kernel - clarify code documentation in vfs_cluster
[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
e381e77c 102
0855a2af
JG
103extern void cpu_heavy_restore(void);
104extern void cpu_lwkt_restore(void);
105extern void cpu_kthread_restore(void);
106extern void cpu_idle_restore(void);
107
b2b3ffcd 108#ifdef __x86_64__
85514115
MD
109
110static int
0855a2af
JG
111jg_tos_ok(struct thread *td)
112{
85514115
MD
113 void *tos;
114 int tos_ok;
115
0855a2af
JG
116 if (td == NULL) {
117 return 1;
118 }
119 KKASSERT(td->td_sp != NULL);
85514115
MD
120 tos = ((void **)td->td_sp)[0];
121 tos_ok = 0;
122 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
123 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
124 tos_ok = 1;
125 }
126 return tos_ok;
127}
128
85514115
MD
129#endif
130
fb0f29c4
MD
131/*
132 * We can make all thread ports use the spin backend instead of the thread
133 * backend. This should only be set to debug the spin backend.
134 */
135TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
136
0f7a3396
MD
137#ifdef INVARIANTS
138SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
139#endif
4b5f931b 140SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
141SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
142 "Successful preemption events");
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
144 "Failed preemption events");
26a0694b 145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
146#ifdef INVARIANTS
147SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
148 &token_contention_count, 0, "spinning due to token contention");
38717797 149#endif
05220613 150
4b5f931b
MD
151/*
152 * These helper procedures handle the runq, they can only be called from
153 * within a critical section.
75cdbe6c
MD
154 *
155 * WARNING! Prior to SMP being brought up it is possible to enqueue and
156 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
157 * instead of 'mycpu' when referencing the globaldata structure. Once
158 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 159 */
f1d1c3fa
MD
160static __inline
161void
162_lwkt_dequeue(thread_t td)
163{
164 if (td->td_flags & TDF_RUNQ) {
4b5f931b 165 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 166 struct globaldata *gd = td->td_gd;
4b5f931b 167
f1d1c3fa 168 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
169 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
170 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
171 }
172}
173
174static __inline
175void
176_lwkt_enqueue(thread_t td)
177{
7f5d7ed7 178 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 179 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 180 struct globaldata *gd = td->td_gd;
4b5f931b 181
f1d1c3fa 182 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
183 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
184 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
185 }
186}
8ad65e08 187
40aaf5fc
NT
188static __boolean_t
189_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
190{
191 struct thread *td = (struct thread *)obj;
192
193 td->td_kstack = NULL;
194 td->td_kstack_size = 0;
195 td->td_flags = TDF_ALLOCATED_THREAD;
196 return (1);
197}
198
199static void
200_lwkt_thread_dtor(void *obj, void *privdata)
201{
202 struct thread *td = (struct thread *)obj;
203
204 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
205 ("_lwkt_thread_dtor: not allocated from objcache"));
206 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
207 td->td_kstack_size > 0,
208 ("_lwkt_thread_dtor: corrupted stack"));
209 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
210}
211
212/*
213 * Initialize the lwkt s/system.
214 */
215void
216lwkt_init(void)
217{
218 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
219 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
220 NULL, CACHE_NTHREADS/2,
221 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
222}
223
37af14fe
MD
224/*
225 * Schedule a thread to run. As the current thread we can always safely
226 * schedule ourselves, and a shortcut procedure is provided for that
227 * function.
228 *
229 * (non-blocking, self contained on a per cpu basis)
230 */
231void
232lwkt_schedule_self(thread_t td)
233{
234 crit_enter_quick(td);
37af14fe 235 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 236 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 237 _lwkt_enqueue(td);
37af14fe
MD
238 crit_exit_quick(td);
239}
240
241/*
242 * Deschedule a thread.
243 *
244 * (non-blocking, self contained on a per cpu basis)
245 */
246void
247lwkt_deschedule_self(thread_t td)
248{
249 crit_enter_quick(td);
37af14fe
MD
250 _lwkt_dequeue(td);
251 crit_exit_quick(td);
252}
253
8ad65e08
MD
254/*
255 * LWKTs operate on a per-cpu basis
256 *
73e4f7b9 257 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
258 */
259void
260lwkt_gdinit(struct globaldata *gd)
261{
4b5f931b
MD
262 int i;
263
264 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
265 TAILQ_INIT(&gd->gd_tdrunq[i]);
266 gd->gd_runqmask = 0;
73e4f7b9 267 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
268}
269
7d0bac62
MD
270/*
271 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
272 * or LWKT start address before it can be scheduled. If the target cpu is
273 * -1 the thread will be created on the current cpu.
0cfcada1
MD
274 *
275 * If you intend to create a thread without a process context this function
276 * does everything except load the startup and switcher function.
7d0bac62
MD
277 */
278thread_t
d3d32139 279lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 280{
c070746a 281 globaldata_t gd = mycpu;
99df837e 282 void *stack;
7d0bac62 283
c070746a
MD
284 /*
285 * If static thread storage is not supplied allocate a thread. Reuse
286 * a cached free thread if possible. gd_freetd is used to keep an exiting
287 * thread intact through the exit.
288 */
ef0fdad1 289 if (td == NULL) {
c070746a
MD
290 if ((td = gd->gd_freetd) != NULL)
291 gd->gd_freetd = NULL;
292 else
293 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
294 KASSERT((td->td_flags &
295 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
296 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
297 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 298 }
c070746a
MD
299
300 /*
301 * Try to reuse cached stack.
302 */
f470d0c8
MD
303 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
304 if (flags & TDF_ALLOCATED_STACK) {
e4846942 305 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
306 stack = NULL;
307 }
308 }
309 if (stack == NULL) {
e4846942 310 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 311 flags |= TDF_ALLOCATED_STACK;
99df837e 312 }
75cdbe6c 313 if (cpu < 0)
c070746a 314 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 315 else
f470d0c8 316 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 317 return(td);
7d0bac62
MD
318}
319
320/*
321 * Initialize a preexisting thread structure. This function is used by
322 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
323 *
f8c3996b
MD
324 * All threads start out in a critical section at a priority of
325 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
326 * appropriate. This function may send an IPI message when the
327 * requested cpu is not the current cpu and consequently gd_tdallq may
328 * not be initialized synchronously from the point of view of the originating
329 * cpu.
