bsd.sys.mk: Remove the i386 restriction for using -Werror with WARNS.
[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>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
7d0bac62 67
99df837e 68#include <machine/stdarg.h>
96728c05 69#include <machine/smp.h>
99df837e 70
d850923c
AE
71#if !defined(KTR_CTXSW)
72#define KTR_CTXSW KTR_ALL
73#endif
74KTR_INFO_MASTER(ctxsw);
75KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "sw %p > %p", 2 * sizeof(struct thread *));
76KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
1541028a 77
40aaf5fc
NT
78static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
79
3824f392
MD
80#ifdef SMP
81static int mplock_countx = 0;
82#endif
0f7a3396
MD
83#ifdef INVARIANTS
84static int panic_on_cscount = 0;
85#endif
05220613
MD
86static __int64_t switch_count = 0;
87static __int64_t preempt_hit = 0;
88static __int64_t preempt_miss = 0;
89static __int64_t preempt_weird = 0;
38717797
HP
90static __int64_t token_contention_count = 0;
91static __int64_t mplock_contention_count = 0;
fb0f29c4 92static int lwkt_use_spin_port;
d2f86ad2 93#ifdef SMP
b9eb1c19 94static int chain_mplock = 0;
e381e77c 95static int bgl_yield = 10;
d2f86ad2 96#endif
40aaf5fc 97static struct objcache *thread_cache;
05220613 98
b9eb1c19
MD
99volatile cpumask_t mp_lock_contention_mask;
100
e381e77c
MD
101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
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
85514115
MD
108#ifdef __amd64__
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
b9eb1c19
MD
140#ifdef SMP
141SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
e381e77c 142SYSCTL_INT(_lwkt, OID_AUTO, bgl_yield_delay, CTLFLAG_RW, &bgl_yield, 0, "");
b9eb1c19 143#endif
4b5f931b 144SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 146SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
151SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
152 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
153#endif
05220613 154
57aa743c
MD
155/*
156 * Kernel Trace
157 */
57aa743c
MD
158#if !defined(KTR_GIANT_CONTENTION)
159#define KTR_GIANT_CONTENTION KTR_ALL
160#endif
161
162KTR_INFO_MASTER(giant);
163KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
164KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
165
166#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
167
4b5f931b
MD
168/*
169 * These helper procedures handle the runq, they can only be called from
170 * within a critical section.
75cdbe6c
MD
171 *
172 * WARNING! Prior to SMP being brought up it is possible to enqueue and
173 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
174 * instead of 'mycpu' when referencing the globaldata structure. Once
175 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 176 */
f1d1c3fa
MD
177static __inline
178void
179_lwkt_dequeue(thread_t td)
180{
181 if (td->td_flags & TDF_RUNQ) {
4b5f931b 182 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 183 struct globaldata *gd = td->td_gd;
4b5f931b 184
f1d1c3fa 185 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
186 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
187 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
188 }
189}
190
191static __inline
192void
193_lwkt_enqueue(thread_t td)
194{
7f5d7ed7 195 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 196 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 197 struct globaldata *gd = td->td_gd;
4b5f931b 198
f1d1c3fa 199 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
200 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
201 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
202 }
203}
8ad65e08 204
40aaf5fc
NT
205static __boolean_t
206_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
207{
208 struct thread *td = (struct thread *)obj;
209
210 td->td_kstack = NULL;
211 td->td_kstack_size = 0;
212 td->td_flags = TDF_ALLOCATED_THREAD;
213 return (1);
214}
215
216static void
217_lwkt_thread_dtor(void *obj, void *privdata)
218{
219 struct thread *td = (struct thread *)obj;
220
221 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
222 ("_lwkt_thread_dtor: not allocated from objcache"));
223 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
224 td->td_kstack_size > 0,
225 ("_lwkt_thread_dtor: corrupted stack"));
226 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
227}
228
229/*
230 * Initialize the lwkt s/system.
231 */
232void
233lwkt_init(void)
234{
235 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
236 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
237 NULL, CACHE_NTHREADS/2,
238 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
239}
240
37af14fe
MD
241/*
242 * Schedule a thread to run. As the current thread we can always safely
243 * schedule ourselves, and a shortcut procedure is provided for that
244 * function.
245 *
246 * (non-blocking, self contained on a per cpu basis)
247 */
248void
249lwkt_schedule_self(thread_t td)
250{
251 crit_enter_quick(td);
37af14fe 252 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 253 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 254 _lwkt_enqueue(td);
37af14fe
MD
255 crit_exit_quick(td);
256}
257
258/*
259 * Deschedule a thread.
260 *
261 * (non-blocking, self contained on a per cpu basis)
262 */
263void
264lwkt_deschedule_self(thread_t td)
265{
266 crit_enter_quick(td);
37af14fe
MD
267 _lwkt_dequeue(td);
268 crit_exit_quick(td);
269}
270
8ad65e08
MD
271/*
272 * LWKTs operate on a per-cpu basis
273 *
73e4f7b9 274 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
275 */
276void
277lwkt_gdinit(struct globaldata *gd)
278{
4b5f931b
MD
279 int i;
280
281 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
282 TAILQ_INIT(&gd->gd_tdrunq[i]);
283 gd->gd_runqmask = 0;
73e4f7b9 284 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
285}
286
7d0bac62
MD
287/*
288 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
289 * or LWKT start address before it can be scheduled. If the target cpu is
290 * -1 the thread will be created on the current cpu.
0cfcada1
MD
291 *
292 * If you intend to create a thread without a process context this function
293 * does everything except load the startup and switcher function.
7d0bac62
MD
294 */
295thread_t
d3d32139 296lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 297{
c070746a 298 globaldata_t gd = mycpu;
99df837e 299 void *stack;
7d0bac62 300
c070746a
MD
301 /*
302 * If static thread storage is not supplied allocate a thread. Reuse
303 * a cached free thread if possible. gd_freetd is used to keep an exiting
304 * thread intact through the exit.
305 */
ef0fdad1 306 if (td == NULL) {
c070746a
MD
307 if ((td = gd->gd_freetd) != NULL)
308 gd->gd_freetd = NULL;
309 else
310 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
311 KASSERT((td->td_flags &
312 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
313 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
314 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 315 }
c070746a
MD
316
317 /*
318 * Try to reuse cached stack.
