vesa(4): Remove definition and usage of the obsolete V_MODEOPTINFO flag.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf 2 * Copyright (c) 2003,2004 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, "");
4b5f931b 141SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
144#ifdef INVARIANTS
145SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
146 &token_contention_count, 0, "spinning due to token contention");
38717797 147#endif
05220613 148
4b5f931b
MD
149/*
150 * These helper procedures handle the runq, they can only be called from
151 * within a critical section.
75cdbe6c
MD
152 *
153 * WARNING! Prior to SMP being brought up it is possible to enqueue and
154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
155 * instead of 'mycpu' when referencing the globaldata structure. Once
156 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 157 */
f1d1c3fa
MD
158static __inline
159void
160_lwkt_dequeue(thread_t td)
161{
162 if (td->td_flags & TDF_RUNQ) {
4b5f931b 163 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 164 struct globaldata *gd = td->td_gd;
4b5f931b 165
f1d1c3fa 166 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
168 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
169 }
170}
171
172static __inline
173void
174_lwkt_enqueue(thread_t td)
175{
7f5d7ed7 176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 177 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 178 struct globaldata *gd = td->td_gd;
4b5f931b 179
f1d1c3fa 180 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
182 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
183 }
184}
8ad65e08 185
40aaf5fc
NT
186static __boolean_t
187_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
188{
189 struct thread *td = (struct thread *)obj;
190
191 td->td_kstack = NULL;
192 td->td_kstack_size = 0;
193 td->td_flags = TDF_ALLOCATED_THREAD;
194 return (1);
195}
196
197static void
198_lwkt_thread_dtor(void *obj, void *privdata)
199{
200 struct thread *td = (struct thread *)obj;
201
202 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
203 ("_lwkt_thread_dtor: not allocated from objcache"));
204 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
205 td->td_kstack_size > 0,
206 ("_lwkt_thread_dtor: corrupted stack"));
207 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
208}
209
210/*
211 * Initialize the lwkt s/system.
212 */
213void
214lwkt_init(void)
215{
216 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
217 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
218 NULL, CACHE_NTHREADS/2,
219 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
220}
221
37af14fe
MD
222/*
223 * Schedule a thread to run. As the current thread we can always safely
224 * schedule ourselves, and a shortcut procedure is provided for that
225 * function.
226 *
227 * (non-blocking, self contained on a per cpu basis)
228 */
229void
230lwkt_schedule_self(thread_t td)
231{
232 crit_enter_quick(td);
37af14fe 233 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 234 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 235 _lwkt_enqueue(td);
37af14fe
MD
236 crit_exit_quick(td);
237}
238
239/*
240 * Deschedule a thread.
241 *
242 * (non-blocking, self contained on a per cpu basis)
243 */
244void
245lwkt_deschedule_self(thread_t td)
246{
247 crit_enter_quick(td);
37af14fe
MD
248 _lwkt_dequeue(td);
249 crit_exit_quick(td);
250}
251
8ad65e08
MD
252/*
253 * LWKTs operate on a per-cpu basis
254 *
73e4f7b9 255 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
256 */
257void
258lwkt_gdinit(struct globaldata *gd)
259{
4b5f931b
MD
260 int i;
261
262 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
263 TAILQ_INIT(&gd->gd_tdrunq[i]);
264 gd->gd_runqmask = 0;
73e4f7b9 265 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
266}
267
7d0bac62
MD
268/*
269 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
270 * or LWKT start address before it can be scheduled. If the target cpu is
271 * -1 the thread will be created on the current cpu.
0cfcada1
MD
272 *
273 * If you intend to create a thread without a process context this function
274 * does everything except load the startup and switcher function.
7d0bac62
MD
275 */
276thread_t
d3d32139 277lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 278{
c070746a 279 globaldata_t gd = mycpu;
99df837e 280 void *stack;
7d0bac62 281
c070746a
MD
282 /*
283 * If static thread storage is not supplied allocate a thread. Reuse
284 * a cached free thread if possible. gd_freetd is used to keep an exiting
285 * thread intact through the exit.
286 */
ef0fdad1 287 if (td == NULL) {
c070746a
MD
288 if ((td = gd->gd_freetd) != NULL)
289 gd->gd_freetd = NULL;
290 else
291 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
292 KASSERT((td->td_flags &
293 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
294 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
295 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 296 }
c070746a
MD
297
298 /*
299 * Try to reuse cached stack.
300 */
f470d0c8
MD
301 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
302 if (flags & TDF_ALLOCATED_STACK) {
e4846942 303 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
304 stack = NULL;
305 }
306 }
307 if (stack == NULL) {
e4846942 308 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 309 flags |= TDF_ALLOCATED_STACK;
99df837e 310 }
75cdbe6c 311 if (cpu < 0)
c070746a 312 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 313 else
f470d0c8 314 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 315 return(td);
7d0bac62
MD
316}
317
318/*
319 * Initialize a preexisting thread structure. This function is used by
320 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
321 *
f8c3996b
MD
322 * All threads start out in a critical section at a priority of
323 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
324 * appropriate. This function may send an IPI message when the
325 * requested cpu is not the current cpu and consequently gd_tdallq may
326 * not be initialized synchronously from the point of view of the originating
327 * cpu.
328 *
329 * NOTE! we have to be careful in regards to creating threads for other cpus
330 * if SMP has not yet been activated.