330 *
331 * NOTE! we have to be careful in regards to creating threads for other cpus
332 * if SMP has not yet been activated.
7d0bac62 333 */
41a01a4d
MD
334#ifdef SMP
335
75cdbe6c
MD
336static void
337lwkt_init_thread_remote(void *arg)
338{
339 thread_t td = arg;
340
52eedfb5
MD
341 /*
342 * Protected by critical section held by IPI dispatch
343 */
75cdbe6c
MD
344 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
345}
346
41a01a4d
MD
347#endif
348
7d0bac62 349void
f470d0c8
MD
350lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
351 struct globaldata *gd)
7d0bac62 352{
37af14fe
MD
353 globaldata_t mygd = mycpu;
354
99df837e
MD
355 bzero(td, sizeof(struct thread));
356 td->td_kstack = stack;
f470d0c8 357 td->td_kstack_size = stksize;
d3d32139 358 td->td_flags = flags;
26a0694b 359 td->td_gd = gd;
f8c3996b 360 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
3b998fa9 361 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
362#ifdef SMP
363 if ((flags & TDF_MPSAFE) == 0)
364 td->td_mpcount = 1;
365#endif
fb0f29c4
MD
366 if (lwkt_use_spin_port)
367 lwkt_initport_spin(&td->td_msgport);
368 else
369 lwkt_initport_thread(&td->td_msgport, td);
99df837e 370 pmap_init_thread(td);
0f7a3396 371#ifdef SMP
5d21b981
MD
372 /*
373 * Normally initializing a thread for a remote cpu requires sending an
374 * IPI. However, the idlethread is setup before the other cpus are
375 * activated so we have to treat it as a special case. XXX manipulation
376 * of gd_tdallq requires the BGL.
377 */
378 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 379 crit_enter_gd(mygd);
75cdbe6c 380 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 381 crit_exit_gd(mygd);
75cdbe6c 382 } else {
2db3b277 383 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 384 }
0f7a3396 385#else
37af14fe 386 crit_enter_gd(mygd);
0f7a3396 387 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 388 crit_exit_gd(mygd);
0f7a3396 389#endif
8c72e3d5
AH
390
391 dsched_new_thread(td);
73e4f7b9
MD
392}
393
394void
395lwkt_set_comm(thread_t td, const char *ctl, ...)
396{
e2565a42 397 __va_list va;
73e4f7b9 398
e2565a42 399 __va_start(va, ctl);
379210cb 400 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 401 __va_end(va);
e7c0dbba 402 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
403}
404
99df837e 405void
73e4f7b9 406lwkt_hold(thread_t td)
99df837e 407{
73e4f7b9
MD
408 ++td->td_refs;
409}
410
411void
412lwkt_rele(thread_t td)
413{
414 KKASSERT(td->td_refs > 0);
415 --td->td_refs;
416}
417
418void
419lwkt_wait_free(thread_t td)
420{
421 while (td->td_refs)
377d4740 422 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
423}
424
425void
426lwkt_free_thread(thread_t td)
427{
d9eea1a5 428 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
429 ("lwkt_free_thread: did not exit! %p", td));
430
40aaf5fc
NT
431 if (td->td_flags & TDF_ALLOCATED_THREAD) {
432 objcache_put(thread_cache, td);
433 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
434 /* client-allocated struct with internally allocated stack */
435 KASSERT(td->td_kstack && td->td_kstack_size > 0,
436 ("lwkt_free_thread: corrupted stack"));
437 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
438 td->td_kstack = NULL;
439 td->td_kstack_size = 0;
99df837e 440 }
e7c0dbba 441 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
442}
443
444
8ad65e08
MD
445/*
446 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
447 * switch to the idlethread. Switching must occur within a critical
448 * section to avoid races with the scheduling queue.
449 *
450 * We always have full control over our cpu's run queue. Other cpus
451 * that wish to manipulate our queue must use the cpu_*msg() calls to
452 * talk to our cpu, so a critical section is all that is needed and
453 * the result is very, very fast thread switching.
454 *
96728c05
MD
455 * The LWKT scheduler uses a fixed priority model and round-robins at
456 * each priority level. User process scheduling is a totally
457 * different beast and LWKT priorities should not be confused with
458 * user process priorities.
f1d1c3fa 459 *
96728c05
MD
460 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
461 * cleans it up. Note that the td_switch() function cannot do anything that
462 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
463 * the target thread (not the current thread). It's nice to have a scheduler
464 * that does not need the MP lock to work because it allows us to do some
465 * really cool high-performance MP lock optimizations.
69d78e99
MD
466 *
467 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
468 * is not called by the current thread in the preemption case, only when
469 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
470 */
471void
472lwkt_switch(void)
473{
37af14fe
MD
474 globaldata_t gd = mycpu;
475 thread_t td = gd->gd_curthread;
8ad65e08 476 thread_t ntd;
8a8d5d85
MD
477#ifdef SMP
478 int mpheld;
479#endif
8ad65e08 480
46a3f46d 481 /*
27e88a6e
MD
482 * Switching from within a 'fast' (non thread switched) interrupt or IPI
483 * is illegal. However, we may have to do it anyway if we hit a fatal
484 * kernel trap or we have paniced.