319 */
f470d0c8
MD
320 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
321 if (flags & TDF_ALLOCATED_STACK) {
e4846942 322 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
323 stack = NULL;
324 }
325 }
326 if (stack == NULL) {
e4846942 327 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 328 flags |= TDF_ALLOCATED_STACK;
99df837e 329 }
75cdbe6c 330 if (cpu < 0)
c070746a 331 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 332 else
f470d0c8 333 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 334 return(td);
7d0bac62
MD
335}
336
337/*
338 * Initialize a preexisting thread structure. This function is used by
339 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
340 *
f8c3996b
MD
341 * All threads start out in a critical section at a priority of
342 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
343 * appropriate. This function may send an IPI message when the
344 * requested cpu is not the current cpu and consequently gd_tdallq may
345 * not be initialized synchronously from the point of view of the originating
346 * cpu.
347 *
348 * NOTE! we have to be careful in regards to creating threads for other cpus
349 * if SMP has not yet been activated.
7d0bac62 350 */
41a01a4d
MD
351#ifdef SMP
352
75cdbe6c
MD
353static void
354lwkt_init_thread_remote(void *arg)
355{
356 thread_t td = arg;
357
52eedfb5
MD
358 /*
359 * Protected by critical section held by IPI dispatch
360 */
75cdbe6c
MD
361 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
362}
363
41a01a4d
MD
364#endif
365
7d0bac62 366void
f470d0c8
MD
367lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
368 struct globaldata *gd)
7d0bac62 369{
37af14fe
MD
370 globaldata_t mygd = mycpu;
371
99df837e
MD
372 bzero(td, sizeof(struct thread));
373 td->td_kstack = stack;
f470d0c8 374 td->td_kstack_size = stksize;
d3d32139 375 td->td_flags = flags;
26a0694b 376 td->td_gd = gd;
f8c3996b 377 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
378#ifdef SMP
379 if ((flags & TDF_MPSAFE) == 0)
380 td->td_mpcount = 1;
381#endif
fb0f29c4
MD
382 if (lwkt_use_spin_port)
383 lwkt_initport_spin(&td->td_msgport);
384 else
385 lwkt_initport_thread(&td->td_msgport, td);
99df837e 386 pmap_init_thread(td);
0f7a3396 387#ifdef SMP
5d21b981
MD
388 /*
389 * Normally initializing a thread for a remote cpu requires sending an
390 * IPI. However, the idlethread is setup before the other cpus are
391 * activated so we have to treat it as a special case. XXX manipulation
392 * of gd_tdallq requires the BGL.
393 */
394 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 395 crit_enter_gd(mygd);
75cdbe6c 396 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 397 crit_exit_gd(mygd);
75cdbe6c 398 } else {
2db3b277 399 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 400 }
0f7a3396 401#else
37af14fe 402 crit_enter_gd(mygd);
0f7a3396 403 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 404 crit_exit_gd(mygd);
0f7a3396 405#endif
73e4f7b9
MD
406}
407
408void
409lwkt_set_comm(thread_t td, const char *ctl, ...)
410{
e2565a42 411 __va_list va;
73e4f7b9 412
e2565a42 413 __va_start(va, ctl);
379210cb 414 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 415 __va_end(va);
7d0bac62
MD
416}
417
99df837e 418void
73e4f7b9 419lwkt_hold(thread_t td)
99df837e 420{
73e4f7b9
MD
421 ++td->td_refs;
422}
423
424void
425lwkt_rele(thread_t td)
426{
427 KKASSERT(td->td_refs > 0);
428 --td->td_refs;
429}
430
431void
432lwkt_wait_free(thread_t td)
433{
434 while (td->td_refs)
377d4740 435 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
436}
437
438void
439lwkt_free_thread(thread_t td)
440{
d9eea1a5 441 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
442 ("lwkt_free_thread: did not exit! %p", td));
443
40aaf5fc
NT
444 if (td->td_flags & TDF_ALLOCATED_THREAD) {
445 objcache_put(thread_cache, td);
446 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
447 /* client-allocated struct with internally allocated stack */
448 KASSERT(td->td_kstack && td->td_kstack_size > 0,
449 ("lwkt_free_thread: corrupted stack"));
450 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
451 td->td_kstack = NULL;
452 td->td_kstack_size = 0;
99df837e
MD
453 }
454}
455
456
8ad65e08
MD
457/*
458 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
459 * switch to the idlethread. Switching must occur within a critical
460 * section to avoid races with the scheduling queue.
461 *
462 * We always have full control over our cpu's run queue. Other cpus
463 * that wish to manipulate our queue must use the cpu_*msg() calls to
464 * talk to our cpu, so a critical section is all that is needed and
465 * the result is very, very fast thread switching.
466 *
96728c05
MD
467 * The LWKT scheduler uses a fixed priority model and round-robins at
468 * each priority level. User process scheduling is a totally
469 * different beast and LWKT priorities should not be confused with
470 * user process priorities.
f1d1c3fa 471 *
96728c05
MD
472 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
473 * cleans it up. Note that the td_switch() function cannot do anything that
474 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
475 * the target thread (not the current thread). It's nice to have a scheduler
476 * that does not need the MP lock to work because it allows us to do some
477 * really cool high-performance MP lock optimizations.
69d78e99
MD
478 *
479 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
480 * is not called by the current thread in the preemption case, only when
481 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
482 */
483void
484lwkt_switch(void)
485{
37af14fe
MD
486 globaldata_t gd = mycpu;
487 thread_t td = gd->gd_curthread;
8ad65e08 488 thread_t ntd;
8a8d5d85
MD
489#ifdef SMP
490 int mpheld;
491#endif
8ad65e08 492
46a3f46d 493 /*
27e88a6e
MD
494 * Switching from within a 'fast' (non thread switched) interrupt or IPI
495 * is illegal. However, we may have to do it anyway if we hit a fatal
496 * kernel trap or we have paniced.