7d0bac62 331 */
41a01a4d
MD
332#ifdef SMP
333
75cdbe6c
MD
334static void
335lwkt_init_thread_remote(void *arg)
336{
337 thread_t td = arg;
338
52eedfb5
MD
339 /*
340 * Protected by critical section held by IPI dispatch
341 */
75cdbe6c
MD
342 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
343}
344
41a01a4d
MD
345#endif
346
7d0bac62 347void
f470d0c8
MD
348lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
349 struct globaldata *gd)
7d0bac62 350{
37af14fe
MD
351 globaldata_t mygd = mycpu;
352
99df837e
MD
353 bzero(td, sizeof(struct thread));
354 td->td_kstack = stack;
f470d0c8 355 td->td_kstack_size = stksize;
d3d32139 356 td->td_flags = flags;
26a0694b 357 td->td_gd = gd;
f8c3996b 358 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
359#ifdef SMP
360 if ((flags & TDF_MPSAFE) == 0)
361 td->td_mpcount = 1;
362#endif
fb0f29c4
MD
363 if (lwkt_use_spin_port)
364 lwkt_initport_spin(&td->td_msgport);
365 else
366 lwkt_initport_thread(&td->td_msgport, td);
99df837e 367 pmap_init_thread(td);
0f7a3396 368#ifdef SMP
5d21b981
MD
369 /*
370 * Normally initializing a thread for a remote cpu requires sending an
371 * IPI. However, the idlethread is setup before the other cpus are
372 * activated so we have to treat it as a special case. XXX manipulation
373 * of gd_tdallq requires the BGL.
374 */
375 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 376 crit_enter_gd(mygd);
75cdbe6c 377 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 378 crit_exit_gd(mygd);
75cdbe6c 379 } else {
2db3b277 380 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 381 }
0f7a3396 382#else
37af14fe 383 crit_enter_gd(mygd);
0f7a3396 384 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 385 crit_exit_gd(mygd);
0f7a3396 386#endif
8c72e3d5
AH
387
388 dsched_new_thread(td);
73e4f7b9
MD
389}
390
391void
392lwkt_set_comm(thread_t td, const char *ctl, ...)
393{
e2565a42 394 __va_list va;
73e4f7b9 395
e2565a42 396 __va_start(va, ctl);
379210cb 397 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 398 __va_end(va);
e7c0dbba 399 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
400}
401
99df837e 402void
73e4f7b9 403lwkt_hold(thread_t td)
99df837e 404{
73e4f7b9
MD
405 ++td->td_refs;
406}
407
408void
409lwkt_rele(thread_t td)
410{
411 KKASSERT(td->td_refs > 0);
412 --td->td_refs;
413}
414
415void
416lwkt_wait_free(thread_t td)
417{
418 while (td->td_refs)
377d4740 419 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
420}
421
422void
423lwkt_free_thread(thread_t td)
424{
d9eea1a5 425 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
426 ("lwkt_free_thread: did not exit! %p", td));
427
40aaf5fc
NT
428 if (td->td_flags & TDF_ALLOCATED_THREAD) {
429 objcache_put(thread_cache, td);
430 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
431 /* client-allocated struct with internally allocated stack */
432 KASSERT(td->td_kstack && td->td_kstack_size > 0,
433 ("lwkt_free_thread: corrupted stack"));
434 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
435 td->td_kstack = NULL;
436 td->td_kstack_size = 0;
99df837e 437 }
e7c0dbba 438 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
439}
440
441
8ad65e08
MD
442/*
443 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
444 * switch to the idlethread. Switching must occur within a critical
445 * section to avoid races with the scheduling queue.
446 *
447 * We always have full control over our cpu's run queue. Other cpus
448 * that wish to manipulate our queue must use the cpu_*msg() calls to
449 * talk to our cpu, so a critical section is all that is needed and
450 * the result is very, very fast thread switching.
451 *
96728c05
MD
452 * The LWKT scheduler uses a fixed priority model and round-robins at
453 * each priority level. User process scheduling is a totally
454 * different beast and LWKT priorities should not be confused with
455 * user process priorities.
f1d1c3fa 456 *
96728c05
MD
457 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
458 * cleans it up. Note that the td_switch() function cannot do anything that
459 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
460 * the target thread (not the current thread). It's nice to have a scheduler
461 * that does not need the MP lock to work because it allows us to do some
462 * really cool high-performance MP lock optimizations.
69d78e99
MD
463 *
464 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
465 * is not called by the current thread in the preemption case, only when
466 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
467 */
468void
469lwkt_switch(void)
470{
37af14fe
MD
471 globaldata_t gd = mycpu;
472 thread_t td = gd->gd_curthread;
8ad65e08 473 thread_t ntd;
8a8d5d85
MD
474#ifdef SMP
475 int mpheld;
476#endif
8ad65e08 477
46a3f46d 478 /*
27e88a6e
MD
479 * Switching from within a 'fast' (non thread switched) interrupt or IPI
480 * is illegal. However, we may have to do it anyway if we hit a fatal
481 * kernel trap or we have paniced.
482 *
483 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 484 */
27e88a6e
MD
485 if (gd->gd_intr_nesting_level) {
486 int savegdnest;
487 int savegdtrap;
488
489 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
490 panic("lwkt_switch: cannot switch from within "
491 "a fast interrupt, yet, td %p\n", td);
492 } else {
493 savegdnest = gd->gd_intr_nesting_level;
494 savegdtrap = gd->gd_trap_nesting_level;
495 gd->gd_intr_nesting_level = 0;
496 gd->gd_trap_nesting_level = 0;
a7422615
MD
497 if ((td->td_flags & TDF_PANICWARN) == 0) {
498 td->td_flags |= TDF_PANICWARN;
6ea70f76 499 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 500 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 501 print_backtrace(-1);
a7422615 502 }
27e88a6e
MD
503 lwkt_switch();
504 gd->gd_intr_nesting_level = savegdnest;
505 gd->gd_trap_nesting_level = savegdtrap;
506 return;
507 }
96728c05 508 }
ef0fdad1 509
cb973d15
MD
510 /*
511 * Passive release (used to transition from user to kernel mode
512 * when we block or switch rather then when we enter the kernel).