485 *
486 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 487 */
27e88a6e
MD
488 if (gd->gd_intr_nesting_level) {
489 int savegdnest;
490 int savegdtrap;
491
492 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
493 panic("lwkt_switch: cannot switch from within "
494 "a fast interrupt, yet, td %p\n", td);
495 } else {
496 savegdnest = gd->gd_intr_nesting_level;
497 savegdtrap = gd->gd_trap_nesting_level;
498 gd->gd_intr_nesting_level = 0;
499 gd->gd_trap_nesting_level = 0;
a7422615
MD
500 if ((td->td_flags & TDF_PANICWARN) == 0) {
501 td->td_flags |= TDF_PANICWARN;
6ea70f76 502 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 503 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 504 print_backtrace(-1);
a7422615 505 }
27e88a6e
MD
506 lwkt_switch();
507 gd->gd_intr_nesting_level = savegdnest;
508 gd->gd_trap_nesting_level = savegdtrap;
509 return;
510 }
96728c05 511 }
ef0fdad1 512
cb973d15
MD
513 /*
514 * Passive release (used to transition from user to kernel mode
515 * when we block or switch rather then when we enter the kernel).
516 * This function is NOT called if we are switching into a preemption
517 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
518 * our current process designation (if we have one) and become a true
519 * LWKT thread, and may also hand the current process designation to
520 * another process and schedule thread.
cb973d15
MD
521 */
522 if (td->td_release)
523 td->td_release(td);
524
37af14fe 525 crit_enter_gd(gd);
3b998fa9 526 if (TD_TOKS_HELD(td))
9d265729
MD
527 lwkt_relalltokens(td);
528
529 /*
b02926de
MD
530 * We had better not be holding any spin locks, but don't get into an
531 * endless panic loop.
9d265729 532 */
bbb31c5d
MD
533 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
534 ("lwkt_switch: still holding a shared spinlock %p!",
535 gd->gd_spinlock_rd));
d666840a
MD
536 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
537 ("lwkt_switch: still holding %d exclusive spinlocks!",
538 gd->gd_spinlocks_wr));
9d265729 539
8a8d5d85
MD
540
541#ifdef SMP
542 /*
543 * td_mpcount cannot be used to determine if we currently hold the
544 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
545 * to get the lock, and switch out if it can't. Our ownership of
546 * the actual lock will remain stable while we are in a critical section
547 * (but, of course, another cpu may own or release the lock so the
548 * actual value of mp_lock is not stable).
8a8d5d85
MD
549 */
550 mpheld = MP_LOCK_HELD();
0f7a3396
MD
551#ifdef INVARIANTS
552 if (td->td_cscount) {
6ea70f76 553 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
554 td);
555 if (panic_on_cscount)
556 panic("switching while mastering cpusync");
557 }
558#endif
8a8d5d85 559#endif
99df837e
MD
560 if ((ntd = td->td_preempted) != NULL) {
561 /*
562 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
563 * thread. This occurs transparently, whether the preempted thread
564 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
565 * itself).
566 *
567 * We have to setup the MP lock for the original thread after backing
568 * out the adjustment that was made to curthread when the original
569 * was preempted.
99df837e 570 */
26a0694b 571 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 572#ifdef SMP
96728c05 573 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 574 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
575 td, ntd, td->td_mpcount, ntd->td_mpcount);
576 }
8a8d5d85
MD
577 if (ntd->td_mpcount) {
578 td->td_mpcount -= ntd->td_mpcount;
579 KKASSERT(td->td_mpcount >= 0);
580 }
581#endif
26a0694b 582 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
583
584 /*
b9eb1c19
MD
585 * The interrupt may have woken a thread up, we need to properly
586 * set the reschedule flag if the originally interrupted thread is
587 * at a lower priority.
8ec60c3f
MD
588 */
589 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
590 need_lwkt_resched();
8a8d5d85 591 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 592 } else {
4b5f931b
MD
593 /*
594 * Priority queue / round-robin at each priority. Note that user
595 * processes run at a fixed, low priority and the user process
596 * scheduler deals with interactions between user processes
597 * by scheduling and descheduling them from the LWKT queue as
598 * necessary.
8a8d5d85
MD
599 *
600 * We have to adjust the MP lock for the target thread. If we
601 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
602 * thread that does not need the MP lock. If we cannot, we spin
603 * instead of HLT.
604 *
605 * A similar issue exists for the tokens held by the target thread.
606 * If we cannot obtain ownership of the tokens we cannot immediately
607 * schedule the thread.
608 */
609
8ec60c3f
MD
610 /*
611 * If an LWKT reschedule was requested, well that is what we are
612 * doing now so clear it.
613 */
614 clear_lwkt_resched();
4b5f931b
MD
615again:
616 if (gd->gd_runqmask) {
617 int nq = bsrl(gd->gd_runqmask);
618 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
619 gd->gd_runqmask &= ~(1 << nq);
620 goto again;
621 }
8a8d5d85 622#ifdef SMP
41a01a4d 623 /*
df6b8ba0
MD
624 * THREAD SELECTION FOR AN SMP MACHINE BUILD
625 *
41a01a4d
MD
626 * If the target needs the MP lock and we couldn't get it,
627 * or if the target is holding tokens and we could not
628 * gain ownership of the tokens, continue looking for a
629 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
630 *
631 * NOTE: the mpheld variable invalid after this conditional, it
632 * can change due to both cpu_try_mplock() returning success
9d265729 633 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
634 * we are trying to check the mpcount of a thread other then
635 * the current thread. Because of this, if the current thread
636 * is not holding td_mpcount, an IPI indirectly run via
9d265729 637 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 638 * cause the core MP lock to be released.
41a01a4d
MD
639 */
640 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
3b998fa9 641 (TD_TOKS_HELD(ntd) && lwkt_getalltokens(ntd) == 0)
41a01a4d 642 ) {
8a8d5d85 643 u_int32_t rqmask = gd->gd_runqmask;
a453459d 644
e0a90d3b
MD
645 cpu_pause();
646
a453459d
MD
647 mpheld = MP_LOCK_HELD();
648 ntd = NULL;
8a8d5d85
MD
649 while (rqmask) {
650 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 651 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 652 /* spinning due to MP lock being held */
41a01a4d 653 continue;
38717797 654 }
a453459d
MD
655
656 /*
9d265729 657 * mpheld state invalid after getalltokens call returns
a453459d
MD
658 * failure, but the variable is only needed for
659 * the loop.