497 *
498 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 499 */
27e88a6e
MD
500 if (gd->gd_intr_nesting_level) {
501 int savegdnest;
502 int savegdtrap;
503
504 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
505 panic("lwkt_switch: cannot switch from within "
506 "a fast interrupt, yet, td %p\n", td);
507 } else {
508 savegdnest = gd->gd_intr_nesting_level;
509 savegdtrap = gd->gd_trap_nesting_level;
510 gd->gd_intr_nesting_level = 0;
511 gd->gd_trap_nesting_level = 0;
a7422615
MD
512 if ((td->td_flags & TDF_PANICWARN) == 0) {
513 td->td_flags |= TDF_PANICWARN;
6ea70f76 514 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 515 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 516 print_backtrace();
a7422615 517 }
27e88a6e
MD
518 lwkt_switch();
519 gd->gd_intr_nesting_level = savegdnest;
520 gd->gd_trap_nesting_level = savegdtrap;
521 return;
522 }
96728c05 523 }
ef0fdad1 524
cb973d15
MD
525 /*
526 * Passive release (used to transition from user to kernel mode
527 * when we block or switch rather then when we enter the kernel).
528 * This function is NOT called if we are switching into a preemption
529 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
530 * our current process designation (if we have one) and become a true
531 * LWKT thread, and may also hand the current process designation to
532 * another process and schedule thread.
cb973d15
MD
533 */
534 if (td->td_release)
535 td->td_release(td);
536
37af14fe 537 crit_enter_gd(gd);
9d265729
MD
538 if (td->td_toks)
539 lwkt_relalltokens(td);
540
541 /*
b02926de
MD
542 * We had better not be holding any spin locks, but don't get into an
543 * endless panic loop.
9d265729 544 */
bbb31c5d
MD
545 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
546 ("lwkt_switch: still holding a shared spinlock %p!",
547 gd->gd_spinlock_rd));
d666840a
MD
548 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
549 ("lwkt_switch: still holding %d exclusive spinlocks!",
550 gd->gd_spinlocks_wr));
9d265729 551
8a8d5d85
MD
552
553#ifdef SMP
554 /*
555 * td_mpcount cannot be used to determine if we currently hold the
556 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
557 * to get the lock, and switch out if it can't. Our ownership of
558 * the actual lock will remain stable while we are in a critical section
559 * (but, of course, another cpu may own or release the lock so the
560 * actual value of mp_lock is not stable).
8a8d5d85
MD
561 */
562 mpheld = MP_LOCK_HELD();
0f7a3396
MD
563#ifdef INVARIANTS
564 if (td->td_cscount) {
6ea70f76 565 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
566 td);
567 if (panic_on_cscount)
568 panic("switching while mastering cpusync");
569 }
570#endif
8a8d5d85 571#endif
99df837e
MD
572 if ((ntd = td->td_preempted) != NULL) {
573 /*
574 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
575 * thread. This occurs transparently, whether the preempted thread
576 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
577 * itself).
578 *
579 * We have to setup the MP lock for the original thread after backing
580 * out the adjustment that was made to curthread when the original
581 * was preempted.
99df837e 582 */
26a0694b 583 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 584#ifdef SMP
96728c05 585 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 586 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
587 td, ntd, td->td_mpcount, ntd->td_mpcount);
588 }
8a8d5d85
MD
589 if (ntd->td_mpcount) {
590 td->td_mpcount -= ntd->td_mpcount;
591 KKASSERT(td->td_mpcount >= 0);
592 }
593#endif
26a0694b 594 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
595
596 /*
b9eb1c19
MD
597 * The interrupt may have woken a thread up, we need to properly
598 * set the reschedule flag if the originally interrupted thread is
599 * at a lower priority.
8ec60c3f
MD
600 */
601 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
602 need_lwkt_resched();
8a8d5d85 603 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 604 } else {
4b5f931b
MD
605 /*
606 * Priority queue / round-robin at each priority. Note that user
607 * processes run at a fixed, low priority and the user process
608 * scheduler deals with interactions between user processes
609 * by scheduling and descheduling them from the LWKT queue as
610 * necessary.
8a8d5d85
MD
611 *
612 * We have to adjust the MP lock for the target thread. If we
613 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
614 * thread that does not need the MP lock. If we cannot, we spin
615 * instead of HLT.
616 *
617 * A similar issue exists for the tokens held by the target thread.
618 * If we cannot obtain ownership of the tokens we cannot immediately
619 * schedule the thread.
620 */
621
8ec60c3f
MD
622 /*
623 * If an LWKT reschedule was requested, well that is what we are
624 * doing now so clear it.
625 */
626 clear_lwkt_resched();
4b5f931b
MD
627again:
628 if (gd->gd_runqmask) {
629 int nq = bsrl(gd->gd_runqmask);
630 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
631 gd->gd_runqmask &= ~(1 << nq);
632 goto again;
633 }
8a8d5d85 634#ifdef SMP
41a01a4d 635 /*
df6b8ba0
MD
636 * THREAD SELECTION FOR AN SMP MACHINE BUILD
637 *
41a01a4d
MD
638 * If the target needs the MP lock and we couldn't get it,
639 * or if the target is holding tokens and we could not
640 * gain ownership of the tokens, continue looking for a
641 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
642 *
643 * NOTE: the mpheld variable invalid after this conditional, it
644 * can change due to both cpu_try_mplock() returning success
9d265729 645 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
646 * we are trying to check the mpcount of a thread other then
647 * the current thread. Because of this, if the current thread
648 * is not holding td_mpcount, an IPI indirectly run via
9d265729 649 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 650 * cause the core MP lock to be released.
41a01a4d
MD
651 */
652 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 653 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 654 ) {
8a8d5d85 655 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
656
657 mpheld = MP_LOCK_HELD();
658 ntd = NULL;
8a8d5d85
MD
659 while (rqmask) {
660 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 661 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 662 /* spinning due to MP lock being held */
38717797 663#ifdef INVARIANTS
a453459d 664 ++mplock_contention_count;
38717797 665#endif
a453459d 666 /* mplock still not held, 'mpheld' still valid */
41a01a4d 667 continue;
38717797 668 }
a453459d
MD
669
670 /*
9d265729 671 * mpheld state invalid after getalltokens call returns
a453459d
MD
672 * failure, but the variable is only needed for
673 * the loop.