513 * This function is NOT called if we are switching into a preemption
514 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
515 * our current process designation (if we have one) and become a true
516 * LWKT thread, and may also hand the current process designation to
517 * another process and schedule thread.
cb973d15
MD
518 */
519 if (td->td_release)
520 td->td_release(td);
521
37af14fe 522 crit_enter_gd(gd);
9d265729
MD
523 if (td->td_toks)
524 lwkt_relalltokens(td);
525
526 /*
b02926de
MD
527 * We had better not be holding any spin locks, but don't get into an
528 * endless panic loop.
9d265729 529 */
bbb31c5d
MD
530 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
531 ("lwkt_switch: still holding a shared spinlock %p!",
532 gd->gd_spinlock_rd));
d666840a
MD
533 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
534 ("lwkt_switch: still holding %d exclusive spinlocks!",
535 gd->gd_spinlocks_wr));
9d265729 536
8a8d5d85
MD
537
538#ifdef SMP
539 /*
540 * td_mpcount cannot be used to determine if we currently hold the
541 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
542 * to get the lock, and switch out if it can't. Our ownership of
543 * the actual lock will remain stable while we are in a critical section
544 * (but, of course, another cpu may own or release the lock so the
545 * actual value of mp_lock is not stable).
8a8d5d85
MD
546 */
547 mpheld = MP_LOCK_HELD();
0f7a3396
MD
548#ifdef INVARIANTS
549 if (td->td_cscount) {
6ea70f76 550 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
551 td);
552 if (panic_on_cscount)
553 panic("switching while mastering cpusync");
554 }
555#endif
8a8d5d85 556#endif
99df837e
MD
557 if ((ntd = td->td_preempted) != NULL) {
558 /*
559 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
560 * thread. This occurs transparently, whether the preempted thread
561 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
562 * itself).
563 *
564 * We have to setup the MP lock for the original thread after backing
565 * out the adjustment that was made to curthread when the original
566 * was preempted.
99df837e 567 */
26a0694b 568 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 569#ifdef SMP
96728c05 570 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 571 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
572 td, ntd, td->td_mpcount, ntd->td_mpcount);
573 }
8a8d5d85
MD
574 if (ntd->td_mpcount) {
575 td->td_mpcount -= ntd->td_mpcount;
576 KKASSERT(td->td_mpcount >= 0);
577 }
578#endif
26a0694b 579 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
580
581 /*
b9eb1c19
MD
582 * The interrupt may have woken a thread up, we need to properly
583 * set the reschedule flag if the originally interrupted thread is
584 * at a lower priority.
8ec60c3f
MD
585 */
586 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
587 need_lwkt_resched();
8a8d5d85 588 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 589 } else {
4b5f931b
MD
590 /*
591 * Priority queue / round-robin at each priority. Note that user
592 * processes run at a fixed, low priority and the user process
593 * scheduler deals with interactions between user processes
594 * by scheduling and descheduling them from the LWKT queue as
595 * necessary.
8a8d5d85
MD
596 *
597 * We have to adjust the MP lock for the target thread. If we
598 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
599 * thread that does not need the MP lock. If we cannot, we spin
600 * instead of HLT.
601 *
602 * A similar issue exists for the tokens held by the target thread.
603 * If we cannot obtain ownership of the tokens we cannot immediately
604 * schedule the thread.
605 */
606
8ec60c3f
MD
607 /*
608 * If an LWKT reschedule was requested, well that is what we are
609 * doing now so clear it.
610 */
611 clear_lwkt_resched();
4b5f931b
MD
612again:
613 if (gd->gd_runqmask) {
614 int nq = bsrl(gd->gd_runqmask);
615 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
616 gd->gd_runqmask &= ~(1 << nq);
617 goto again;
618 }
8a8d5d85 619#ifdef SMP
41a01a4d 620 /*
df6b8ba0
MD
621 * THREAD SELECTION FOR AN SMP MACHINE BUILD
622 *
41a01a4d
MD
623 * If the target needs the MP lock and we couldn't get it,
624 * or if the target is holding tokens and we could not
625 * gain ownership of the tokens, continue looking for a
626 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
627 *
628 * NOTE: the mpheld variable invalid after this conditional, it
629 * can change due to both cpu_try_mplock() returning success
9d265729 630 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
631 * we are trying to check the mpcount of a thread other then
632 * the current thread. Because of this, if the current thread
633 * is not holding td_mpcount, an IPI indirectly run via
9d265729 634 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 635 * cause the core MP lock to be released.
41a01a4d
MD
636 */
637 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 638 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 639 ) {
8a8d5d85 640 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
641
642 mpheld = MP_LOCK_HELD();
643 ntd = NULL;
8a8d5d85
MD
644 while (rqmask) {
645 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 646 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 647 /* spinning due to MP lock being held */
41a01a4d 648 continue;
38717797 649 }
a453459d
MD
650
651 /*
9d265729 652 * mpheld state invalid after getalltokens call returns
a453459d
MD
653 * failure, but the variable is only needed for
654 * the loop.
655 */
9d265729 656 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 657 /* spinning due to token contention */
38717797 658#ifdef INVARIANTS
a453459d 659 ++token_contention_count;
38717797 660#endif
a453459d 661 mpheld = MP_LOCK_HELD();
41a01a4d 662 continue;
38717797 663 }
41a01a4d 664 break;
8a8d5d85
MD
665 }
666 if (ntd)
667 break;
668 rqmask &= ~(1 << nq);
669 nq = bsrl(rqmask);
b9eb1c19
MD
670
671 /*
672 * We have two choices. We can either refuse to run a
673 * user thread when a kernel thread needs the MP lock
674 * but could not get it, or we can allow it to run but
675 * then expect an IPI (hopefully) later on to force a
676 * reschedule when the MP lock might become available.