660 */
3b998fa9 661 if (TD_TOKS_HELD(ntd) && !lwkt_getalltokens(ntd)) {
a453459d 662 /* spinning due to token contention */
38717797 663#ifdef INVARIANTS
a453459d 664 ++token_contention_count;
38717797 665#endif
a453459d 666 mpheld = MP_LOCK_HELD();
41a01a4d 667 continue;
38717797 668 }
41a01a4d 669 break;
8a8d5d85
MD
670 }
671 if (ntd)
672 break;
673 rqmask &= ~(1 << nq);
674 nq = bsrl(rqmask);
b9eb1c19
MD
675
676 /*
677 * We have two choices. We can either refuse to run a
678 * user thread when a kernel thread needs the MP lock
679 * but could not get it, or we can allow it to run but
680 * then expect an IPI (hopefully) later on to force a
681 * reschedule when the MP lock might become available.
682 */
683 if (nq < TDPRI_KERN_LPSCHED) {
684a93c4
MD
684 break; /* for now refuse to run */
685#if 0
b9eb1c19
MD
686 if (chain_mplock == 0)
687 break;
b9eb1c19 688 /* continue loop, allow user threads to be scheduled */
684a93c4 689#endif
b9eb1c19 690 }
8a8d5d85 691 }
684a93c4
MD
692
693 /*
694 * Case where a (kernel) thread needed the MP lock and could
695 * not get one, and we may or may not have found another
696 * thread which does not need the MP lock to run while
697 * we wait (ntd).
698 */
8a8d5d85 699 if (ntd == NULL) {
a2a5ad0d
MD
700 ntd = &gd->gd_idlethread;
701 ntd->td_flags |= TDF_IDLE_NOHLT;
684a93c4
MD
702 set_mplock_contention_mask(gd);
703 cpu_mplock_contested();
df6b8ba0 704 goto using_idle_thread;
8a8d5d85 705 } else {
684a93c4 706 clr_mplock_contention_mask(gd);
344ad853 707 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
708 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
709 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 }
711 } else {
684a93c4 712 clr_mplock_contention_mask(gd);
344ad853 713 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
714 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
715 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
716 }
717#else
df6b8ba0
MD
718 /*
719 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
720 * worry about tokens or the BGL. However, we still have
721 * to call lwkt_getalltokens() in order to properly detect
722 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 723 */
7eb611ef 724 lwkt_getalltokens(ntd);
344ad853 725 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
726 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
727 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 728#endif
4b5f931b 729 } else {
3c23a41a 730 /*
60f945af
MD
731 * We have nothing to run but only let the idle loop halt
732 * the cpu if there are no pending interrupts.
3c23a41a 733 */
a2a5ad0d 734 ntd = &gd->gd_idlethread;
60f945af 735 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 736 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 737#ifdef SMP
df6b8ba0
MD
738using_idle_thread:
739 /*
740 * The idle thread should not be holding the MP lock unless we
741 * are trapping in the kernel or in a panic. Since we select the
742 * idle thread unconditionally when no other thread is available,
743 * if the MP lock is desired during a panic or kernel trap, we
744 * have to loop in the scheduler until we get it.
745 */
746 if (ntd->td_mpcount) {
747 mpheld = MP_LOCK_HELD();
684a93c4 748 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
df6b8ba0 749 panic("Idle thread %p was holding the BGL!", ntd);
684a93c4 750 if (mpheld == 0)
df6b8ba0
MD
751 goto again;
752 }
a453459d 753#endif
4b5f931b 754 }
f1d1c3fa 755 }
26a0694b
MD
756 KASSERT(ntd->td_pri >= TDPRI_CRIT,
757 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
758
759 /*
760 * Do the actual switch. If the new target does not need the MP lock
761 * and we are holding it, release the MP lock. If the new target requires
762 * the MP lock we have already acquired it for the target.
763 */
764#ifdef SMP
765 if (ntd->td_mpcount == 0 ) {
766 if (MP_LOCK_HELD())
767 cpu_rel_mplock();
768 } else {
a453459d 769 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
770 }
771#endif
94f6d86e
MD
772 if (td != ntd) {
773 ++switch_count;
b2b3ffcd 774#ifdef __x86_64__
f64b567c
SW
775 {
776 int tos_ok __debugvar = jg_tos_ok(ntd);
777 KKASSERT(tos_ok);
778 }
85514115 779#endif
a1f0fb66 780 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 781 td->td_switch(ntd);
94f6d86e 782 }
37af14fe
MD
783 /* NOTE: current cpu may have changed after switch */
784 crit_exit_quick(td);
8ad65e08
MD
785}
786
b68b7282 787/*
96728c05
MD
788 * Request that the target thread preempt the current thread. Preemption
789 * only works under a specific set of conditions:
b68b7282 790 *
96728c05
MD
791 * - We are not preempting ourselves
792 * - The target thread is owned by the current cpu
793 * - We are not currently being preempted
794 * - The target is not currently being preempted
d3d1cbc8
MD
795 * - We are not holding any spin locks
796 * - The target thread is not holding any tokens
96728c05
MD
797 * - We are able to satisfy the target's MP lock requirements (if any).
798 *
799 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
800 * this is called via lwkt_schedule() through the td_preemptable callback.
801 * critpri is the managed critical priority that we should ignore in order
802 * to determine whether preemption is possible (aka usually just the crit
803 * priority of lwkt_schedule() itself).
b68b7282 804 *
26a0694b
MD
805 * XXX at the moment we run the target thread in a critical section during
806 * the preemption in order to prevent the target from taking interrupts
807 * that *WE* can't. Preemption is strictly limited to interrupt threads
808 * and interrupt-like threads, outside of a critical section, and the
809 * preempted source thread will be resumed the instant the target blocks
810 * whether or not the source is scheduled (i.e. preemption is supposed to
811 * be as transparent as possible).
