LWKT - Rename lwkt_yield_quick() to a more appropriate splz_check().
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
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:
8c10bfcf 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.
20 *
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.
8c10bfcf 33 *
0aa16b5d 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/kernel.h>
47#include <sys/proc.h>
48#include <sys/rtprio.h>
49#include <sys/queue.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f6bf3af1 54#include <sys/caps.h>
9d265729 55#include <sys/spinlock.h>
57aa743c 56#include <sys/ktr.h>
9d265729
MD
57
58#include <sys/thread2.h>
59#include <sys/spinlock2.h>
f1d1c3fa 60
7d0bac62
MD
61#include <vm/vm.h>
62#include <vm/vm_param.h>
63#include <vm/vm_kern.h>
64#include <vm/vm_object.h>
65#include <vm/vm_page.h>
66#include <vm/vm_map.h>
67#include <vm/vm_pager.h>
68#include <vm/vm_extern.h>
7d0bac62 69
99df837e 70#include <machine/stdarg.h>
96728c05 71#include <machine/smp.h>
99df837e 72
1541028a 73
40aaf5fc
NT
74static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
75
3824f392
MD
76#ifdef SMP
77static int mplock_countx = 0;
78#endif
0f7a3396
MD
79#ifdef INVARIANTS
80static int panic_on_cscount = 0;
81#endif
05220613
MD
82static __int64_t switch_count = 0;
83static __int64_t preempt_hit = 0;
84static __int64_t preempt_miss = 0;
85static __int64_t preempt_weird = 0;
38717797
HP
86static __int64_t token_contention_count = 0;
87static __int64_t mplock_contention_count = 0;
fb0f29c4 88static int lwkt_use_spin_port;
d2f86ad2 89#ifdef SMP
b9eb1c19 90static int chain_mplock = 0;
d2f86ad2 91#endif
40aaf5fc 92static struct objcache *thread_cache;
05220613 93
b9eb1c19
MD
94volatile cpumask_t mp_lock_contention_mask;
95
0855a2af
JG
96extern void cpu_heavy_restore(void);
97extern void cpu_lwkt_restore(void);
98extern void cpu_kthread_restore(void);
99extern void cpu_idle_restore(void);
100
85514115
MD
101#ifdef __amd64__
102
103static int
0855a2af
JG
104jg_tos_ok(struct thread *td)
105{
85514115
MD
106 void *tos;
107 int tos_ok;
108
0855a2af
JG
109 if (td == NULL) {
110 return 1;
111 }
112 KKASSERT(td->td_sp != NULL);
85514115
MD
113 tos = ((void **)td->td_sp)[0];
114 tos_ok = 0;
115 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
116 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
117 tos_ok = 1;
118 }
119 return tos_ok;
120}
121
85514115
MD
122#endif
123
fb0f29c4
MD
124/*
125 * We can make all thread ports use the spin backend instead of the thread
126 * backend. This should only be set to debug the spin backend.
127 */
128TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
129
0f7a3396
MD
130#ifdef INVARIANTS
131SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
132#endif
b9eb1c19
MD
133#ifdef SMP
134SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
135#endif
4b5f931b 136SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 137SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 138SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 139SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
140#ifdef INVARIANTS
141SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
142 &token_contention_count, 0, "spinning due to token contention");
143SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
144 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
145#endif
05220613 146
57aa743c
MD
147/*
148 * Kernel Trace
149 */
57aa743c
MD
150#if !defined(KTR_GIANT_CONTENTION)
151#define KTR_GIANT_CONTENTION KTR_ALL
152#endif
153
154KTR_INFO_MASTER(giant);
155KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
156KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
157
158#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
159
4b5f931b
MD
160/*
161 * These helper procedures handle the runq, they can only be called from
162 * within a critical section.
75cdbe6c
MD
163 *
164 * WARNING! Prior to SMP being brought up it is possible to enqueue and
165 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
166 * instead of 'mycpu' when referencing the globaldata structure. Once
167 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 168 */
f1d1c3fa
MD
169static __inline
170void
171_lwkt_dequeue(thread_t td)
172{
173 if (td->td_flags & TDF_RUNQ) {
4b5f931b 174 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 175 struct globaldata *gd = td->td_gd;
4b5f931b 176
f1d1c3fa 177 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
178 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
179 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
180 }
181}
182
183static __inline
184void
185_lwkt_enqueue(thread_t td)
186{
344ad853 187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 188 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 189 struct globaldata *gd = td->td_gd;
4b5f931b 190
f1d1c3fa 191 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
192 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
193 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
194 }
195}
8ad65e08 196
40aaf5fc
NT
197static __boolean_t
198_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
199{
200 struct thread *td = (struct thread *)obj;
201
202 td->td_kstack = NULL;
203 td->td_kstack_size = 0;
204 td->td_flags = TDF_ALLOCATED_THREAD;
205 return (1);
206}
207
208static void
209_lwkt_thread_dtor(void *obj, void *privdata)
210{
211 struct thread *td = (struct thread *)obj;
212
213 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
214 ("_lwkt_thread_dtor: not allocated from objcache"));
215 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
216 td->td_kstack_size > 0,
217 ("_lwkt_thread_dtor: corrupted stack"));
218 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
219}
220
221/*
222 * Initialize the lwkt s/system.
223 */
224void
225lwkt_init(void)
226{
227 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
228 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
229 NULL, CACHE_NTHREADS/2,
230 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
231}
232
37af14fe
MD
233/*
234 * Schedule a thread to run. As the current thread we can always safely
235 * schedule ourselves, and a shortcut procedure is provided for that
236 * function.