677 */
678 if (nq < TDPRI_KERN_LPSCHED) {
684a93c4
MD
679 break; /* for now refuse to run */
680#if 0
b9eb1c19
MD
681 if (chain_mplock == 0)
682 break;
b9eb1c19 683 /* continue loop, allow user threads to be scheduled */
684a93c4 684#endif
b9eb1c19 685 }
8a8d5d85 686 }
684a93c4
MD
687
688 /*
689 * Case where a (kernel) thread needed the MP lock and could
690 * not get one, and we may or may not have found another
691 * thread which does not need the MP lock to run while
692 * we wait (ntd).
693 */
8a8d5d85 694 if (ntd == NULL) {
a2a5ad0d
MD
695 ntd = &gd->gd_idlethread;
696 ntd->td_flags |= TDF_IDLE_NOHLT;
684a93c4
MD
697 set_mplock_contention_mask(gd);
698 cpu_mplock_contested();
df6b8ba0 699 goto using_idle_thread;
8a8d5d85 700 } else {
684a93c4 701 clr_mplock_contention_mask(gd);
344ad853 702 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
703 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
704 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 }
706 } else {
684a93c4 707 clr_mplock_contention_mask(gd);
344ad853 708 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
709 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
711 }
712#else
df6b8ba0
MD
713 /*
714 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
715 * worry about tokens or the BGL. However, we still have
716 * to call lwkt_getalltokens() in order to properly detect
717 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 718 */
7eb611ef 719 lwkt_getalltokens(ntd);
344ad853 720 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
721 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
722 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 723#endif
4b5f931b 724 } else {
3c23a41a 725 /*
60f945af
MD
726 * We have nothing to run but only let the idle loop halt
727 * the cpu if there are no pending interrupts.
3c23a41a 728 */
a2a5ad0d 729 ntd = &gd->gd_idlethread;
60f945af 730 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 731 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 732#ifdef SMP
df6b8ba0
MD
733using_idle_thread:
734 /*
735 * The idle thread should not be holding the MP lock unless we
736 * are trapping in the kernel or in a panic. Since we select the
737 * idle thread unconditionally when no other thread is available,
738 * if the MP lock is desired during a panic or kernel trap, we
739 * have to loop in the scheduler until we get it.
740 */
741 if (ntd->td_mpcount) {
742 mpheld = MP_LOCK_HELD();
684a93c4 743 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
df6b8ba0 744 panic("Idle thread %p was holding the BGL!", ntd);
684a93c4 745 if (mpheld == 0)
df6b8ba0
MD
746 goto again;
747 }
a453459d 748#endif
4b5f931b 749 }
f1d1c3fa 750 }
26a0694b
MD
751 KASSERT(ntd->td_pri >= TDPRI_CRIT,
752 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
753
754 /*
755 * Do the actual switch. If the new target does not need the MP lock
756 * and we are holding it, release the MP lock. If the new target requires
757 * the MP lock we have already acquired it for the target.
758 */
759#ifdef SMP
760 if (ntd->td_mpcount == 0 ) {
761 if (MP_LOCK_HELD())
762 cpu_rel_mplock();
763 } else {
a453459d 764 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
765 }
766#endif
94f6d86e
MD
767 if (td != ntd) {
768 ++switch_count;
b2b3ffcd 769#ifdef __x86_64__
f64b567c
SW
770 {
771 int tos_ok __debugvar = jg_tos_ok(ntd);
772 KKASSERT(tos_ok);
773 }
85514115 774#endif
a1f0fb66 775 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 776 td->td_switch(ntd);
94f6d86e 777 }
37af14fe
MD
778 /* NOTE: current cpu may have changed after switch */
779 crit_exit_quick(td);
8ad65e08
MD
780}
781
b68b7282 782/*
96728c05
MD
783 * Request that the target thread preempt the current thread. Preemption
784 * only works under a specific set of conditions:
b68b7282 785 *
96728c05
MD
786 * - We are not preempting ourselves
787 * - The target thread is owned by the current cpu
788 * - We are not currently being preempted
789 * - The target is not currently being preempted
d3d1cbc8
MD
790 * - We are not holding any spin locks
791 * - The target thread is not holding any tokens
96728c05
MD
792 * - We are able to satisfy the target's MP lock requirements (if any).
793 *
794 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
795 * this is called via lwkt_schedule() through the td_preemptable callback.
796 * critpri is the managed critical priority that we should ignore in order
797 * to determine whether preemption is possible (aka usually just the crit
798 * priority of lwkt_schedule() itself).
b68b7282 799 *
26a0694b
MD
800 * XXX at the moment we run the target thread in a critical section during
801 * the preemption in order to prevent the target from taking interrupts
802 * that *WE* can't. Preemption is strictly limited to interrupt threads
803 * and interrupt-like threads, outside of a critical section, and the
804 * preempted source thread will be resumed the instant the target blocks
805 * whether or not the source is scheduled (i.e. preemption is supposed to
806 * be as transparent as possible).
4b5f931b 807 *
8a8d5d85
MD
808 * The target thread inherits our MP count (added to its own) for the
809 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
810 * MP lock during the preemption. Therefore, any preempting targets must be
811 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
812 * out of sync with the physical mp_lock, but we do not have to preserve
813 * the original ownership of the lock if it was out of synch (that is, we
814 * can leave it synchronized on return).
b68b7282
MD
815 */
816void
96728c05 817lwkt_preempt(thread_t ntd, int critpri)
b68b7282 818{
46a3f46d 819 struct globaldata *gd = mycpu;
0a3f9b47 820 thread_t td;
8a8d5d85
MD
821#ifdef SMP
822 int mpheld;
57c254db 823 int savecnt;
8a8d5d85 824#endif
b68b7282 825
26a0694b 826 /*
96728c05
MD
827 * The caller has put us in a critical section. We can only preempt
828 * if the caller of the caller was not in a critical section (basically
d666840a 829 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 830 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 831 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
832 *
833 * YYY The target thread must be in a critical section (else it must
834 * inherit our critical section? I dunno yet).