674 */
9d265729 675 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 676 /* spinning due to token contention */
38717797 677#ifdef INVARIANTS
a453459d 678 ++token_contention_count;
38717797 679#endif
a453459d 680 mpheld = MP_LOCK_HELD();
41a01a4d 681 continue;
38717797 682 }
41a01a4d 683 break;
8a8d5d85
MD
684 }
685 if (ntd)
686 break;
687 rqmask &= ~(1 << nq);
688 nq = bsrl(rqmask);
b9eb1c19
MD
689
690 /*
691 * We have two choices. We can either refuse to run a
692 * user thread when a kernel thread needs the MP lock
693 * but could not get it, or we can allow it to run but
694 * then expect an IPI (hopefully) later on to force a
695 * reschedule when the MP lock might become available.
696 */
697 if (nq < TDPRI_KERN_LPSCHED) {
698 if (chain_mplock == 0)
699 break;
700 atomic_set_int(&mp_lock_contention_mask,
701 gd->gd_cpumask);
702 /* continue loop, allow user threads to be scheduled */
703 }
8a8d5d85
MD
704 }
705 if (ntd == NULL) {
b402c633 706 cpu_mplock_contested();
a2a5ad0d
MD
707 ntd = &gd->gd_idlethread;
708 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 709 goto using_idle_thread;
8a8d5d85 710 } else {
344ad853 711 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
712 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
713 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
714 }
715 } else {
3824f392
MD
716 if (ntd->td_mpcount)
717 ++mplock_countx;
344ad853 718 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
719 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
720 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
721 }
722#else
df6b8ba0
MD
723 /*
724 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
725 * worry about tokens or the BGL. However, we still have
726 * to call lwkt_getalltokens() in order to properly detect
727 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 728 */
7eb611ef 729 lwkt_getalltokens(ntd);
344ad853 730 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
731 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
732 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 733#endif
4b5f931b 734 } else {
3c23a41a 735 /*
60f945af
MD
736 * We have nothing to run but only let the idle loop halt
737 * the cpu if there are no pending interrupts.
3c23a41a 738 */
a2a5ad0d 739 ntd = &gd->gd_idlethread;
60f945af 740 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 741 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 742#ifdef SMP
df6b8ba0
MD
743using_idle_thread:
744 /*
745 * The idle thread should not be holding the MP lock unless we
746 * are trapping in the kernel or in a panic. Since we select the
747 * idle thread unconditionally when no other thread is available,
748 * if the MP lock is desired during a panic or kernel trap, we
749 * have to loop in the scheduler until we get it.
750 */
751 if (ntd->td_mpcount) {
752 mpheld = MP_LOCK_HELD();
b402c633 753 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 754 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
755 } else if (mpheld == 0) {
756 cpu_mplock_contested();
df6b8ba0 757 goto again;
b402c633 758 }
df6b8ba0 759 }
a453459d 760#endif
4b5f931b 761 }
f1d1c3fa 762 }
26a0694b
MD
763 KASSERT(ntd->td_pri >= TDPRI_CRIT,
764 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
765
766 /*
767 * Do the actual switch. If the new target does not need the MP lock
768 * and we are holding it, release the MP lock. If the new target requires
769 * the MP lock we have already acquired it for the target.
770 */
771#ifdef SMP
772 if (ntd->td_mpcount == 0 ) {
773 if (MP_LOCK_HELD())
774 cpu_rel_mplock();
775 } else {
a453459d 776 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
777 }
778#endif
94f6d86e
MD
779 if (td != ntd) {
780 ++switch_count;
85514115 781#ifdef __amd64__
0855a2af 782 KKASSERT(jg_tos_ok(ntd));
85514115 783#endif
d850923c 784 KTR_LOG(ctxsw_sw, td, ntd);
f1d1c3fa 785 td->td_switch(ntd);
94f6d86e 786 }
37af14fe
MD
787 /* NOTE: current cpu may have changed after switch */
788 crit_exit_quick(td);
8ad65e08
MD
789}
790
b68b7282 791/*
96728c05
MD
792 * Request that the target thread preempt the current thread. Preemption
793 * only works under a specific set of conditions:
b68b7282 794 *
96728c05
MD
795 * - We are not preempting ourselves
796 * - The target thread is owned by the current cpu
797 * - We are not currently being preempted
798 * - The target is not currently being preempted
d3d1cbc8
MD
799 * - We are not holding any spin locks
800 * - The target thread is not holding any tokens
96728c05
MD
801 * - We are able to satisfy the target's MP lock requirements (if any).
802 *
803 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
804 * this is called via lwkt_schedule() through the td_preemptable callback.
805 * critpri is the managed critical priority that we should ignore in order
806 * to determine whether preemption is possible (aka usually just the crit
807 * priority of lwkt_schedule() itself).
b68b7282 808 *
26a0694b
MD
809 * XXX at the moment we run the target thread in a critical section during
810 * the preemption in order to prevent the target from taking interrupts
811 * that *WE* can't. Preemption is strictly limited to interrupt threads
812 * and interrupt-like threads, outside of a critical section, and the
813 * preempted source thread will be resumed the instant the target blocks
814 * whether or not the source is scheduled (i.e. preemption is supposed to
815 * be as transparent as possible).
4b5f931b 816 *
8a8d5d85
MD
817 * The target thread inherits our MP count (added to its own) for the
818 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
819 * MP lock during the preemption. Therefore, any preempting targets must be
820 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
821 * out of sync with the physical mp_lock, but we do not have to preserve
822 * the original ownership of the lock if it was out of synch (that is, we
823 * can leave it synchronized on return).
b68b7282
MD
824 */
825void
96728c05 826lwkt_preempt(thread_t ntd, int critpri)
b68b7282 827{
46a3f46d 828 struct globaldata *gd = mycpu;
0a3f9b47 829 thread_t td;
8a8d5d85
MD
830#ifdef SMP
831 int mpheld;
57c254db 832 int savecnt;
8a8d5d85 833#endif
b68b7282 834
26a0694b 835 /*
96728c05
MD
836 * The caller has put us in a critical section. We can only preempt
837 * if the caller of the caller was not in a critical section (basically
d666840a 838 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 839 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 840 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
841 *
842 * YYY The target thread must be in a critical section (else it must
843 * inherit our critical section? I dunno yet).