237 *
238 * (non-blocking, self contained on a per cpu basis)
239 */
240void
241lwkt_schedule_self(thread_t td)
242{
243 crit_enter_quick(td);
37af14fe 244 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 245 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 246 _lwkt_enqueue(td);
37af14fe
MD
247 crit_exit_quick(td);
248}
249
250/*
251 * Deschedule a thread.
252 *
253 * (non-blocking, self contained on a per cpu basis)
254 */
255void
256lwkt_deschedule_self(thread_t td)
257{
258 crit_enter_quick(td);
37af14fe
MD
259 _lwkt_dequeue(td);
260 crit_exit_quick(td);
261}
262
8ad65e08
MD
263/*
264 * LWKTs operate on a per-cpu basis
265 *
73e4f7b9 266 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
267 */
268void
269lwkt_gdinit(struct globaldata *gd)
270{
4b5f931b
MD
271 int i;
272
273 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
274 TAILQ_INIT(&gd->gd_tdrunq[i]);
275 gd->gd_runqmask = 0;
73e4f7b9 276 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
277}
278
7d0bac62
MD
279/*
280 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
281 * or LWKT start address before it can be scheduled. If the target cpu is
282 * -1 the thread will be created on the current cpu.
0cfcada1
MD
283 *
284 * If you intend to create a thread without a process context this function
285 * does everything except load the startup and switcher function.
7d0bac62
MD
286 */
287thread_t
d3d32139 288lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 289{
c070746a 290 globaldata_t gd = mycpu;
99df837e 291 void *stack;
7d0bac62 292
c070746a
MD
293 /*
294 * If static thread storage is not supplied allocate a thread. Reuse
295 * a cached free thread if possible. gd_freetd is used to keep an exiting
296 * thread intact through the exit.
297 */
ef0fdad1 298 if (td == NULL) {
c070746a
MD
299 if ((td = gd->gd_freetd) != NULL)
300 gd->gd_freetd = NULL;
301 else
302 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
303 KASSERT((td->td_flags &
304 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
305 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
306 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 307 }
c070746a
MD
308
309 /*
310 * Try to reuse cached stack.
311 */
f470d0c8
MD
312 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
313 if (flags & TDF_ALLOCATED_STACK) {
e4846942 314 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
315 stack = NULL;
316 }
317 }
318 if (stack == NULL) {
e4846942 319 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 320 flags |= TDF_ALLOCATED_STACK;
99df837e 321 }
75cdbe6c 322 if (cpu < 0)
c070746a 323 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 324 else
f470d0c8 325 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 326 return(td);
7d0bac62
MD
327}
328
329/*
330 * Initialize a preexisting thread structure. This function is used by
331 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
332 *
f8c3996b
MD
333 * All threads start out in a critical section at a priority of
334 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
335 * appropriate. This function may send an IPI message when the
336 * requested cpu is not the current cpu and consequently gd_tdallq may
337 * not be initialized synchronously from the point of view of the originating
338 * cpu.
339 *
340 * NOTE! we have to be careful in regards to creating threads for other cpus
341 * if SMP has not yet been activated.
7d0bac62 342 */
41a01a4d
MD
343#ifdef SMP
344
75cdbe6c
MD
345static void
346lwkt_init_thread_remote(void *arg)
347{
348 thread_t td = arg;
349
52eedfb5
MD
350 /*
351 * Protected by critical section held by IPI dispatch
352 */
75cdbe6c
MD
353 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
354}
355
41a01a4d
MD
356#endif
357
7d0bac62 358void
f470d0c8
MD
359lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
360 struct globaldata *gd)
7d0bac62 361{
37af14fe
MD
362 globaldata_t mygd = mycpu;
363
99df837e
MD
364 bzero(td, sizeof(struct thread));
365 td->td_kstack = stack;
f470d0c8 366 td->td_kstack_size = stksize;
d3d32139 367 td->td_flags = flags;
26a0694b 368 td->td_gd = gd;
f8c3996b 369 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
370#ifdef SMP
371 if ((flags & TDF_MPSAFE) == 0)
372 td->td_mpcount = 1;
373#endif
fb0f29c4
MD
374 if (lwkt_use_spin_port)
375 lwkt_initport_spin(&td->td_msgport);
376 else
377 lwkt_initport_thread(&td->td_msgport, td);
99df837e 378 pmap_init_thread(td);
0f7a3396 379#ifdef SMP
5d21b981
MD
380 /*
381 * Normally initializing a thread for a remote cpu requires sending an
382 * IPI. However, the idlethread is setup before the other cpus are
383 * activated so we have to treat it as a special case. XXX manipulation
384 * of gd_tdallq requires the BGL.
385 */
386 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 387 crit_enter_gd(mygd);
75cdbe6c 388 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 389 crit_exit_gd(mygd);
75cdbe6c 390 } else {
2db3b277 391 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 392 }
0f7a3396 393#else
37af14fe 394 crit_enter_gd(mygd);
0f7a3396 395 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 396 crit_exit_gd(mygd);
0f7a3396 397#endif
73e4f7b9
MD
398}
399
400void
401lwkt_set_comm(thread_t td, const char *ctl, ...)