4b5f931b 812 *
8a8d5d85
MD
813 * The target thread inherits our MP count (added to its own) for the
814 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
815 * MP lock during the preemption. Therefore, any preempting targets must be
816 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
817 * out of sync with the physical mp_lock, but we do not have to preserve
818 * the original ownership of the lock if it was out of synch (that is, we
819 * can leave it synchronized on return).
b68b7282
MD
820 */
821void
96728c05 822lwkt_preempt(thread_t ntd, int critpri)
b68b7282 823{
46a3f46d 824 struct globaldata *gd = mycpu;
0a3f9b47 825 thread_t td;
8a8d5d85
MD
826#ifdef SMP
827 int mpheld;
57c254db 828 int savecnt;
8a8d5d85 829#endif
b68b7282 830
26a0694b 831 /*
96728c05
MD
832 * The caller has put us in a critical section. We can only preempt
833 * if the caller of the caller was not in a critical section (basically
d666840a 834 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 835 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 836 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
837 *
838 * YYY The target thread must be in a critical section (else it must
839 * inherit our critical section? I dunno yet).
41a01a4d 840 *
0a3f9b47 841 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
842 */
843 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 844
0a3f9b47 845 td = gd->gd_curthread;
0a3f9b47 846 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
847 ++preempt_miss;
848 return;
849 }
96728c05
MD
850 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
851 ++preempt_miss;
8ec60c3f 852 need_lwkt_resched();
96728c05
MD
853 return;
854 }
855#ifdef SMP
46a3f46d 856 if (ntd->td_gd != gd) {
96728c05 857 ++preempt_miss;
8ec60c3f 858 need_lwkt_resched();
96728c05
MD
859 return;
860 }
861#endif
41a01a4d 862 /*
d3d1cbc8 863 * Take the easy way out and do not preempt if we are holding
d666840a 864 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
865 * preempted interlock against the target thread's tokens and whether
866 * we can get all the target thread's tokens, but this situation
867 * should not occur very often so its easier to simply not preempt.
d666840a
MD
868 * Also, plain spinlocks are impossible to figure out at this point so
869 * just don't preempt.
d3d1cbc8
MD
870 *
871 * Do not try to preempt if the target thread is holding any tokens.
872 * We could try to acquire the tokens but this case is so rare there
873 * is no need to support it.
41a01a4d 874 */
bbb31c5d 875 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 876 ++preempt_miss;
8ec60c3f 877 need_lwkt_resched();
41a01a4d
MD
878 return;
879 }
3b998fa9 880 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
881 ++preempt_miss;
882 need_lwkt_resched();
883 return;
884 }
26a0694b
MD
885 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
886 ++preempt_weird;
8ec60c3f 887 need_lwkt_resched();
26a0694b
MD
888 return;
889 }
890 if (ntd->td_preempted) {
4b5f931b 891 ++preempt_hit;
8ec60c3f 892 need_lwkt_resched();
26a0694b 893 return;
b68b7282 894 }
8a8d5d85 895#ifdef SMP
a2a5ad0d
MD
896 /*
897 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
898 * to or from the MP lock. In this case td_mpcount will be pre-disposed
899 * (non-zero) but not actually synchronized with the actual state of the
900 * lock. We can use it to imply an MP lock requirement for the
901 * preemption but we cannot use it to test whether we hold the MP lock
902 * or not.
a2a5ad0d 903 */
96728c05 904 savecnt = td->td_mpcount;
71ef2f5c 905 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
906 ntd->td_mpcount += td->td_mpcount;
907 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
908 ntd->td_mpcount -= td->td_mpcount;
909 ++preempt_miss;
8ec60c3f 910 need_lwkt_resched();
8a8d5d85
MD
911 return;
912 }
913#endif
26a0694b 914
8ec60c3f
MD
915 /*
916 * Since we are able to preempt the current thread, there is no need to
917 * call need_lwkt_resched().
918 */
26a0694b
MD
919 ++preempt_hit;
920 ntd->td_preempted = td;
921 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 922 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 923 td->td_switch(ntd);
b9eb1c19 924
26a0694b 925 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
926#ifdef SMP
927 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
928 mpheld = MP_LOCK_HELD();
929 if (mpheld && td->td_mpcount == 0)
96728c05 930 cpu_rel_mplock();
71ef2f5c 931 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
932 panic("lwkt_preempt(): MP lock was not held through");
933#endif
26a0694b
MD
934 ntd->td_preempted = NULL;
935 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
936}
937
f1d1c3fa 938/*
faaeffac 939 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 940 *
faaeffac
MD
941 * td_nest_count prevents deep nesting via splz() or doreti() which
942 * might otherwise blow out the kernel stack. Note that except for
943 * this special case, we MUST call splz() here to handle any
944 * pending ints, particularly after we switch, or we might accidently
945 * halt the cpu with interrupts pending.
ef0fdad1 946 *
f1d1c3fa
MD
947 * (self contained on a per cpu basis)
948 */
949void
faaeffac 950splz_check(void)
f1d1c3fa 951{
7966cb69
MD
952 globaldata_t gd = mycpu;
953 thread_t td = gd->gd_curthread;
ef0fdad1 954
46a3f46d 955 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 956 splz();
f1d1c3fa
MD
957}
958
8ad65e08 959/*
faaeffac
MD
960 * This implements a normal yield which will yield to equal priority
961 * threads as well as higher priority threads. Note that gd_reqflags
962 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
963 *
964 * (self contained on a per cpu basis)
8ad65e08
MD
965 */
966void
f1d1c3fa 967lwkt_yield(void)
8ad65e08 968{
37af14fe 969 lwkt_schedule_self(curthread);
f1d1c3fa
MD
970 lwkt_switch();
971}
972
3824f392
MD
973/*
974 * This function is used along with the lwkt_passive_recover() inline
975 * by the trap code to negotiate a passive release of the current
976 * process/lwp designation with the user scheduler.
977 */
978void
979lwkt_passive_release(struct thread *td)
980{
981 struct lwp *lp = td->td_lwp;
982
983 td->td_release = NULL;
984 lwkt_setpri_self(TDPRI_KERN_USER);
985 lp->lwp_proc->p_usched->release_curproc(lp);
986}
987
988/*
989 * Make a kernel thread act as if it were in user mode with regards
990 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
991 * loops which may be potentially cpu-bound can call lwkt_user_yield().
