349 KKASSERT(pri >= 0);
1350 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1351 td->td_pri = pri;
03bd0a5e
MD
1352}
1353
26a0694b
MD
1354void
1355lwkt_setpri_self(int pri)
1356{
1357 thread_t td = curthread;
1358
4b5f931b
MD
1359 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1360 crit_enter();
1361 if (td->td_flags & TDF_RUNQ) {
1362 _lwkt_dequeue(td);
f9235b6d 1363 td->td_pri = pri;
4b5f931b
MD
1364 _lwkt_enqueue(td);
1365 } else {
f9235b6d 1366 td->td_pri = pri;
4b5f931b
MD
1367 }
1368 crit_exit();
1369}
1370
f9235b6d
MD
1371/*
1372 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1373 *
1374 * Example: two competing threads, same priority N. decrement by (2*N)
1375 * increment by N*8, each thread will get 4 ticks.
1376 */
1377void
1378lwkt_fairq_schedulerclock(thread_t td)
1379{
1380 if (fairq_enable) {
1381 while (td) {
1382 if (td != &td->td_gd->gd_idlethread) {
1383 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1384 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1385 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1386 if (td->td_fairq_accum < 0)
1387 need_lwkt_resched();
1388 td->td_fairq_lticks = ticks;
1389 }
1390 td = td->td_preempted;
1391 }
1392 }
1393}
1394
1395static void
1396lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1397{
1398 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1399 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1400 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1401}
1402
5d21b981 1403/*
52eedfb5
MD
1404 * Migrate the current thread to the specified cpu.
1405 *
1406 * This is accomplished by descheduling ourselves from the current cpu,
1407 * moving our thread to the tdallq of the target cpu, IPI messaging the
1408 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1409 * races while the thread is being migrated.
ae8e83e6
MD
1410 *
1411 * We must be sure to remove ourselves from the current cpu's tsleepq
1412 * before potentially moving to another queue. The thread can be on
1413 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1414 */
3d28ff59 1415#ifdef SMP
5d21b981 1416static void lwkt_setcpu_remote(void *arg);
3d28ff59 1417#endif
5d21b981
MD
1418
1419void
1420lwkt_setcpu_self(globaldata_t rgd)
1421{
1422#ifdef SMP
1423 thread_t td = curthread;
1424
1425 if (td->td_gd != rgd) {
1426 crit_enter_quick(td);
ae8e83e6 1427 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1428 tsleep_remove(td);
5d21b981
MD
1429 td->td_flags |= TDF_MIGRATING;
1430 lwkt_deschedule_self(td);
52eedfb5 1431 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1432 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1433 lwkt_switch();
1434 /* we are now on the target cpu */
52eedfb5 1435 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1436 crit_exit_quick(td);
1437 }
1438#endif
1439}
1440
ecdefdda
MD
1441void
1442lwkt_migratecpu(int cpuid)
1443{
1444#ifdef SMP
1445 globaldata_t rgd;
1446
1447 rgd = globaldata_find(cpuid);
1448 lwkt_setcpu_self(rgd);
1449#endif
1450}
1451
5d21b981
MD
1452/*
1453 * Remote IPI for cpu migration (called while in a critical section so we
1454 * do not have to enter another one). The thread has already been moved to
1455 * our cpu's allq, but we must wait for the thread to be completely switched
1456 * out on the originating cpu before we schedule it on ours or the stack
1457 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1458 * change to main memory.
1459 *
1460 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1461 * against wakeups. It is best if this interface is used only when there
1462 * are no pending events that might try to schedule the thread.
1463 */
3d28ff59 1464#ifdef SMP
5d21b981
MD
1465static void
1466lwkt_setcpu_remote(void *arg)
1467{
1468 thread_t td = arg;
1469 globaldata_t gd = mycpu;
1470
df910c23
MD
1471 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1472#ifdef SMP
1473 lwkt_process_ipiq();
1474#endif
35238fa5 1475 cpu_lfence();
df910c23 1476 }
5d21b981 1477 td->td_gd = gd;
35238fa5 1478 cpu_sfence();
5d21b981 1479 td->td_flags &= ~TDF_MIGRATING;
9388413d 1480 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1481 _lwkt_enqueue(td);
1482}
3d28ff59 1483#endif
5d21b981 1484
553ea3c8 1485struct lwp *
4b5f931b
MD
1486lwkt_preempted_proc(void)
1487{
73e4f7b9 1488 thread_t td = curthread;
4b5f931b
MD
1489 while (td->td_preempted)
1490 td = td->td_preempted;
553ea3c8 1491 return(td->td_lwp);
4b5f931b
MD
1492}
1493
99df837e
MD
1494/*
1495 * Create a kernel process/thread/whatever. It shares it's address space
1496 * with proc0 - ie: kernel only.