402{
e2565a42 403 __va_list va;
73e4f7b9 404
e2565a42 405 __va_start(va, ctl);
379210cb 406 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 407 __va_end(va);
7d0bac62
MD
408}
409
99df837e 410void
73e4f7b9 411lwkt_hold(thread_t td)
99df837e 412{
73e4f7b9
MD
413 ++td->td_refs;
414}
415
416void
417lwkt_rele(thread_t td)
418{
419 KKASSERT(td->td_refs > 0);
420 --td->td_refs;
421}
422
423void
424lwkt_wait_free(thread_t td)
425{
426 while (td->td_refs)
377d4740 427 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
428}
429
430void
431lwkt_free_thread(thread_t td)
432{
d9eea1a5 433 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
434 ("lwkt_free_thread: did not exit! %p", td));
435
40aaf5fc
NT
436 if (td->td_flags & TDF_ALLOCATED_THREAD) {
437 objcache_put(thread_cache, td);
438 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
439 /* client-allocated struct with internally allocated stack */
440 KASSERT(td->td_kstack && td->td_kstack_size > 0,
441 ("lwkt_free_thread: corrupted stack"));
442 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
443 td->td_kstack = NULL;
444 td->td_kstack_size = 0;
99df837e
MD
445 }
446}
447
448
8ad65e08
MD
449/*
450 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
451 * switch to the idlethread. Switching must occur within a critical
452 * section to avoid races with the scheduling queue.
453 *
454 * We always have full control over our cpu's run queue. Other cpus
455 * that wish to manipulate our queue must use the cpu_*msg() calls to
456 * talk to our cpu, so a critical section is all that is needed and
457 * the result is very, very fast thread switching.
458 *
96728c05
MD
459 * The LWKT scheduler uses a fixed priority model and round-robins at
460 * each priority level. User process scheduling is a totally
461 * different beast and LWKT priorities should not be confused with
462 * user process priorities.
f1d1c3fa 463 *
96728c05
MD
464 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
465 * cleans it up. Note that the td_switch() function cannot do anything that
466 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
467 * the target thread (not the current thread). It's nice to have a scheduler
468 * that does not need the MP lock to work because it allows us to do some
469 * really cool high-performance MP lock optimizations.
69d78e99
MD
470 *
471 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
472 * is not called by the current thread in the preemption case, only when
473 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
474 */
475void
476lwkt_switch(void)
477{
37af14fe
MD
478 globaldata_t gd = mycpu;
479 thread_t td = gd->gd_curthread;
8ad65e08 480 thread_t ntd;
8a8d5d85
MD
481#ifdef SMP
482 int mpheld;
483#endif
8ad65e08 484
46a3f46d 485 /*
27e88a6e
MD
486 * Switching from within a 'fast' (non thread switched) interrupt or IPI
487 * is illegal. However, we may have to do it anyway if we hit a fatal
488 * kernel trap or we have paniced.
489 *
490 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 491 */
27e88a6e
MD
492 if (gd->gd_intr_nesting_level) {
493 int savegdnest;
494 int savegdtrap;
495
496 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
497 panic("lwkt_switch: cannot switch from within "
498 "a fast interrupt, yet, td %p\n", td);
499 } else {
500 savegdnest = gd->gd_intr_nesting_level;
501 savegdtrap = gd->gd_trap_nesting_level;
502 gd->gd_intr_nesting_level = 0;
503 gd->gd_trap_nesting_level = 0;
a7422615
MD
504 if ((td->td_flags & TDF_PANICWARN) == 0) {
505 td->td_flags |= TDF_PANICWARN;
6ea70f76 506 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 507 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 508 print_backtrace();
a7422615 509 }
27e88a6e
MD
510 lwkt_switch();
511 gd->gd_intr_nesting_level = savegdnest;
512 gd->gd_trap_nesting_level = savegdtrap;
513 return;
514 }
96728c05 515 }
ef0fdad1 516
cb973d15
MD
517 /*
518 * Passive release (used to transition from user to kernel mode
519 * when we block or switch rather then when we enter the kernel).
520 * This function is NOT called if we are switching into a preemption
521 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
522 * our current process designation (if we have one) and become a true
523 * LWKT thread, and may also hand the current process designation to
524 * another process and schedule thread.
cb973d15
MD
525 */
526 if (td->td_release)
527 td->td_release(td);
528
37af14fe 529 crit_enter_gd(gd);
9d265729
MD
530 if (td->td_toks)
531 lwkt_relalltokens(td);
532
533 /*
b02926de
MD
534 * We had better not be holding any spin locks, but don't get into an
535 * endless panic loop.
9d265729 536 */
bbb31c5d
MD
537 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
538 ("lwkt_switch: still holding a shared spinlock %p!",
539 gd->gd_spinlock_rd));
d666840a
MD
540 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
541 ("lwkt_switch: still holding %d exclusive spinlocks!",
542 gd->gd_spinlocks_wr));
9d265729 543
8a8d5d85
MD
544
545#ifdef SMP
546 /*
547 * td_mpcount cannot be used to determine if we currently hold the
548 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
549 * to get the lock, and switch out if it can't. Our ownership of
550 * the actual lock will remain stable while we are in a critical section
551 * (but, of course, another cpu may own or release the lock so the
552 * actual value of mp_lock is not stable).
8a8d5d85
MD
553 */
554 mpheld = MP_LOCK_HELD();
0f7a3396
MD
555#ifdef INVARIANTS
556 if (td->td_cscount) {
6ea70f76 557 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
558 td);
559 if (panic_on_cscount)
560 panic("switching while mastering cpusync");
561 }
562#endif
8a8d5d85 563#endif
99df837e
MD
564 if ((ntd = td->td_preempted) != NULL) {
565 /*
566 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
567 * thread. This occurs transparently, whether the preempted thread
568 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
569 * itself).
570 *
571 * We have to setup the MP lock for the original thread after backing
572 * out the adjustment that was made to curthread when the original
573 * was preempted.