992 *
993 * The lwkt_user_yield() function is designed to have very low overhead
994 * if no yield is determined to be needed.
995 */
996void
997lwkt_user_yield(void)
998{
999 thread_t td = curthread;
1000 struct lwp *lp = td->td_lwp;
1001
1002#ifdef SMP
1003 /*
1004 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1005 * kernel can prevent other cpus from servicing interrupt threads
1006 * which still require the MP lock (which is a lot of them). This
1007 * has a chaining effect since if the interrupt is blocked, so is
1008 * the event, so normal scheduling will not pick up on the problem.
1009 */
684a93c4
MD
1010 if (mp_lock_contention_mask && td->td_mpcount) {
1011 yield_mplock(td);
3824f392
MD
1012 }
1013#endif
1014
1015 /*
1016 * Another kernel thread wants the cpu
1017 */
1018 if (lwkt_resched_wanted())
1019 lwkt_switch();
1020
1021 /*
1022 * If the user scheduler has asynchronously determined that the current
1023 * process (when running in user mode) needs to lose the cpu then make
1024 * sure we are released.
1025 */
1026 if (user_resched_wanted()) {
1027 if (td->td_release)
1028 td->td_release(td);
1029 }
1030
1031 /*
1032 * If we are released reduce our priority
1033 */
1034 if (td->td_release == NULL) {
1035 if (lwkt_check_resched(td) > 0)
1036 lwkt_switch();
e381e77c
MD
1037 if (lp) {
1038 lp->lwp_proc->p_usched->acquire_curproc(lp);
1039 td->td_release = lwkt_passive_release;
1040 lwkt_setpri_self(TDPRI_USER_NORM);
1041 }
3824f392
MD
1042 }
1043}
1044
b9eb1c19
MD
1045/*
1046 * Return 0 if no runnable threads are pending at the same or higher
1047 * priority as the passed thread.
1048 *
1049 * Return 1 if runnable threads are pending at the same priority.
1050 *
1051 * Return 2 if runnable threads are pending at a higher priority.
1052 */
1053int
1054lwkt_check_resched(thread_t td)
1055{
1056 int pri = td->td_pri & TDPRI_MASK;
1057
1058 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1059 return(2);
1060 if (TAILQ_NEXT(td, td_threadq))
1061 return(1);
1062 return(0);
1063}
1064
8ad65e08 1065/*
f1d1c3fa
MD
1066 * Generic schedule. Possibly schedule threads belonging to other cpus and
1067 * deal with threads that might be blocked on a wait queue.
1068 *
0a3f9b47
MD
1069 * We have a little helper inline function which does additional work after
1070 * the thread has been enqueued, including dealing with preemption and
1071 * setting need_lwkt_resched() (which prevents the kernel from returning
1072 * to userland until it has processed higher priority threads).
6330a558
MD
1073 *
1074 * It is possible for this routine to be called after a failed _enqueue
1075 * (due to the target thread migrating, sleeping, or otherwise blocked).
1076 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1077 *
1078 * reschedok is an optimized constant propagated from lwkt_schedule() or
1079 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1080 * reschedule to be requested if the target thread has a higher priority.
1081 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1082 * be 0, prevented undesired reschedules.
8ad65e08 1083 */
0a3f9b47
MD
1084static __inline
1085void
361d01dd 1086_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1087{
b9eb1c19 1088 thread_t otd;
c730be20 1089
6330a558 1090 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1091 if (ntd->td_preemptable && reschedok) {
6330a558 1092 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1093 } else if (reschedok) {
b9eb1c19
MD
1094 otd = curthread;
1095 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1096 need_lwkt_resched();
6330a558 1097 }
0a3f9b47
MD
1098 }
1099}
1100
361d01dd 1101static __inline
8ad65e08 1102void
361d01dd 1103_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1104{
37af14fe
MD
1105 globaldata_t mygd = mycpu;
1106
41a01a4d 1107 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1108 crit_enter_gd(mygd);
9388413d 1109 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1110 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1111 _lwkt_enqueue(td);
1112 } else {
f1d1c3fa 1113 /*
7cd8d145
MD
1114 * If we own the thread, there is no race (since we are in a
1115 * critical section). If we do not own the thread there might
1116 * be a race but the target cpu will deal with it.
f1d1c3fa 1117 */
0f7a3396 1118#ifdef SMP
7cd8d145 1119 if (td->td_gd == mygd) {
9d265729 1120 _lwkt_enqueue(td);
361d01dd 1121 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1122 } else {
e381e77c 1123 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1124 }
0f7a3396 1125#else
7cd8d145 1126 _lwkt_enqueue(td);
361d01dd 1127 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1128#endif
8ad65e08 1129 }
37af14fe 1130 crit_exit_gd(mygd);
8ad65e08
MD
1131}
1132
361d01dd
MD
1133void
1134lwkt_schedule(thread_t td)
1135{
1136 _lwkt_schedule(td, 1);
1137}
1138
1139void
1140lwkt_schedule_noresched(thread_t td)
1141{
1142 _lwkt_schedule(td, 0);
1143}
1144
88ebb169
SW
1145#ifdef SMP
1146
e381e77c
MD
1147/*
1148 * When scheduled remotely if frame != NULL the IPIQ is being
1149 * run via doreti or an interrupt then preemption can be allowed.
1150 *
1151 * To allow preemption we have to drop the critical section so only
1152 * one is present in _lwkt_schedule_post.
1153 */
1154static void
1155lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1156{
1157 thread_t td = curthread;
1158 thread_t ntd = arg;
1159
1160 if (frame && ntd->td_preemptable) {
1161 crit_exit_noyield(td);
1162 _lwkt_schedule(ntd, 1);
1163 crit_enter_quick(td);
1164 } else {
1165 _lwkt_schedule(ntd, 1);
1166 }
1167}
1168
d9eea1a5 1169/*
52eedfb5
MD
1170 * Thread migration using a 'Pull' method. The thread may or may not be
1171 * the current thread. It MUST be descheduled and in a stable state.
1172 * lwkt_giveaway() must be called on the cpu owning the thread.