41a01a4d 835 *
0a3f9b47 836 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
837 */
838 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 839
0a3f9b47 840 td = gd->gd_curthread;
0a3f9b47 841 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
842 ++preempt_miss;
843 return;
844 }
96728c05
MD
845 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
846 ++preempt_miss;
8ec60c3f 847 need_lwkt_resched();
96728c05
MD
848 return;
849 }
850#ifdef SMP
46a3f46d 851 if (ntd->td_gd != gd) {
96728c05 852 ++preempt_miss;
8ec60c3f 853 need_lwkt_resched();
96728c05
MD
854 return;
855 }
856#endif
41a01a4d 857 /*
d3d1cbc8 858 * Take the easy way out and do not preempt if we are holding
d666840a 859 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
860 * preempted interlock against the target thread's tokens and whether
861 * we can get all the target thread's tokens, but this situation
862 * should not occur very often so its easier to simply not preempt.
d666840a
MD
863 * Also, plain spinlocks are impossible to figure out at this point so
864 * just don't preempt.
d3d1cbc8
MD
865 *
866 * Do not try to preempt if the target thread is holding any tokens.
867 * We could try to acquire the tokens but this case is so rare there
868 * is no need to support it.
41a01a4d 869 */
bbb31c5d 870 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 871 ++preempt_miss;
8ec60c3f 872 need_lwkt_resched();
41a01a4d
MD
873 return;
874 }
d3d1cbc8
MD
875 if (ntd->td_toks) {
876 ++preempt_miss;
877 need_lwkt_resched();
878 return;
879 }
26a0694b
MD
880 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
881 ++preempt_weird;
8ec60c3f 882 need_lwkt_resched();
26a0694b
MD
883 return;
884 }
885 if (ntd->td_preempted) {
4b5f931b 886 ++preempt_hit;
8ec60c3f 887 need_lwkt_resched();
26a0694b 888 return;
b68b7282 889 }
8a8d5d85 890#ifdef SMP
a2a5ad0d
MD
891 /*
892 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
893 * to or from the MP lock. In this case td_mpcount will be pre-disposed
894 * (non-zero) but not actually synchronized with the actual state of the
895 * lock. We can use it to imply an MP lock requirement for the
896 * preemption but we cannot use it to test whether we hold the MP lock
897 * or not.
a2a5ad0d 898 */
96728c05 899 savecnt = td->td_mpcount;
71ef2f5c 900 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
901 ntd->td_mpcount += td->td_mpcount;
902 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
903 ntd->td_mpcount -= td->td_mpcount;
904 ++preempt_miss;
8ec60c3f 905 need_lwkt_resched();
8a8d5d85
MD
906 return;
907 }
908#endif
26a0694b 909
8ec60c3f
MD
910 /*
911 * Since we are able to preempt the current thread, there is no need to
912 * call need_lwkt_resched().
913 */
26a0694b
MD
914 ++preempt_hit;
915 ntd->td_preempted = td;
916 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 917 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 918 td->td_switch(ntd);
b9eb1c19 919
26a0694b 920 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
921#ifdef SMP
922 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
923 mpheld = MP_LOCK_HELD();
924 if (mpheld && td->td_mpcount == 0)
96728c05 925 cpu_rel_mplock();
71ef2f5c 926 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
927 panic("lwkt_preempt(): MP lock was not held through");
928#endif
26a0694b
MD
929 ntd->td_preempted = NULL;
930 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
931}
932
f1d1c3fa 933/*
faaeffac 934 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 935 *
faaeffac
MD
936 * td_nest_count prevents deep nesting via splz() or doreti() which
937 * might otherwise blow out the kernel stack. Note that except for
938 * this special case, we MUST call splz() here to handle any
939 * pending ints, particularly after we switch, or we might accidently
940 * halt the cpu with interrupts pending.
ef0fdad1 941 *
f1d1c3fa
MD
942 * (self contained on a per cpu basis)
943 */
944void
faaeffac 945splz_check(void)
f1d1c3fa 946{
7966cb69
MD
947 globaldata_t gd = mycpu;
948 thread_t td = gd->gd_curthread;
ef0fdad1 949
46a3f46d 950 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 951 splz();
f1d1c3fa
MD
952}
953
8ad65e08 954/*
faaeffac
MD
955 * This implements a normal yield which will yield to equal priority
956 * threads as well as higher priority threads. Note that gd_reqflags
957 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
958 *
959 * (self contained on a per cpu basis)
8ad65e08
MD
960 */
961void
f1d1c3fa 962lwkt_yield(void)
8ad65e08 963{
37af14fe 964 lwkt_schedule_self(curthread);
f1d1c3fa
MD
965 lwkt_switch();
966}
967
3824f392
MD
968/*
969 * This function is used along with the lwkt_passive_recover() inline
970 * by the trap code to negotiate a passive release of the current
971 * process/lwp designation with the user scheduler.
972 */
973void
974lwkt_passive_release(struct thread *td)
975{
976 struct lwp *lp = td->td_lwp;
977
978 td->td_release = NULL;
979 lwkt_setpri_self(TDPRI_KERN_USER);
980 lp->lwp_proc->p_usched->release_curproc(lp);
981}
982
983/*
984 * Make a kernel thread act as if it were in user mode with regards
985 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
986 * loops which may be potentially cpu-bound can call lwkt_user_yield().
987 *
988 * The lwkt_user_yield() function is designed to have very low overhead
989 * if no yield is determined to be needed.
990 */
991void
992lwkt_user_yield(void)
993{
994 thread_t td = curthread;
995 struct lwp *lp = td->td_lwp;
996
997#ifdef SMP
998 /*
999 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1000 * kernel can prevent other cpus from servicing interrupt threads
1001 * which still require the MP lock (which is a lot of them). This
1002 * has a chaining effect since if the interrupt is blocked, so is
1003 * the event, so normal scheduling will not pick up on the problem.