99df837e 574 */
26a0694b 575 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 576#ifdef SMP
96728c05 577 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 578 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
579 td, ntd, td->td_mpcount, ntd->td_mpcount);
580 }
8a8d5d85
MD
581 if (ntd->td_mpcount) {
582 td->td_mpcount -= ntd->td_mpcount;
583 KKASSERT(td->td_mpcount >= 0);
584 }
585#endif
26a0694b 586 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
587
588 /*
b9eb1c19
MD
589 * The interrupt may have woken a thread up, we need to properly
590 * set the reschedule flag if the originally interrupted thread is
591 * at a lower priority.
8ec60c3f
MD
592 */
593 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
594 need_lwkt_resched();
8a8d5d85 595 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 596 } else {
4b5f931b
MD
597 /*
598 * Priority queue / round-robin at each priority. Note that user
599 * processes run at a fixed, low priority and the user process
600 * scheduler deals with interactions between user processes
601 * by scheduling and descheduling them from the LWKT queue as
602 * necessary.
8a8d5d85
MD
603 *
604 * We have to adjust the MP lock for the target thread. If we
605 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
606 * thread that does not need the MP lock. If we cannot, we spin
607 * instead of HLT.
608 *
609 * A similar issue exists for the tokens held by the target thread.
610 * If we cannot obtain ownership of the tokens we cannot immediately
611 * schedule the thread.
612 */
613
8ec60c3f
MD
614 /*
615 * If an LWKT reschedule was requested, well that is what we are
616 * doing now so clear it.
617 */
618 clear_lwkt_resched();
4b5f931b
MD
619again:
620 if (gd->gd_runqmask) {
621 int nq = bsrl(gd->gd_runqmask);
622 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
623 gd->gd_runqmask &= ~(1 << nq);
624 goto again;
625 }
8a8d5d85 626#ifdef SMP
41a01a4d 627 /*
df6b8ba0
MD
628 * THREAD SELECTION FOR AN SMP MACHINE BUILD
629 *
41a01a4d
MD
630 * If the target needs the MP lock and we couldn't get it,
631 * or if the target is holding tokens and we could not
632 * gain ownership of the tokens, continue looking for a
633 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
634 *
635 * NOTE: the mpheld variable invalid after this conditional, it
636 * can change due to both cpu_try_mplock() returning success
9d265729 637 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
638 * we are trying to check the mpcount of a thread other then
639 * the current thread. Because of this, if the current thread
640 * is not holding td_mpcount, an IPI indirectly run via
9d265729 641 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 642 * cause the core MP lock to be released.
41a01a4d
MD
643 */
644 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 645 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 646 ) {
8a8d5d85 647 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
648
649 mpheld = MP_LOCK_HELD();
650 ntd = NULL;
8a8d5d85
MD
651 while (rqmask) {
652 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 653 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 654 /* spinning due to MP lock being held */
38717797 655#ifdef INVARIANTS
a453459d 656 ++mplock_contention_count;
38717797 657#endif
a453459d 658 /* mplock still not held, 'mpheld' still valid */
41a01a4d 659 continue;
38717797 660 }
a453459d
MD
661
662 /*
9d265729 663 * mpheld state invalid after getalltokens call returns
a453459d
MD
664 * failure, but the variable is only needed for
665 * the loop.
666 */
9d265729 667 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 668 /* spinning due to token contention */
38717797 669#ifdef INVARIANTS
a453459d 670 ++token_contention_count;
38717797 671#endif
a453459d 672 mpheld = MP_LOCK_HELD();
41a01a4d 673 continue;
38717797 674 }
41a01a4d 675 break;
8a8d5d85
MD
676 }
677 if (ntd)
678 break;
679 rqmask &= ~(1 << nq);
680 nq = bsrl(rqmask);
b9eb1c19
MD
681
682 /*
683 * We have two choices. We can either refuse to run a
684 * user thread when a kernel thread needs the MP lock
685 * but could not get it, or we can allow it to run but
686 * then expect an IPI (hopefully) later on to force a
687 * reschedule when the MP lock might become available.
688 */
689 if (nq < TDPRI_KERN_LPSCHED) {
690 if (chain_mplock == 0)
691 break;
692 atomic_set_int(&mp_lock_contention_mask,
693 gd->gd_cpumask);
694 /* continue loop, allow user threads to be scheduled */
695 }
8a8d5d85
MD
696 }
697 if (ntd == NULL) {
b402c633 698 cpu_mplock_contested();
a2a5ad0d
MD
699 ntd = &gd->gd_idlethread;
700 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 701 goto using_idle_thread;
8a8d5d85 702 } else {
344ad853 703 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
704 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
706 }
707 } else {
3824f392
MD
708 if (ntd->td_mpcount)
709 ++mplock_countx;
344ad853 710 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
711 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
713 }
714#else
df6b8ba0
MD
715 /*
716 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
717 * worry about tokens or the BGL. However, we still have
718 * to call lwkt_getalltokens() in order to properly detect
719 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 720 */
7eb611ef 721 lwkt_getalltokens(ntd);
344ad853 722 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
723 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
724 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 725#endif
4b5f931b 726 } else {
3c23a41a 727 /*
60f945af
MD
728 * We have nothing to run but only let the idle loop halt
729 * the cpu if there are no pending interrupts.
3c23a41a 730 */
a2a5ad0d 731 ntd = &gd->gd_idlethread;
60f945af 732 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 733 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 734#ifdef SMP
df6b8ba0
MD
735using_idle_thread:
736 /*
737 * The idle thread should not be holding the MP lock unless we
738 * are trapping in the kernel or in a panic. Since we select the
739 * idle thread unconditionally when no other thread is available,
740 * if the MP lock is desired during a panic or kernel trap, we
741 * have to loop in the scheduler until we get it.