1497 *
365fa13f
MD
1498 * NOTE! By default new threads are created with the MP lock held. A
1499 * thread which does not require the MP lock should release it by calling
1500 * rel_mplock() at the start of the new thread.
99df837e
MD
1501 */
1502int
1503lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1504 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1505 const char *fmt, ...)
99df837e 1506{
73e4f7b9 1507 thread_t td;
e2565a42 1508 __va_list ap;
99df837e 1509
d3d32139 1510 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1511 tdflags);
a2a5ad0d
MD
1512 if (tdp)
1513 *tdp = td;
709799ea 1514 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1515
1516 /*
1517 * Set up arg0 for 'ps' etc
1518 */
e2565a42 1519 __va_start(ap, fmt);
379210cb 1520 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1521 __va_end(ap);
99df837e
MD
1522
1523 /*
1524 * Schedule the thread to run
1525 */
ef0fdad1
MD
1526 if ((td->td_flags & TDF_STOPREQ) == 0)
1527 lwkt_schedule(td);
1528 else
1529 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1530 return 0;
1531}
1532
1533/*
1534 * Destroy an LWKT thread. Warning! This function is not called when
1535 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1536 * uses a different reaping mechanism.
1537 */
1538void
1539lwkt_exit(void)
1540{
1541 thread_t td = curthread;
c070746a 1542 thread_t std;
8826f33a 1543 globaldata_t gd;
99df837e
MD
1544
1545 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1546 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1547 caps_exit(td);
c070746a
MD
1548
1549 /*
1550 * Get us into a critical section to interlock gd_freetd and loop
1551 * until we can get it freed.
1552 *
1553 * We have to cache the current td in gd_freetd because objcache_put()ing
1554 * it would rip it out from under us while our thread is still active.
1555 */
1556 gd = mycpu;
37af14fe 1557 crit_enter_quick(td);
c070746a
MD
1558 while ((std = gd->gd_freetd) != NULL) {
1559 gd->gd_freetd = NULL;
1560 objcache_put(thread_cache, std);
1561 }
3b4192fb
MD
1562
1563 /*
1564 * Remove thread resources from kernel lists and deschedule us for
1565 * the last time.
1566 */
1567 if (td->td_flags & TDF_TSLEEPQ)
1568 tsleep_remove(td);
79eae878 1569 biosched_done(td);
f8abf63c 1570 dsched_exit_thread(td);
37af14fe 1571 lwkt_deschedule_self(td);
e56e4dea 1572 lwkt_remove_tdallq(td);
c070746a
MD
1573 if (td->td_flags & TDF_ALLOCATED_THREAD)
1574 gd->gd_freetd = td;
99df837e
MD
1575 cpu_thread_exit();
1576}
1577
e56e4dea
MD
1578void
1579lwkt_remove_tdallq(thread_t td)
1580{
1581 KKASSERT(td->td_gd == mycpu);
1582 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1583}
1584
2d93b37a
MD
1585void
1586crit_panic(void)
1587{
1588 thread_t td = curthread;
1589 int lpri = td->td_pri;
1590
1591 td->td_pri = 0;
1592 panic("td_pri is/would-go negative! %p %d", td, lpri);
1593}
1594
d165e668
MD
1595#ifdef SMP
1596
bd8015ca
MD
1597/*
1598 * Called from debugger/panic on cpus which have been stopped. We must still
1599 * process the IPIQ while stopped, even if we were stopped while in a critical
1600 * section (XXX).
1601 *
1602 * If we are dumping also try to process any pending interrupts. This may
1603 * or may not work depending on the state of the cpu at the point it was
1604 * stopped.
1605 */
1606void
1607lwkt_smp_stopped(void)
1608{
1609 globaldata_t gd = mycpu;
1610
1611 crit_enter_gd(gd);
1612 if (dumping) {
1613 lwkt_process_ipiq();
1614 splz();
1615 } else {
1616 lwkt_process_ipiq();
1617 }
1618 crit_exit_gd(gd);
1619}
1620
d165e668 1621#endif