41a01a4d 844 *
0a3f9b47 845 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
846 */
847 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 848
0a3f9b47 849 td = gd->gd_curthread;
0a3f9b47 850 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
851 ++preempt_miss;
852 return;
853 }
96728c05
MD
854 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
855 ++preempt_miss;
8ec60c3f 856 need_lwkt_resched();
96728c05
MD
857 return;
858 }
859#ifdef SMP
46a3f46d 860 if (ntd->td_gd != gd) {
96728c05 861 ++preempt_miss;
8ec60c3f 862 need_lwkt_resched();
96728c05
MD
863 return;
864 }
865#endif
41a01a4d 866 /*
d3d1cbc8 867 * Take the easy way out and do not preempt if we are holding
d666840a 868 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
869 * preempted interlock against the target thread's tokens and whether
870 * we can get all the target thread's tokens, but this situation
871 * should not occur very often so its easier to simply not preempt.
d666840a
MD
872 * Also, plain spinlocks are impossible to figure out at this point so
873 * just don't preempt.
d3d1cbc8
MD
874 *
875 * Do not try to preempt if the target thread is holding any tokens.
876 * We could try to acquire the tokens but this case is so rare there
877 * is no need to support it.
41a01a4d 878 */
bbb31c5d 879 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 880 ++preempt_miss;
8ec60c3f 881 need_lwkt_resched();
41a01a4d
MD
882 return;
883 }
d3d1cbc8
MD
884 if (ntd->td_toks) {
885 ++preempt_miss;
886 need_lwkt_resched();
887 return;
888 }
26a0694b
MD
889 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
890 ++preempt_weird;
8ec60c3f 891 need_lwkt_resched();
26a0694b
MD
892 return;
893 }
894 if (ntd->td_preempted) {
4b5f931b 895 ++preempt_hit;
8ec60c3f 896 need_lwkt_resched();
26a0694b 897 return;
b68b7282 898 }
8a8d5d85 899#ifdef SMP
a2a5ad0d
MD
900 /*
901 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
902 * to or from the MP lock. In this case td_mpcount will be pre-disposed
903 * (non-zero) but not actually synchronized with the actual state of the
904 * lock. We can use it to imply an MP lock requirement for the
905 * preemption but we cannot use it to test whether we hold the MP lock
906 * or not.
a2a5ad0d 907 */
96728c05 908 savecnt = td->td_mpcount;
71ef2f5c 909 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
910 ntd->td_mpcount += td->td_mpcount;
911 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
912 ntd->td_mpcount -= td->td_mpcount;
913 ++preempt_miss;
8ec60c3f 914 need_lwkt_resched();
8a8d5d85
MD
915 return;
916 }
917#endif
26a0694b 918
8ec60c3f
MD
919 /*
920 * Since we are able to preempt the current thread, there is no need to
921 * call need_lwkt_resched().
922 */
26a0694b
MD
923 ++preempt_hit;
924 ntd->td_preempted = td;
925 td->td_flags |= TDF_PREEMPT_LOCK;
d850923c 926 KTR_LOG(ctxsw_pre, td, ntd);
26a0694b 927 td->td_switch(ntd);
b9eb1c19 928
26a0694b 929 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
930#ifdef SMP
931 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
932 mpheld = MP_LOCK_HELD();
933 if (mpheld && td->td_mpcount == 0)
96728c05 934 cpu_rel_mplock();
71ef2f5c 935 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
936 panic("lwkt_preempt(): MP lock was not held through");
937#endif
26a0694b
MD
938 ntd->td_preempted = NULL;
939 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
940}
941
f1d1c3fa 942/*
faaeffac 943 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 944 *
faaeffac
MD
945 * td_nest_count prevents deep nesting via splz() or doreti() which
946 * might otherwise blow out the kernel stack. Note that except for
947 * this special case, we MUST call splz() here to handle any
948 * pending ints, particularly after we switch, or we might accidently
949 * halt the cpu with interrupts pending.
ef0fdad1 950 *
f1d1c3fa
MD
951 * (self contained on a per cpu basis)
952 */
953void
faaeffac 954splz_check(void)
f1d1c3fa 955{
7966cb69
MD
956 globaldata_t gd = mycpu;
957 thread_t td = gd->gd_curthread;
ef0fdad1 958
46a3f46d 959 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 960 splz();
f1d1c3fa
MD
961}
962
8ad65e08 963/*
faaeffac
MD
964 * This implements a normal yield which will yield to equal priority
965 * threads as well as higher priority threads. Note that gd_reqflags
966 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
967 *
968 * (self contained on a per cpu basis)
8ad65e08
MD
969 */
970void
f1d1c3fa 971lwkt_yield(void)
8ad65e08 972{
37af14fe 973 lwkt_schedule_self(curthread);
f1d1c3fa
MD
974 lwkt_switch();
975}
976
3824f392
MD
977/*
978 * This function is used along with the lwkt_passive_recover() inline
979 * by the trap code to negotiate a passive release of the current
980 * process/lwp designation with the user scheduler.
981 */
982void
983lwkt_passive_release(struct thread *td)
984{
985 struct lwp *lp = td->td_lwp;
986
987 td->td_release = NULL;
988 lwkt_setpri_self(TDPRI_KERN_USER);
989 lp->lwp_proc->p_usched->release_curproc(lp);
990}
991
992/*
993 * Make a kernel thread act as if it were in user mode with regards
994 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
995 * loops which may be potentially cpu-bound can call lwkt_user_yield().
996 *
997 * The lwkt_user_yield() function is designed to have very low overhead
998 * if no yield is determined to be needed.
999 */
1000void
1001lwkt_user_yield(void)
1002{
1003 thread_t td = curthread;
1004 struct lwp *lp = td->td_lwp;
1005
1006#ifdef SMP
1007 /*
1008 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1009 * kernel can prevent other cpus from servicing interrupt threads
1010 * which still require the MP lock (which is a lot of them). This
1011 * has a chaining effect since if the interrupt is blocked, so is
1012 * the event, so normal scheduling will not pick up on the problem.