742 */
743 if (ntd->td_mpcount) {
744 mpheld = MP_LOCK_HELD();
b402c633 745 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 746 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
747 } else if (mpheld == 0) {
748 cpu_mplock_contested();
df6b8ba0 749 goto again;
b402c633 750 }
df6b8ba0 751 }
a453459d 752#endif
4b5f931b 753 }
f1d1c3fa 754 }
26a0694b
MD
755 KASSERT(ntd->td_pri >= TDPRI_CRIT,
756 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
757
758 /*
759 * Do the actual switch. If the new target does not need the MP lock
760 * and we are holding it, release the MP lock. If the new target requires
761 * the MP lock we have already acquired it for the target.
762 */
763#ifdef SMP
764 if (ntd->td_mpcount == 0 ) {
765 if (MP_LOCK_HELD())
766 cpu_rel_mplock();
767 } else {
a453459d 768 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
769 }
770#endif
94f6d86e
MD
771 if (td != ntd) {
772 ++switch_count;
85514115 773#ifdef __amd64__
0855a2af 774 KKASSERT(jg_tos_ok(ntd));
85514115 775#endif
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;
917 td->td_switch(ntd);
b9eb1c19 918
26a0694b 919 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
920#ifdef SMP
921 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
922 mpheld = MP_LOCK_HELD();
923 if (mpheld && td->td_mpcount == 0)
96728c05 924 cpu_rel_mplock();
71ef2f5c 925 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
926 panic("lwkt_preempt(): MP lock was not held through");
927#endif
26a0694b
MD
928 ntd->td_preempted = NULL;
929 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
930}
931
f1d1c3fa 932/*
faaeffac 933 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 934 *
faaeffac
MD
935 * td_nest_count prevents deep nesting via splz() or doreti() which
936 * might otherwise blow out the kernel stack. Note that except for
937 * this special case, we MUST call splz() here to handle any
938 * pending ints, particularly after we switch, or we might accidently
939 * halt the cpu with interrupts pending.
ef0fdad1 940 *
f1d1c3fa
MD
941 * (self contained on a per cpu basis)
942 */
943void
faaeffac 944splz_check(void)
f1d1c3fa 945{
7966cb69
MD
946 globaldata_t gd = mycpu;
947 thread_t td = gd->gd_curthread;
ef0fdad1 948
46a3f46d 949 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 950 splz();
f1d1c3fa
MD
951}
952
8ad65e08 953/*
faaeffac
MD
954 * This implements a normal yield which will yield to equal priority
955 * threads as well as higher priority threads. Note that gd_reqflags
956 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
957 *
958 * (self contained on a per cpu basis)
8ad65e08
MD
959 */
960void
f1d1c3fa 961lwkt_yield(void)
8ad65e08 962{
37af14fe 963 lwkt_schedule_self(curthread);
f1d1c3fa
MD
964 lwkt_switch();
965}
966
3824f392
MD
967/*
968 * This function is used along with the lwkt_passive_recover() inline
969 * by the trap code to negotiate a passive release of the current
970 * process/lwp designation with the user scheduler.
971 */
972void
973lwkt_passive_release(struct thread *td)
974{
975 struct lwp *lp = td->td_lwp;
976
977 td->td_release = NULL;
978 lwkt_setpri_self(TDPRI_KERN_USER);
979 lp->lwp_proc->p_usched->release_curproc(lp);
980}
981
982/*
983 * Make a kernel thread act as if it were in user mode with regards
984 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
985 * loops which may be potentially cpu-bound can call lwkt_user_yield().
986 *
987 * The lwkt_user_yield() function is designed to have very low overhead
988 * if no yield is determined to be needed.
989 */
990void
991lwkt_user_yield(void)
992{
993 thread_t td = curthread;
994 struct lwp *lp = td->td_lwp;
995
996#ifdef SMP
997 /*
998 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
999 * kernel can prevent other cpus from servicing interrupt threads
1000 * which still require the MP lock (which is a lot of them). This
1001 * has a chaining effect since if the interrupt is blocked, so is
1002 * the event, so normal scheduling will not pick up on the problem.
1003 */
1004 if (mplock_countx && td->td_mpcount) {
1005 int savecnt = td->td_mpcount;
1006
1007 td->td_mpcount = 1;
1008 rel_mplock();
1009 DELAY(10);
1010 get_mplock();
1011 td->td_mpcount = savecnt;
1012 mplock_countx = 0;
1013 }
1014#endif
1015
1016 /*
1017 * Another kernel thread wants the cpu
1018 */
1019 if (lwkt_resched_wanted())
1020 lwkt_switch();
1021
1022 /*
1023 * If the user scheduler has asynchronously determined that the current
1024 * process (when running in user mode) needs to lose the cpu then make
1025 * sure we are released.
1026 */
1027 if (user_resched_wanted()) {
1028 if (td->td_release)
1029 td->td_release(td);
1030 }
1031
1032 /*
1033 * If we are released reduce our priority
1034 */
1035 if (td->td_release == NULL) {
1036 if (lwkt_check_resched(td) > 0)
1037 lwkt_switch();
1038 lp->lwp_proc->p_usched->acquire_curproc(lp);
1039 td->td_release = lwkt_passive_release;
1040 lwkt_setpri_self(TDPRI_USER_NORM);
1041 }
1042}
1043
b9eb1c19
MD
1044/*
1045 * Return 0 if no runnable threads are pending at the same or higher
1046 * priority as the passed thread.
1047 *
1048 * Return 1 if runnable threads are pending at the same priority.