1004 */
684a93c4
MD
1005 if (mp_lock_contention_mask && td->td_mpcount) {
1006 yield_mplock(td);
3824f392
MD
1007 }
1008#endif
1009
1010 /*
1011 * Another kernel thread wants the cpu
1012 */
1013 if (lwkt_resched_wanted())
1014 lwkt_switch();
1015
1016 /*
1017 * If the user scheduler has asynchronously determined that the current
1018 * process (when running in user mode) needs to lose the cpu then make
1019 * sure we are released.
1020 */
1021 if (user_resched_wanted()) {
1022 if (td->td_release)
1023 td->td_release(td);
1024 }
1025
1026 /*
1027 * If we are released reduce our priority
1028 */
1029 if (td->td_release == NULL) {
1030 if (lwkt_check_resched(td) > 0)
1031 lwkt_switch();
e381e77c
MD
1032 if (lp) {
1033 lp->lwp_proc->p_usched->acquire_curproc(lp);
1034 td->td_release = lwkt_passive_release;
1035 lwkt_setpri_self(TDPRI_USER_NORM);
1036 }
3824f392
MD
1037 }
1038}
1039
b9eb1c19
MD
1040/*
1041 * Return 0 if no runnable threads are pending at the same or higher
1042 * priority as the passed thread.
1043 *
1044 * Return 1 if runnable threads are pending at the same priority.
1045 *
1046 * Return 2 if runnable threads are pending at a higher priority.
1047 */
1048int
1049lwkt_check_resched(thread_t td)
1050{
1051 int pri = td->td_pri & TDPRI_MASK;
1052
1053 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1054 return(2);
1055 if (TAILQ_NEXT(td, td_threadq))
1056 return(1);
1057 return(0);
1058}
1059
8ad65e08 1060/*
f1d1c3fa
MD
1061 * Generic schedule. Possibly schedule threads belonging to other cpus and
1062 * deal with threads that might be blocked on a wait queue.
1063 *
0a3f9b47
MD
1064 * We have a little helper inline function which does additional work after
1065 * the thread has been enqueued, including dealing with preemption and
1066 * setting need_lwkt_resched() (which prevents the kernel from returning
1067 * to userland until it has processed higher priority threads).
6330a558
MD
1068 *
1069 * It is possible for this routine to be called after a failed _enqueue
1070 * (due to the target thread migrating, sleeping, or otherwise blocked).
1071 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1072 *
1073 * reschedok is an optimized constant propagated from lwkt_schedule() or
1074 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1075 * reschedule to be requested if the target thread has a higher priority.
1076 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1077 * be 0, prevented undesired reschedules.
8ad65e08 1078 */
0a3f9b47
MD
1079static __inline
1080void
361d01dd 1081_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1082{
b9eb1c19 1083 thread_t otd;
c730be20 1084
6330a558 1085 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1086 if (ntd->td_preemptable && reschedok) {
6330a558 1087 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1088 } else if (reschedok) {
b9eb1c19
MD
1089 otd = curthread;
1090 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1091 need_lwkt_resched();
6330a558 1092 }
0a3f9b47
MD
1093 }
1094}
1095
361d01dd 1096static __inline
8ad65e08 1097void
361d01dd 1098_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1099{
37af14fe
MD
1100 globaldata_t mygd = mycpu;
1101
41a01a4d 1102 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1103 crit_enter_gd(mygd);
9388413d 1104 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1105 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1106 _lwkt_enqueue(td);
1107 } else {
f1d1c3fa 1108 /*
7cd8d145
MD
1109 * If we own the thread, there is no race (since we are in a
1110 * critical section). If we do not own the thread there might
1111 * be a race but the target cpu will deal with it.
f1d1c3fa 1112 */
0f7a3396 1113#ifdef SMP
7cd8d145 1114 if (td->td_gd == mygd) {
9d265729 1115 _lwkt_enqueue(td);
361d01dd 1116 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1117 } else {
e381e77c 1118 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1119 }
0f7a3396 1120#else
7cd8d145 1121 _lwkt_enqueue(td);
361d01dd 1122 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1123#endif
8ad65e08 1124 }
37af14fe 1125 crit_exit_gd(mygd);
8ad65e08
MD
1126}
1127
361d01dd
MD
1128void
1129lwkt_schedule(thread_t td)
1130{
1131 _lwkt_schedule(td, 1);
1132}
1133
1134void
1135lwkt_schedule_noresched(thread_t td)
1136{
1137 _lwkt_schedule(td, 0);
1138}
1139
88ebb169
SW
1140#ifdef SMP
1141
e381e77c
MD
1142/*
1143 * When scheduled remotely if frame != NULL the IPIQ is being
1144 * run via doreti or an interrupt then preemption can be allowed.
1145 *
1146 * To allow preemption we have to drop the critical section so only
1147 * one is present in _lwkt_schedule_post.
1148 */
1149static void
1150lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1151{
1152 thread_t td = curthread;
1153 thread_t ntd = arg;
1154
1155 if (frame && ntd->td_preemptable) {
1156 crit_exit_noyield(td);
1157 _lwkt_schedule(ntd, 1);
1158 crit_enter_quick(td);
1159 } else {
1160 _lwkt_schedule(ntd, 1);
1161 }
1162}
1163
d9eea1a5 1164/*
52eedfb5
MD
1165 * Thread migration using a 'Pull' method. The thread may or may not be
1166 * the current thread. It MUST be descheduled and in a stable state.
1167 * lwkt_giveaway() must be called on the cpu owning the thread.
1168 *
1169 * At any point after lwkt_giveaway() is called, the target cpu may
1170 * 'pull' the thread by calling lwkt_acquire().
1171 *
ae8e83e6
MD
1172 * We have to make sure the thread is not sitting on a per-cpu tsleep
1173 * queue or it will blow up when it moves to another cpu.