1013 */
1014 if (mplock_countx && td->td_mpcount) {
1015 int savecnt = td->td_mpcount;
1016
1017 td->td_mpcount = 1;
e381e77c 1018 mplock_countx = 0;
3824f392 1019 rel_mplock();
e381e77c 1020 DELAY(bgl_yield);
3824f392
MD
1021 get_mplock();
1022 td->td_mpcount = savecnt;
3824f392
MD
1023 }
1024#endif
1025
1026 /*
1027 * Another kernel thread wants the cpu
1028 */
1029 if (lwkt_resched_wanted())
1030 lwkt_switch();
1031
1032 /*
1033 * If the user scheduler has asynchronously determined that the current
1034 * process (when running in user mode) needs to lose the cpu then make
1035 * sure we are released.
1036 */
1037 if (user_resched_wanted()) {
1038 if (td->td_release)
1039 td->td_release(td);
1040 }
1041
1042 /*
1043 * If we are released reduce our priority
1044 */
1045 if (td->td_release == NULL) {
1046 if (lwkt_check_resched(td) > 0)
1047 lwkt_switch();
e381e77c
MD
1048 if (lp) {
1049 lp->lwp_proc->p_usched->acquire_curproc(lp);
1050 td->td_release = lwkt_passive_release;
1051 lwkt_setpri_self(TDPRI_USER_NORM);
1052 }
3824f392
MD
1053 }
1054}
1055
b9eb1c19
MD
1056/*
1057 * Return 0 if no runnable threads are pending at the same or higher
1058 * priority as the passed thread.
1059 *
1060 * Return 1 if runnable threads are pending at the same priority.
1061 *
1062 * Return 2 if runnable threads are pending at a higher priority.
1063 */
1064int
1065lwkt_check_resched(thread_t td)
1066{
1067 int pri = td->td_pri & TDPRI_MASK;
1068
1069 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1070 return(2);
1071 if (TAILQ_NEXT(td, td_threadq))
1072 return(1);
1073 return(0);
1074}
1075
8ad65e08 1076/*
f1d1c3fa
MD
1077 * Generic schedule. Possibly schedule threads belonging to other cpus and
1078 * deal with threads that might be blocked on a wait queue.
1079 *
0a3f9b47
MD
1080 * We have a little helper inline function which does additional work after
1081 * the thread has been enqueued, including dealing with preemption and
1082 * setting need_lwkt_resched() (which prevents the kernel from returning
1083 * to userland until it has processed higher priority threads).
6330a558
MD
1084 *
1085 * It is possible for this routine to be called after a failed _enqueue
1086 * (due to the target thread migrating, sleeping, or otherwise blocked).
1087 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1088 *
1089 * reschedok is an optimized constant propagated from lwkt_schedule() or
1090 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1091 * reschedule to be requested if the target thread has a higher priority.
1092 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1093 * be 0, prevented undesired reschedules.
8ad65e08 1094 */
0a3f9b47
MD
1095static __inline
1096void
361d01dd 1097_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1098{
b9eb1c19 1099 thread_t otd;
c730be20 1100
6330a558 1101 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1102 if (ntd->td_preemptable && reschedok) {
6330a558 1103 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1104 } else if (reschedok) {
b9eb1c19
MD
1105 otd = curthread;
1106 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1107 need_lwkt_resched();
6330a558 1108 }
0a3f9b47
MD
1109 }
1110}
1111
361d01dd 1112static __inline
8ad65e08 1113void
361d01dd 1114_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1115{
37af14fe
MD
1116 globaldata_t mygd = mycpu;
1117
41a01a4d 1118 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1119 crit_enter_gd(mygd);
9388413d 1120 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1121 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1122 _lwkt_enqueue(td);
1123 } else {
f1d1c3fa 1124 /*
7cd8d145
MD
1125 * If we own the thread, there is no race (since we are in a
1126 * critical section). If we do not own the thread there might
1127 * be a race but the target cpu will deal with it.
f1d1c3fa 1128 */
0f7a3396 1129#ifdef SMP
7cd8d145 1130 if (td->td_gd == mygd) {
9d265729 1131 _lwkt_enqueue(td);
361d01dd 1132 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1133 } else {
e381e77c 1134 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1135 }
0f7a3396 1136#else
7cd8d145 1137 _lwkt_enqueue(td);
361d01dd 1138 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1139#endif
8ad65e08 1140 }
37af14fe 1141 crit_exit_gd(mygd);
8ad65e08
MD
1142}
1143
361d01dd
MD
1144void
1145lwkt_schedule(thread_t td)
1146{
1147 _lwkt_schedule(td, 1);
1148}
1149
1150void
1151lwkt_schedule_noresched(thread_t td)
1152{
1153 _lwkt_schedule(td, 0);
1154}
1155
e381e77c
MD
1156/*
1157 * When scheduled remotely if frame != NULL the IPIQ is being
1158 * run via doreti or an interrupt then preemption can be allowed.
1159 *
1160 * To allow preemption we have to drop the critical section so only
1161 * one is present in _lwkt_schedule_post.
1162 */
1163static void
1164lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1165{
1166 thread_t td = curthread;
1167 thread_t ntd = arg;
1168
1169 if (frame && ntd->td_preemptable) {
1170 crit_exit_noyield(td);
1171 _lwkt_schedule(ntd, 1);
1172 crit_enter_quick(td);
1173 } else {
1174 _lwkt_schedule(ntd, 1);
1175 }
1176}
1177
52eedfb5
MD
1178#ifdef SMP
1179
d9eea1a5 1180/*
52eedfb5
MD
1181 * Thread migration using a 'Pull' method. The thread may or may not be
1182 * the current thread. It MUST be descheduled and in a stable state.
1183 * lwkt_giveaway() must be called on the cpu owning the thread.
1184 *
1185 * At any point after lwkt_giveaway() is called, the target cpu may
1186 * 'pull' the thread by calling lwkt_acquire().
1187 *
ae8e83e6
MD
1188 * We have to make sure the thread is not sitting on a per-cpu tsleep
1189 * queue or it will blow up when it moves to another cpu.