1049 *
1050 * Return 2 if runnable threads are pending at a higher priority.
1051 */
1052int
1053lwkt_check_resched(thread_t td)
1054{
1055 int pri = td->td_pri & TDPRI_MASK;
1056
1057 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1058 return(2);
1059 if (TAILQ_NEXT(td, td_threadq))
1060 return(1);
1061 return(0);
1062}
1063
8ad65e08 1064/*
f1d1c3fa
MD
1065 * Generic schedule. Possibly schedule threads belonging to other cpus and
1066 * deal with threads that might be blocked on a wait queue.
1067 *
0a3f9b47
MD
1068 * We have a little helper inline function which does additional work after
1069 * the thread has been enqueued, including dealing with preemption and
1070 * setting need_lwkt_resched() (which prevents the kernel from returning
1071 * to userland until it has processed higher priority threads).
6330a558
MD
1072 *
1073 * It is possible for this routine to be called after a failed _enqueue
1074 * (due to the target thread migrating, sleeping, or otherwise blocked).
1075 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1076 *
1077 * reschedok is an optimized constant propagated from lwkt_schedule() or
1078 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1079 * reschedule to be requested if the target thread has a higher priority.
1080 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1081 * be 0, prevented undesired reschedules.
8ad65e08 1082 */
0a3f9b47
MD
1083static __inline
1084void
361d01dd 1085_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1086{
b9eb1c19 1087 thread_t otd;
c730be20 1088
6330a558 1089 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1090 if (ntd->td_preemptable && reschedok) {
6330a558 1091 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1092 } else if (reschedok) {
b9eb1c19
MD
1093 otd = curthread;
1094 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1095 need_lwkt_resched();
6330a558 1096 }
0a3f9b47
MD
1097 }
1098}
1099
361d01dd 1100static __inline
8ad65e08 1101void
361d01dd 1102_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1103{
37af14fe
MD
1104 globaldata_t mygd = mycpu;
1105
41a01a4d 1106 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1107 crit_enter_gd(mygd);
9388413d 1108 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1109 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1110 _lwkt_enqueue(td);
1111 } else {
f1d1c3fa 1112 /*
7cd8d145
MD
1113 * If we own the thread, there is no race (since we are in a
1114 * critical section). If we do not own the thread there might
1115 * be a race but the target cpu will deal with it.
f1d1c3fa 1116 */
0f7a3396 1117#ifdef SMP
7cd8d145 1118 if (td->td_gd == mygd) {
9d265729 1119 _lwkt_enqueue(td);
361d01dd 1120 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1121 } else {
7cd8d145
MD
1122 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1123 }
0f7a3396 1124#else
7cd8d145 1125 _lwkt_enqueue(td);
361d01dd 1126 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1127#endif
8ad65e08 1128 }
37af14fe 1129 crit_exit_gd(mygd);
8ad65e08
MD
1130}
1131
361d01dd
MD
1132void
1133lwkt_schedule(thread_t td)
1134{
1135 _lwkt_schedule(td, 1);
1136}
1137
1138void
1139lwkt_schedule_noresched(thread_t td)
1140{
1141 _lwkt_schedule(td, 0);
1142}
1143
52eedfb5
MD
1144#ifdef SMP
1145
d9eea1a5 1146/*
52eedfb5
MD
1147 * Thread migration using a 'Pull' method. The thread may or may not be
1148 * the current thread. It MUST be descheduled and in a stable state.
1149 * lwkt_giveaway() must be called on the cpu owning the thread.
1150 *
1151 * At any point after lwkt_giveaway() is called, the target cpu may
1152 * 'pull' the thread by calling lwkt_acquire().
1153 *
ae8e83e6
MD
1154 * We have to make sure the thread is not sitting on a per-cpu tsleep
1155 * queue or it will blow up when it moves to another cpu.
1156 *
52eedfb5 1157 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1158 */
52eedfb5
MD
1159void
1160lwkt_giveaway(thread_t td)
1161{
1162 globaldata_t gd = mycpu;
1163
1164 crit_enter_gd(gd);
ae8e83e6
MD
1165 if (td->td_flags & TDF_TSLEEPQ)
1166 tsleep_remove(td);
52eedfb5
MD
1167 KKASSERT(td->td_gd == gd);
1168 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1169 td->td_flags |= TDF_MIGRATING;
1170 crit_exit_gd(gd);
1171}
1172
a2a5ad0d
MD
1173void
1174lwkt_acquire(thread_t td)
1175{
37af14fe
MD
1176 globaldata_t gd;
1177 globaldata_t mygd;
a2a5ad0d 1178
52eedfb5 1179 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1180 gd = td->td_gd;
37af14fe 1181 mygd = mycpu;
52eedfb5 1182 if (gd != mycpu) {
35238fa5 1183 cpu_lfence();
52eedfb5 1184 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1185 crit_enter_gd(mygd);
df910c23
MD
1186 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1187#ifdef SMP
1188 lwkt_process_ipiq();
1189#endif
52eedfb5 1190 cpu_lfence();
df910c23 1191 }
37af14fe 1192 td->td_gd = mygd;
52eedfb5
MD
1193 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1194 td->td_flags &= ~TDF_MIGRATING;
1195 crit_exit_gd(mygd);
1196 } else {
1197 crit_enter_gd(mygd);
1198 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1199 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1200 crit_exit_gd(mygd);
a2a5ad0d
MD
1201 }
1202}
1203
52eedfb5
MD
1204#endif
1205
f1d1c3fa
MD
1206/*
1207 * Generic deschedule. Descheduling threads other then your own should be
1208 * done only in carefully controlled circumstances. Descheduling is
1209 * asynchronous.