1174 *
52eedfb5 1175 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1176 */
52eedfb5
MD
1177void
1178lwkt_giveaway(thread_t td)
1179{
3b4192fb 1180 globaldata_t gd = mycpu;
52eedfb5 1181
3b4192fb
MD
1182 crit_enter_gd(gd);
1183 if (td->td_flags & TDF_TSLEEPQ)
1184 tsleep_remove(td);
1185 KKASSERT(td->td_gd == gd);
1186 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1187 td->td_flags |= TDF_MIGRATING;
1188 crit_exit_gd(gd);
52eedfb5
MD
1189}
1190
a2a5ad0d
MD
1191void
1192lwkt_acquire(thread_t td)
1193{
37af14fe
MD
1194 globaldata_t gd;
1195 globaldata_t mygd;
a2a5ad0d 1196
52eedfb5 1197 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1198 gd = td->td_gd;
37af14fe 1199 mygd = mycpu;
52eedfb5 1200 if (gd != mycpu) {
35238fa5 1201 cpu_lfence();
52eedfb5 1202 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1203 crit_enter_gd(mygd);
df910c23
MD
1204 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1205#ifdef SMP
1206 lwkt_process_ipiq();
1207#endif
52eedfb5 1208 cpu_lfence();
df910c23 1209 }
37af14fe 1210 td->td_gd = mygd;
52eedfb5
MD
1211 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1212 td->td_flags &= ~TDF_MIGRATING;
1213 crit_exit_gd(mygd);
1214 } else {
1215 crit_enter_gd(mygd);
1216 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1217 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1218 crit_exit_gd(mygd);
a2a5ad0d
MD
1219 }
1220}
1221
52eedfb5
MD
1222#endif
1223
f1d1c3fa
MD
1224/*
1225 * Generic deschedule. Descheduling threads other then your own should be
1226 * done only in carefully controlled circumstances. Descheduling is
1227 * asynchronous.
1228 *
1229 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1230 */
1231void
1232lwkt_deschedule(thread_t td)
1233{
f1d1c3fa 1234 crit_enter();
b8a98473 1235#ifdef SMP
f1d1c3fa
MD
1236 if (td == curthread) {
1237 _lwkt_dequeue(td);
1238 } else {
a72187e9 1239 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1240 _lwkt_dequeue(td);
1241 } else {
b8a98473 1242 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1243 }
1244 }
b8a98473
MD
1245#else
1246 _lwkt_dequeue(td);
1247#endif
f1d1c3fa
MD
1248 crit_exit();
1249}
1250
4b5f931b
MD
1251/*
1252 * Set the target thread's priority. This routine does not automatically
1253 * switch to a higher priority thread, LWKT threads are not designed for
1254 * continuous priority changes. Yield if you want to switch.
1255 *
1256 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1257 * of the td_pri field. The specified priority may also indicate zero or
1258 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1259 *
1260 * Note that we requeue the thread whether it winds up on a different runq
1261 * or not. uio_yield() depends on this and the routine is not normally
1262 * called with the same priority otherwise.
4b5f931b
MD
1263 */
1264void
1265lwkt_setpri(thread_t td, int pri)
1266{
26a0694b 1267 KKASSERT(pri >= 0);
a72187e9 1268 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1269 crit_enter();
1270 if (td->td_flags & TDF_RUNQ) {
1271 _lwkt_dequeue(td);
1272 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1273 _lwkt_enqueue(td);
1274 } else {
1275 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1276 }
1277 crit_exit();
1278}
1279
03bd0a5e
MD
1280/*
1281 * Set the initial priority for a thread prior to it being scheduled for
1282 * the first time. The thread MUST NOT be scheduled before or during
1283 * this call. The thread may be assigned to a cpu other then the current
1284 * cpu.
1285 *
1286 * Typically used after a thread has been created with TDF_STOPPREQ,
1287 * and before the thread is initially scheduled.
1288 */
1289void
1290lwkt_setpri_initial(thread_t td, int pri)
1291{
1292 KKASSERT(pri >= 0);
1293 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1294 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1295}
1296
26a0694b
MD
1297void
1298lwkt_setpri_self(int pri)
1299{
1300 thread_t td = curthread;
1301
4b5f931b
MD
1302 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1303 crit_enter();
1304 if (td->td_flags & TDF_RUNQ) {
1305 _lwkt_dequeue(td);
1306 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1307 _lwkt_enqueue(td);
1308 } else {
1309 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1310 }
1311 crit_exit();
1312}
1313
5d21b981 1314/*
52eedfb5
MD
1315 * Migrate the current thread to the specified cpu.
1316 *
1317 * This is accomplished by descheduling ourselves from the current cpu,
1318 * moving our thread to the tdallq of the target cpu, IPI messaging the
1319 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1320 * races while the thread is being migrated.
ae8e83e6
MD
1321 *
1322 * We must be sure to remove ourselves from the current cpu's tsleepq
1323 * before potentially moving to another queue. The thread can be on
1324 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1325 */
3d28ff59 1326#ifdef SMP
5d21b981 1327static void lwkt_setcpu_remote(void *arg);
3d28ff59 1328#endif
5d21b981
MD
1329
1330void
1331lwkt_setcpu_self(globaldata_t rgd)
1332{
1333#ifdef SMP
1334 thread_t td = curthread;
1335
1336 if (td->td_gd != rgd) {
1337 crit_enter_quick(td);
ae8e83e6 1338 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1339 tsleep_remove(td);
5d21b981
MD
1340 td->td_flags |= TDF_MIGRATING;
1341 lwkt_deschedule_self(td);
52eedfb5 1342 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1343 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1344 lwkt_switch();
1345 /* we are now on the target cpu */
52eedfb5 1346 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1347 crit_exit_quick(td);
1348 }
1349#endif
1350}
1351
ecdefdda
MD
1352void
1353lwkt_migratecpu(int cpuid)
1354{
1355#ifdef SMP
1356 globaldata_t rgd;
1357
1358 rgd = globaldata_find(cpuid);
1359 lwkt_setcpu_self(rgd);
1360#endif
1361}
1362
5d21b981
MD
1363/*
1364 * Remote IPI for cpu migration (called while in a critical section so we
1365 * do not have to enter another one). The thread has already been moved to
1366 * our cpu's allq, but we must wait for the thread to be completely switched
1367 * out on the originating cpu before we schedule it on ours or the stack
1368 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1369 * change to main memory.