1190 *
52eedfb5 1191 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1192 */
52eedfb5
MD
1193void
1194lwkt_giveaway(thread_t td)
1195{
3b4192fb 1196 globaldata_t gd = mycpu;
52eedfb5 1197
3b4192fb
MD
1198 crit_enter_gd(gd);
1199 if (td->td_flags & TDF_TSLEEPQ)
1200 tsleep_remove(td);
1201 KKASSERT(td->td_gd == gd);
1202 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1203 td->td_flags |= TDF_MIGRATING;
1204 crit_exit_gd(gd);
52eedfb5
MD
1205}
1206
a2a5ad0d
MD
1207void
1208lwkt_acquire(thread_t td)
1209{
37af14fe
MD
1210 globaldata_t gd;
1211 globaldata_t mygd;
a2a5ad0d 1212
52eedfb5 1213 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1214 gd = td->td_gd;
37af14fe 1215 mygd = mycpu;
52eedfb5 1216 if (gd != mycpu) {
35238fa5 1217 cpu_lfence();
52eedfb5 1218 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1219 crit_enter_gd(mygd);
df910c23
MD
1220 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1221#ifdef SMP
1222 lwkt_process_ipiq();
1223#endif
52eedfb5 1224 cpu_lfence();
df910c23 1225 }
37af14fe 1226 td->td_gd = mygd;
52eedfb5
MD
1227 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1228 td->td_flags &= ~TDF_MIGRATING;
1229 crit_exit_gd(mygd);
1230 } else {
1231 crit_enter_gd(mygd);
1232 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1233 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1234 crit_exit_gd(mygd);
a2a5ad0d
MD
1235 }
1236}
1237
52eedfb5
MD
1238#endif
1239
f1d1c3fa
MD
1240/*
1241 * Generic deschedule. Descheduling threads other then your own should be
1242 * done only in carefully controlled circumstances. Descheduling is
1243 * asynchronous.
1244 *
1245 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1246 */
1247void
1248lwkt_deschedule(thread_t td)
1249{
f1d1c3fa 1250 crit_enter();
b8a98473 1251#ifdef SMP
f1d1c3fa
MD
1252 if (td == curthread) {
1253 _lwkt_dequeue(td);
1254 } else {
a72187e9 1255 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1256 _lwkt_dequeue(td);
1257 } else {
b8a98473 1258 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1259 }
1260 }
b8a98473
MD
1261#else
1262 _lwkt_dequeue(td);
1263#endif
f1d1c3fa
MD
1264 crit_exit();
1265}
1266
4b5f931b
MD
1267/*
1268 * Set the target thread's priority. This routine does not automatically
1269 * switch to a higher priority thread, LWKT threads are not designed for
1270 * continuous priority changes. Yield if you want to switch.
1271 *
1272 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1273 * of the td_pri field. The specified priority may also indicate zero or
1274 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1275 *
1276 * Note that we requeue the thread whether it winds up on a different runq
1277 * or not. uio_yield() depends on this and the routine is not normally
1278 * called with the same priority otherwise.
4b5f931b
MD
1279 */
1280void
1281lwkt_setpri(thread_t td, int pri)
1282{
26a0694b 1283 KKASSERT(pri >= 0);
a72187e9 1284 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1285 crit_enter();
1286 if (td->td_flags & TDF_RUNQ) {
1287 _lwkt_dequeue(td);
1288 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1289 _lwkt_enqueue(td);
1290 } else {
1291 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1292 }
1293 crit_exit();
1294}
1295
1296void
1297lwkt_setpri_self(int pri)
1298{
1299 thread_t td = curthread;
1300
4b5f931b
MD
1301 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1302 crit_enter();
1303 if (td->td_flags & TDF_RUNQ) {
1304 _lwkt_dequeue(td);
1305 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1306 _lwkt_enqueue(td);
1307 } else {
1308 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1309 }
1310 crit_exit();
1311}
1312
5d21b981 1313/*
52eedfb5
MD
1314 * Migrate the current thread to the specified cpu.
1315 *
1316 * This is accomplished by descheduling ourselves from the current cpu,
1317 * moving our thread to the tdallq of the target cpu, IPI messaging the
1318 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1319 * races while the thread is being migrated.
ae8e83e6
MD
1320 *
1321 * We must be sure to remove ourselves from the current cpu's tsleepq
1322 * before potentially moving to another queue. The thread can be on
1323 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1324 */
3d28ff59 1325#ifdef SMP
5d21b981 1326static void lwkt_setcpu_remote(void *arg);
3d28ff59 1327#endif
5d21b981
MD
1328
1329void
1330lwkt_setcpu_self(globaldata_t rgd)
1331{
1332#ifdef SMP
1333 thread_t td = curthread;
1334
1335 if (td->td_gd != rgd) {
1336 crit_enter_quick(td);
ae8e83e6 1337 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1338 tsleep_remove(td);
5d21b981
MD
1339 td->td_flags |= TDF_MIGRATING;
1340 lwkt_deschedule_self(td);
52eedfb5 1341 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1342 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1343 lwkt_switch();
1344 /* we are now on the target cpu */
52eedfb5 1345 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1346 crit_exit_quick(td);
1347 }
1348#endif
1349}
1350
ecdefdda
MD
1351void
1352lwkt_migratecpu(int cpuid)
1353{
1354#ifdef SMP
1355 globaldata_t rgd;
1356
1357 rgd = globaldata_find(cpuid);
1358 lwkt_setcpu_self(rgd);
1359#endif
1360}
1361
5d21b981
MD
1362/*
1363 * Remote IPI for cpu migration (called while in a critical section so we
1364 * do not have to enter another one). The thread has already been moved to
1365 * our cpu's allq, but we must wait for the thread to be completely switched
1366 * out on the originating cpu before we schedule it on ours or the stack
1367 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1368 * change to main memory.
1369 *
1370 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1371 * against wakeups. It is best if this interface is used only when there
1372 * are no pending events that might try to schedule the thread.