1210 *
1211 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1212 */
1213void
1214lwkt_deschedule(thread_t td)
1215{
f1d1c3fa 1216 crit_enter();
b8a98473 1217#ifdef SMP
f1d1c3fa
MD
1218 if (td == curthread) {
1219 _lwkt_dequeue(td);
1220 } else {
a72187e9 1221 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1222 _lwkt_dequeue(td);
1223 } else {
b8a98473 1224 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1225 }
1226 }
b8a98473
MD
1227#else
1228 _lwkt_dequeue(td);
1229#endif
f1d1c3fa
MD
1230 crit_exit();
1231}
1232
4b5f931b
MD
1233/*
1234 * Set the target thread's priority. This routine does not automatically
1235 * switch to a higher priority thread, LWKT threads are not designed for
1236 * continuous priority changes. Yield if you want to switch.
1237 *
1238 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1239 * of the td_pri field. The specified priority may also indicate zero or
1240 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1241 *
1242 * Note that we requeue the thread whether it winds up on a different runq
1243 * or not. uio_yield() depends on this and the routine is not normally
1244 * called with the same priority otherwise.
4b5f931b
MD
1245 */
1246void
1247lwkt_setpri(thread_t td, int pri)
1248{
26a0694b 1249 KKASSERT(pri >= 0);
a72187e9 1250 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1251 crit_enter();
1252 if (td->td_flags & TDF_RUNQ) {
1253 _lwkt_dequeue(td);
1254 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1255 _lwkt_enqueue(td);
1256 } else {
1257 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1258 }
1259 crit_exit();
1260}
1261
1262void
1263lwkt_setpri_self(int pri)
1264{
1265 thread_t td = curthread;
1266
4b5f931b
MD
1267 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1268 crit_enter();
1269 if (td->td_flags & TDF_RUNQ) {
1270 _lwkt_dequeue(td);
1271 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1272 _lwkt_enqueue(td);
1273 } else {
1274 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1275 }
1276 crit_exit();
1277}
1278
5d21b981 1279/*
52eedfb5
MD
1280 * Migrate the current thread to the specified cpu.
1281 *
1282 * This is accomplished by descheduling ourselves from the current cpu,
1283 * moving our thread to the tdallq of the target cpu, IPI messaging the
1284 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1285 * races while the thread is being migrated.
ae8e83e6
MD
1286 *
1287 * We must be sure to remove ourselves from the current cpu's tsleepq
1288 * before potentially moving to another queue. The thread can be on
1289 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1290 */
3d28ff59 1291#ifdef SMP
5d21b981 1292static void lwkt_setcpu_remote(void *arg);
3d28ff59 1293#endif
5d21b981
MD
1294
1295void
1296lwkt_setcpu_self(globaldata_t rgd)
1297{
1298#ifdef SMP
1299 thread_t td = curthread;
1300
1301 if (td->td_gd != rgd) {
1302 crit_enter_quick(td);
ae8e83e6
MD
1303 if (td->td_flags & TDF_TSLEEPQ)
1304 tsleep_remove(td);
5d21b981
MD
1305 td->td_flags |= TDF_MIGRATING;
1306 lwkt_deschedule_self(td);
52eedfb5 1307 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1308 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1309 lwkt_switch();
1310 /* we are now on the target cpu */
52eedfb5 1311 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1312 crit_exit_quick(td);
1313 }
1314#endif
1315}
1316
ecdefdda
MD
1317void
1318lwkt_migratecpu(int cpuid)
1319{
1320#ifdef SMP
1321 globaldata_t rgd;
1322
1323 rgd = globaldata_find(cpuid);
1324 lwkt_setcpu_self(rgd);
1325#endif
1326}
1327
5d21b981
MD
1328/*
1329 * Remote IPI for cpu migration (called while in a critical section so we
1330 * do not have to enter another one). The thread has already been moved to
1331 * our cpu's allq, but we must wait for the thread to be completely switched
1332 * out on the originating cpu before we schedule it on ours or the stack
1333 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1334 * change to main memory.
1335 *
1336 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1337 * against wakeups. It is best if this interface is used only when there
1338 * are no pending events that might try to schedule the thread.
1339 */
3d28ff59 1340#ifdef SMP
5d21b981
MD
1341static void
1342lwkt_setcpu_remote(void *arg)
1343{
1344 thread_t td = arg;
1345 globaldata_t gd = mycpu;
1346
df910c23
MD
1347 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1348#ifdef SMP
1349 lwkt_process_ipiq();
1350#endif
35238fa5 1351 cpu_lfence();
df910c23 1352 }
5d21b981 1353 td->td_gd = gd;
35238fa5 1354 cpu_sfence();
5d21b981 1355 td->td_flags &= ~TDF_MIGRATING;
9388413d 1356 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1357 _lwkt_enqueue(td);
1358}
3d28ff59 1359#endif
5d21b981 1360
553ea3c8 1361struct lwp *
4b5f931b
MD
1362lwkt_preempted_proc(void)
1363{
73e4f7b9 1364 thread_t td = curthread;
4b5f931b
MD
1365 while (td->td_preempted)
1366 td = td->td_preempted;
553ea3c8 1367 return(td->td_lwp);
4b5f931b
MD
1368}
1369
99df837e
MD
1370/*
1371 * Create a kernel process/thread/whatever. It shares it's address space
1372 * with proc0 - ie: kernel only.
1373 *
365fa13f
MD
1374 * NOTE! By default new threads are created with the MP lock held. A
1375 * thread which does not require the MP lock should release it by calling
1376 * rel_mplock() at the start of the new thread.