1173 *
1174 * At any point after lwkt_giveaway() is called, the target cpu may
1175 * 'pull' the thread by calling lwkt_acquire().
1176 *
ae8e83e6
MD
1177 * We have to make sure the thread is not sitting on a per-cpu tsleep
1178 * queue or it will blow up when it moves to another cpu.
1179 *
52eedfb5 1180 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1181 */
52eedfb5
MD
1182void
1183lwkt_giveaway(thread_t td)
1184{
3b4192fb 1185 globaldata_t gd = mycpu;
52eedfb5 1186
3b4192fb
MD
1187 crit_enter_gd(gd);
1188 if (td->td_flags & TDF_TSLEEPQ)
1189 tsleep_remove(td);
1190 KKASSERT(td->td_gd == gd);
1191 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1192 td->td_flags |= TDF_MIGRATING;
1193 crit_exit_gd(gd);
52eedfb5
MD
1194}
1195
a2a5ad0d
MD
1196void
1197lwkt_acquire(thread_t td)
1198{
37af14fe
MD
1199 globaldata_t gd;
1200 globaldata_t mygd;
a2a5ad0d 1201
52eedfb5 1202 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1203 gd = td->td_gd;
37af14fe 1204 mygd = mycpu;
52eedfb5 1205 if (gd != mycpu) {
35238fa5 1206 cpu_lfence();
52eedfb5 1207 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1208 crit_enter_gd(mygd);
df910c23
MD
1209 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1210#ifdef SMP
1211 lwkt_process_ipiq();
1212#endif
52eedfb5 1213 cpu_lfence();
df910c23 1214 }
37af14fe 1215 td->td_gd = mygd;
52eedfb5
MD
1216 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1217 td->td_flags &= ~TDF_MIGRATING;
1218 crit_exit_gd(mygd);
1219 } else {
1220 crit_enter_gd(mygd);
1221 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1222 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1223 crit_exit_gd(mygd);
a2a5ad0d
MD
1224 }
1225}
1226
52eedfb5
MD
1227#endif
1228
f1d1c3fa
MD
1229/*
1230 * Generic deschedule. Descheduling threads other then your own should be
1231 * done only in carefully controlled circumstances. Descheduling is
1232 * asynchronous.
1233 *
1234 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1235 */
1236void
1237lwkt_deschedule(thread_t td)
1238{
f1d1c3fa 1239 crit_enter();
b8a98473 1240#ifdef SMP
f1d1c3fa
MD
1241 if (td == curthread) {
1242 _lwkt_dequeue(td);
1243 } else {
a72187e9 1244 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1245 _lwkt_dequeue(td);
1246 } else {
b8a98473 1247 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1248 }
1249 }
b8a98473
MD
1250#else
1251 _lwkt_dequeue(td);
1252#endif
f1d1c3fa
MD
1253 crit_exit();
1254}
1255
4b5f931b
MD
1256/*
1257 * Set the target thread's priority. This routine does not automatically
1258 * switch to a higher priority thread, LWKT threads are not designed for
1259 * continuous priority changes. Yield if you want to switch.
1260 *
1261 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1262 * of the td_pri field. The specified priority may also indicate zero or
1263 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1264 *
1265 * Note that we requeue the thread whether it winds up on a different runq
1266 * or not. uio_yield() depends on this and the routine is not normally
1267 * called with the same priority otherwise.
4b5f931b
MD
1268 */
1269void
1270lwkt_setpri(thread_t td, int pri)
1271{
26a0694b 1272 KKASSERT(pri >= 0);
a72187e9 1273 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1274 crit_enter();
1275 if (td->td_flags & TDF_RUNQ) {
1276 _lwkt_dequeue(td);
1277 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1278 _lwkt_enqueue(td);
1279 } else {
1280 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1281 }
1282 crit_exit();
1283}
1284
03bd0a5e
MD
1285/*
1286 * Set the initial priority for a thread prior to it being scheduled for
1287 * the first time. The thread MUST NOT be scheduled before or during
1288 * this call. The thread may be assigned to a cpu other then the current
1289 * cpu.
1290 *
1291 * Typically used after a thread has been created with TDF_STOPPREQ,
1292 * and before the thread is initially scheduled.
1293 */
1294void
1295lwkt_setpri_initial(thread_t td, int pri)
1296{
1297 KKASSERT(pri >= 0);
1298 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1299 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1300}
1301
26a0694b
MD
1302void
1303lwkt_setpri_self(int pri)
1304{
1305 thread_t td = curthread;
1306
4b5f931b
MD
1307 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1308 crit_enter();
1309 if (td->td_flags & TDF_RUNQ) {
1310 _lwkt_dequeue(td);
1311 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1312 _lwkt_enqueue(td);
1313 } else {
1314 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1315 }
1316 crit_exit();
1317}
1318
5d21b981 1319/*
52eedfb5
MD
1320 * Migrate the current thread to the specified cpu.
1321 *
1322 * This is accomplished by descheduling ourselves from the current cpu,
1323 * moving our thread to the tdallq of the target cpu, IPI messaging the
1324 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1325 * races while the thread is being migrated.
ae8e83e6
MD
1326 *
1327 * We must be sure to remove ourselves from the current cpu's tsleepq
1328 * before potentially moving to another queue. The thread can be on
1329 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1330 */
3d28ff59 1331#ifdef SMP
5d21b981 1332static void lwkt_setcpu_remote(void *arg);
3d28ff59 1333#endif
5d21b981
MD
1334
1335void
1336lwkt_setcpu_self(globaldata_t rgd)
1337{
1338#ifdef SMP
1339 thread_t td = curthread;
1340
1341 if (td->td_gd != rgd) {
1342 crit_enter_quick(td);
ae8e83e6 1343 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1344 tsleep_remove(td);
5d21b981
MD
1345 td->td_flags |= TDF_MIGRATING;
1346 lwkt_deschedule_self(td);
52eedfb5 1347 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1348 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1349 lwkt_switch();
1350 /* we are now on the target cpu */
52eedfb5 1351 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1352 crit_exit_quick(td);
1353 }
1354#endif
1355}
1356
ecdefdda
MD
1357void
1358lwkt_migratecpu(int cpuid)
1359{
1360#ifdef SMP
1361 globaldata_t rgd;
1362
1363 rgd = globaldata_find(cpuid);
1364 lwkt_setcpu_self(rgd);
1365#endif
1366}
1367
5d21b981
MD
1368/*
1369 * Remote IPI for cpu migration (called while in a critical section so we
1370 * do not have to enter another one). The thread has already been moved to
1371 * our cpu's allq, but we must wait for the thread to be completely switched
1372 * out on the originating cpu before we schedule it on ours or the stack
1373 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1374 * change to main memory.