1370 *
1371 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1372 * against wakeups. It is best if this interface is used only when there
1373 * are no pending events that might try to schedule the thread.
1374 */
3d28ff59 1375#ifdef SMP
5d21b981
MD
1376static void
1377lwkt_setcpu_remote(void *arg)
1378{
1379 thread_t td = arg;
1380 globaldata_t gd = mycpu;
1381
df910c23
MD
1382 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1383#ifdef SMP
1384 lwkt_process_ipiq();
1385#endif
35238fa5 1386 cpu_lfence();
df910c23 1387 }
5d21b981 1388 td->td_gd = gd;
35238fa5 1389 cpu_sfence();
5d21b981 1390 td->td_flags &= ~TDF_MIGRATING;
9388413d 1391 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1392 _lwkt_enqueue(td);
1393}
3d28ff59 1394#endif
5d21b981 1395
553ea3c8 1396struct lwp *
4b5f931b
MD
1397lwkt_preempted_proc(void)
1398{
73e4f7b9 1399 thread_t td = curthread;
4b5f931b
MD
1400 while (td->td_preempted)
1401 td = td->td_preempted;
553ea3c8 1402 return(td->td_lwp);
4b5f931b
MD
1403}
1404
99df837e
MD
1405/*
1406 * Create a kernel process/thread/whatever. It shares it's address space
1407 * with proc0 - ie: kernel only.
1408 *
365fa13f
MD
1409 * NOTE! By default new threads are created with the MP lock held. A
1410 * thread which does not require the MP lock should release it by calling
1411 * rel_mplock() at the start of the new thread.
99df837e
MD
1412 */
1413int
1414lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1415 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1416 const char *fmt, ...)
99df837e 1417{
73e4f7b9 1418 thread_t td;
e2565a42 1419 __va_list ap;
99df837e 1420
d3d32139 1421 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1422 tdflags);
a2a5ad0d
MD
1423 if (tdp)
1424 *tdp = td;
709799ea 1425 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1426
1427 /*
1428 * Set up arg0 for 'ps' etc
1429 */
e2565a42 1430 __va_start(ap, fmt);
379210cb 1431 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1432 __va_end(ap);
99df837e
MD
1433
1434 /*
1435 * Schedule the thread to run
1436 */
ef0fdad1
MD
1437 if ((td->td_flags & TDF_STOPREQ) == 0)
1438 lwkt_schedule(td);
1439 else
1440 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1441 return 0;
1442}
1443
1444/*
1445 * Destroy an LWKT thread. Warning! This function is not called when
1446 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1447 * uses a different reaping mechanism.
1448 */
1449void
1450lwkt_exit(void)
1451{
1452 thread_t td = curthread;
c070746a 1453 thread_t std;
8826f33a 1454 globaldata_t gd;
99df837e
MD
1455
1456 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1457 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1458 caps_exit(td);
c070746a
MD
1459
1460 /*
1461 * Get us into a critical section to interlock gd_freetd and loop
1462 * until we can get it freed.
1463 *
1464 * We have to cache the current td in gd_freetd because objcache_put()ing
1465 * it would rip it out from under us while our thread is still active.
1466 */
1467 gd = mycpu;
37af14fe 1468 crit_enter_quick(td);
c070746a
MD
1469 while ((std = gd->gd_freetd) != NULL) {
1470 gd->gd_freetd = NULL;
1471 objcache_put(thread_cache, std);
1472 }
3b4192fb
MD
1473
1474 /*
1475 * Remove thread resources from kernel lists and deschedule us for
1476 * the last time.
1477 */
1478 if (td->td_flags & TDF_TSLEEPQ)
1479 tsleep_remove(td);
79eae878 1480 biosched_done(td);
f8abf63c 1481 dsched_exit_thread(td);
37af14fe 1482 lwkt_deschedule_self(td);
e56e4dea 1483 lwkt_remove_tdallq(td);
c070746a
MD
1484 if (td->td_flags & TDF_ALLOCATED_THREAD)
1485 gd->gd_freetd = td;
99df837e
MD
1486 cpu_thread_exit();
1487}
1488
e56e4dea
MD
1489void
1490lwkt_remove_tdallq(thread_t td)
1491{
1492 KKASSERT(td->td_gd == mycpu);
1493 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1494}
1495
2d93b37a
MD
1496void
1497crit_panic(void)
1498{
1499 thread_t td = curthread;
1500 int lpri = td->td_pri;
1501
1502 td->td_pri = 0;
1503 panic("td_pri is/would-go negative! %p %d", td, lpri);
1504}
1505
d165e668
MD
1506#ifdef SMP
1507
bd8015ca
MD
1508/*
1509 * Called from debugger/panic on cpus which have been stopped. We must still
1510 * process the IPIQ while stopped, even if we were stopped while in a critical
1511 * section (XXX).
1512 *
1513 * If we are dumping also try to process any pending interrupts. This may
1514 * or may not work depending on the state of the cpu at the point it was
1515 * stopped.
1516 */
1517void
1518lwkt_smp_stopped(void)
1519{
1520 globaldata_t gd = mycpu;
1521
1522 crit_enter_gd(gd);
1523 if (dumping) {
1524 lwkt_process_ipiq();
1525 splz();
1526 } else {
1527 lwkt_process_ipiq();
1528 }
1529 crit_exit_gd(gd);
1530}
1531
d165e668 1532#endif