99df837e
MD
1377 */
1378int
1379lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1380 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1381 const char *fmt, ...)
99df837e 1382{
73e4f7b9 1383 thread_t td;
e2565a42 1384 __va_list ap;
99df837e 1385
d3d32139 1386 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1387 tdflags);
a2a5ad0d
MD
1388 if (tdp)
1389 *tdp = td;
709799ea 1390 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1391
1392 /*
1393 * Set up arg0 for 'ps' etc
1394 */
e2565a42 1395 __va_start(ap, fmt);
379210cb 1396 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1397 __va_end(ap);
99df837e
MD
1398
1399 /*
1400 * Schedule the thread to run
1401 */
ef0fdad1
MD
1402 if ((td->td_flags & TDF_STOPREQ) == 0)
1403 lwkt_schedule(td);
1404 else
1405 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1406 return 0;
1407}
1408
1409/*
1410 * Destroy an LWKT thread. Warning! This function is not called when
1411 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1412 * uses a different reaping mechanism.
1413 */
1414void
1415lwkt_exit(void)
1416{
1417 thread_t td = curthread;
c070746a 1418 thread_t std;
8826f33a 1419 globaldata_t gd;
99df837e
MD
1420
1421 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1422 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1423 caps_exit(td);
c070746a
MD
1424
1425 /*
1426 * Get us into a critical section to interlock gd_freetd and loop
1427 * until we can get it freed.
1428 *
1429 * We have to cache the current td in gd_freetd because objcache_put()ing
1430 * it would rip it out from under us while our thread is still active.
1431 */
1432 gd = mycpu;
37af14fe 1433 crit_enter_quick(td);
c070746a
MD
1434 while ((std = gd->gd_freetd) != NULL) {
1435 gd->gd_freetd = NULL;
1436 objcache_put(thread_cache, std);
1437 }
37af14fe 1438 lwkt_deschedule_self(td);
e56e4dea 1439 lwkt_remove_tdallq(td);
c070746a
MD
1440 if (td->td_flags & TDF_ALLOCATED_THREAD)
1441 gd->gd_freetd = td;
99df837e
MD
1442 cpu_thread_exit();
1443}
1444
e56e4dea
MD
1445void
1446lwkt_remove_tdallq(thread_t td)
1447{
1448 KKASSERT(td->td_gd == mycpu);
1449 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1450}
1451
2d93b37a
MD
1452void
1453crit_panic(void)
1454{
1455 thread_t td = curthread;
1456 int lpri = td->td_pri;
1457
1458 td->td_pri = 0;
1459 panic("td_pri is/would-go negative! %p %d", td, lpri);
1460}
1461
d165e668
MD
1462#ifdef SMP
1463
bd8015ca
MD
1464/*
1465 * Called from debugger/panic on cpus which have been stopped. We must still
1466 * process the IPIQ while stopped, even if we were stopped while in a critical
1467 * section (XXX).
1468 *
1469 * If we are dumping also try to process any pending interrupts. This may
1470 * or may not work depending on the state of the cpu at the point it was
1471 * stopped.
1472 */
1473void
1474lwkt_smp_stopped(void)
1475{
1476 globaldata_t gd = mycpu;
1477
1478 crit_enter_gd(gd);
1479 if (dumping) {
1480 lwkt_process_ipiq();
1481 splz();
1482 } else {
1483 lwkt_process_ipiq();
1484 }
1485 crit_exit_gd(gd);
1486}
1487
57aa743c
MD
1488/*
1489 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1490 * get_mplock() has already incremented td_mpcount. We must block and
1491 * not return until giant is held.
1492 *
1493 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1494 * reschedule the thread until it can obtain the giant lock for it.
1495 */
1496void
1497lwkt_mp_lock_contested(void)
1498{
3824f392 1499 ++mplock_countx;
57aa743c 1500 loggiant(beg);
57aa743c 1501 lwkt_switch();
57aa743c 1502 loggiant(end);
57aa743c
MD
1503}
1504
b9eb1c19
MD
1505/*
1506 * The rel_mplock() code will call this function after releasing the
1507 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1508 *
1509 * We then chain an IPI to a single other cpu potentially needing the
1510 * lock. This is a bit heuristical and we can wind up with IPIs flying
1511 * all over the place.
1512 */
1513static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1514
1515void
1516lwkt_mp_lock_uncontested(void)
1517{
1518 globaldata_t gd;
1519 globaldata_t dgd;
1520 cpumask_t mask;
1521 cpumask_t tmpmask;
1522 int cpuid;
1523
1524 if (chain_mplock) {
1525 gd = mycpu;
1526 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1527 mask = mp_lock_contention_mask;
1528 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1529
1530 if (mask) {
1531 if (mask & tmpmask)
1532 cpuid = bsfl(mask & tmpmask);
1533 else
1534 cpuid = bsfl(mask);
1535 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1536 dgd = globaldata_find(cpuid);
1537 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1538 }
1539 }
1540}
1541
1542/*
1543 * The idea is for this IPI to interrupt a potentially lower priority
1544 * thread, such as a user thread, to allow the scheduler to reschedule
1545 * a higher priority kernel thread that needs the MP lock.
1546 *
1547 * For now we set the LWKT reschedule flag which generates an AST in
1548 * doreti, though theoretically it is also possible to possibly preempt
1549 * here if the underlying thread was operating in user mode. Nah.
1550 */
1551static void
1552lwkt_mp_lock_uncontested_remote(void *arg __unused)
1553{
1554 need_lwkt_resched();
1555}
1556
d165e668 1557#endif