1375 *
1376 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1377 * against wakeups. It is best if this interface is used only when there
1378 * are no pending events that might try to schedule the thread.
1379 */
3d28ff59 1380#ifdef SMP
5d21b981
MD
1381static void
1382lwkt_setcpu_remote(void *arg)
1383{
1384 thread_t td = arg;
1385 globaldata_t gd = mycpu;
1386
df910c23
MD
1387 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1388#ifdef SMP
1389 lwkt_process_ipiq();
1390#endif
35238fa5 1391 cpu_lfence();
df910c23 1392 }
5d21b981 1393 td->td_gd = gd;
35238fa5 1394 cpu_sfence();
5d21b981 1395 td->td_flags &= ~TDF_MIGRATING;
9388413d 1396 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1397 _lwkt_enqueue(td);
1398}
3d28ff59 1399#endif
5d21b981 1400
553ea3c8 1401struct lwp *
4b5f931b
MD
1402lwkt_preempted_proc(void)
1403{
73e4f7b9 1404 thread_t td = curthread;
4b5f931b
MD
1405 while (td->td_preempted)
1406 td = td->td_preempted;
553ea3c8 1407 return(td->td_lwp);
4b5f931b
MD
1408}
1409
99df837e
MD
1410/*
1411 * Create a kernel process/thread/whatever. It shares it's address space
1412 * with proc0 - ie: kernel only.
1413 *
365fa13f
MD
1414 * NOTE! By default new threads are created with the MP lock held. A
1415 * thread which does not require the MP lock should release it by calling
1416 * rel_mplock() at the start of the new thread.
99df837e
MD
1417 */
1418int
1419lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1420 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1421 const char *fmt, ...)
99df837e 1422{
73e4f7b9 1423 thread_t td;
e2565a42 1424 __va_list ap;
99df837e 1425
d3d32139 1426 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1427 tdflags);
a2a5ad0d
MD
1428 if (tdp)
1429 *tdp = td;
709799ea 1430 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1431
1432 /*
1433 * Set up arg0 for 'ps' etc
1434 */
e2565a42 1435 __va_start(ap, fmt);
379210cb 1436 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1437 __va_end(ap);
99df837e
MD
1438
1439 /*
1440 * Schedule the thread to run
1441 */
ef0fdad1
MD
1442 if ((td->td_flags & TDF_STOPREQ) == 0)
1443 lwkt_schedule(td);
1444 else
1445 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1446 return 0;
1447}
1448
1449/*
1450 * Destroy an LWKT thread. Warning! This function is not called when
1451 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1452 * uses a different reaping mechanism.
1453 */
1454void
1455lwkt_exit(void)
1456{
1457 thread_t td = curthread;
c070746a 1458 thread_t std;
8826f33a 1459 globaldata_t gd;
99df837e
MD
1460
1461 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1462 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1463 caps_exit(td);
c070746a
MD
1464
1465 /*
1466 * Get us into a critical section to interlock gd_freetd and loop
1467 * until we can get it freed.
1468 *
1469 * We have to cache the current td in gd_freetd because objcache_put()ing
1470 * it would rip it out from under us while our thread is still active.
1471 */
1472 gd = mycpu;
37af14fe 1473 crit_enter_quick(td);
c070746a
MD
1474 while ((std = gd->gd_freetd) != NULL) {
1475 gd->gd_freetd = NULL;
1476 objcache_put(thread_cache, std);
1477 }
3b4192fb
MD
1478
1479 /*
1480 * Remove thread resources from kernel lists and deschedule us for
1481 * the last time.
1482 */
1483 if (td->td_flags & TDF_TSLEEPQ)
1484 tsleep_remove(td);
79eae878 1485 biosched_done(td);
f8abf63c 1486 dsched_exit_thread(td);
37af14fe 1487 lwkt_deschedule_self(td);
e56e4dea 1488 lwkt_remove_tdallq(td);
c070746a
MD
1489 if (td->td_flags & TDF_ALLOCATED_THREAD)
1490 gd->gd_freetd = td;
99df837e
MD
1491 cpu_thread_exit();
1492}
1493
e56e4dea
MD
1494void
1495lwkt_remove_tdallq(thread_t td)
1496{
1497 KKASSERT(td->td_gd == mycpu);
1498 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1499}
1500
2d93b37a
MD
1501void
1502crit_panic(void)
1503{
1504 thread_t td = curthread;
1505 int lpri = td->td_pri;
1506
1507 td->td_pri = 0;
1508 panic("td_pri is/would-go negative! %p %d", td, lpri);
1509}
1510
d165e668
MD
1511#ifdef SMP
1512
bd8015ca
MD
1513/*
1514 * Called from debugger/panic on cpus which have been stopped. We must still
1515 * process the IPIQ while stopped, even if we were stopped while in a critical
1516 * section (XXX).
1517 *
1518 * If we are dumping also try to process any pending interrupts. This may
1519 * or may not work depending on the state of the cpu at the point it was
1520 * stopped.
1521 */
1522void
1523lwkt_smp_stopped(void)
1524{
1525 globaldata_t gd = mycpu;
1526
1527 crit_enter_gd(gd);
1528 if (dumping) {
1529 lwkt_process_ipiq();
1530 splz();
1531 } else {
1532 lwkt_process_ipiq();
1533 }
1534 crit_exit_gd(gd);
1535}
1536
d165e668 1537#endif