1373 */
3d28ff59 1374#ifdef SMP
5d21b981
MD
1375static void
1376lwkt_setcpu_remote(void *arg)
1377{
1378 thread_t td = arg;
1379 globaldata_t gd = mycpu;
1380
df910c23
MD
1381 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1382#ifdef SMP
1383 lwkt_process_ipiq();
1384#endif
35238fa5 1385 cpu_lfence();
df910c23 1386 }
5d21b981 1387 td->td_gd = gd;
35238fa5 1388 cpu_sfence();
5d21b981 1389 td->td_flags &= ~TDF_MIGRATING;
9388413d 1390 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1391 _lwkt_enqueue(td);
1392}
3d28ff59 1393#endif
5d21b981 1394
553ea3c8 1395struct lwp *
4b5f931b
MD
1396lwkt_preempted_proc(void)
1397{
73e4f7b9 1398 thread_t td = curthread;
4b5f931b
MD
1399 while (td->td_preempted)
1400 td = td->td_preempted;
553ea3c8 1401 return(td->td_lwp);
4b5f931b
MD
1402}
1403
99df837e
MD
1404/*
1405 * Create a kernel process/thread/whatever. It shares it's address space
1406 * with proc0 - ie: kernel only.
1407 *
365fa13f
MD
1408 * NOTE! By default new threads are created with the MP lock held. A
1409 * thread which does not require the MP lock should release it by calling
1410 * rel_mplock() at the start of the new thread.
99df837e
MD
1411 */
1412int
1413lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1414 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1415 const char *fmt, ...)
99df837e 1416{
73e4f7b9 1417 thread_t td;
e2565a42 1418 __va_list ap;
99df837e 1419
d3d32139 1420 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1421 tdflags);
a2a5ad0d
MD
1422 if (tdp)
1423 *tdp = td;
709799ea 1424 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1425
1426 /*
1427 * Set up arg0 for 'ps' etc
1428 */
e2565a42 1429 __va_start(ap, fmt);
379210cb 1430 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1431 __va_end(ap);
99df837e
MD
1432
1433 /*
1434 * Schedule the thread to run
1435 */
ef0fdad1
MD
1436 if ((td->td_flags & TDF_STOPREQ) == 0)
1437 lwkt_schedule(td);
1438 else
1439 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1440 return 0;
1441}
1442
1443/*
1444 * Destroy an LWKT thread. Warning! This function is not called when
1445 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1446 * uses a different reaping mechanism.
1447 */
1448void
1449lwkt_exit(void)
1450{
1451 thread_t td = curthread;
c070746a 1452 thread_t std;
8826f33a 1453 globaldata_t gd;
99df837e
MD
1454
1455 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1456 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1457 caps_exit(td);
c070746a
MD
1458
1459 /*
1460 * Get us into a critical section to interlock gd_freetd and loop
1461 * until we can get it freed.
1462 *
1463 * We have to cache the current td in gd_freetd because objcache_put()ing
1464 * it would rip it out from under us while our thread is still active.
1465 */
1466 gd = mycpu;
37af14fe 1467 crit_enter_quick(td);
c070746a
MD
1468 while ((std = gd->gd_freetd) != NULL) {
1469 gd->gd_freetd = NULL;
1470 objcache_put(thread_cache, std);
1471 }
3b4192fb
MD
1472
1473 /*
1474 * Remove thread resources from kernel lists and deschedule us for
1475 * the last time.
1476 */
1477 if (td->td_flags & TDF_TSLEEPQ)
1478 tsleep_remove(td);
79eae878 1479 biosched_done(td);
37af14fe 1480 lwkt_deschedule_self(td);
e56e4dea 1481 lwkt_remove_tdallq(td);
c070746a
MD
1482 if (td->td_flags & TDF_ALLOCATED_THREAD)
1483 gd->gd_freetd = td;
99df837e
MD
1484 cpu_thread_exit();
1485}
1486
e56e4dea
MD
1487void
1488lwkt_remove_tdallq(thread_t td)
1489{
1490 KKASSERT(td->td_gd == mycpu);
1491 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1492}
1493
2d93b37a
MD
1494void
1495crit_panic(void)
1496{
1497 thread_t td = curthread;
1498 int lpri = td->td_pri;
1499
1500 td->td_pri = 0;
1501 panic("td_pri is/would-go negative! %p %d", td, lpri);
1502}
1503
d165e668
MD
1504#ifdef SMP
1505
bd8015ca
MD
1506/*
1507 * Called from debugger/panic on cpus which have been stopped. We must still
1508 * process the IPIQ while stopped, even if we were stopped while in a critical
1509 * section (XXX).
1510 *
1511 * If we are dumping also try to process any pending interrupts. This may
1512 * or may not work depending on the state of the cpu at the point it was
1513 * stopped.
1514 */
1515void
1516lwkt_smp_stopped(void)
1517{
1518 globaldata_t gd = mycpu;
1519
1520 crit_enter_gd(gd);
1521 if (dumping) {
1522 lwkt_process_ipiq();
1523 splz();
1524 } else {
1525 lwkt_process_ipiq();
1526 }
1527 crit_exit_gd(gd);
1528}
1529
57aa743c
MD
1530/*
1531 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1532 * get_mplock() has already incremented td_mpcount. We must block and
1533 * not return until giant is held.
1534 *
1535 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1536 * reschedule the thread until it can obtain the giant lock for it.
1537 */
1538void
1539lwkt_mp_lock_contested(void)
1540{
3824f392 1541 ++mplock_countx;
57aa743c 1542 loggiant(beg);
57aa743c 1543 lwkt_switch();
57aa743c 1544 loggiant(end);
57aa743c
MD
1545}
1546
b9eb1c19
MD
1547/*
1548 * The rel_mplock() code will call this function after releasing the
1549 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1550 *
1551 * We then chain an IPI to a single other cpu potentially needing the
1552 * lock. This is a bit heuristical and we can wind up with IPIs flying
1553 * all over the place.
1554 */
1555static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1556
1557void
1558lwkt_mp_lock_uncontested(void)
1559{
1560 globaldata_t gd;
1561 globaldata_t dgd;
1562 cpumask_t mask;
1563 cpumask_t tmpmask;
1564 int cpuid;
1565
1566 if (chain_mplock) {
1567 gd = mycpu;
1568 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1569 mask = mp_lock_contention_mask;
1570 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1571
1572 if (mask) {
1573 if (mask & tmpmask)
1574 cpuid = bsfl(mask & tmpmask);
1575 else
1576 cpuid = bsfl(mask);
1577 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1578 dgd = globaldata_find(cpuid);
1579 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1580 }
1581 }
1582}
1583
1584/*
1585 * The idea is for this IPI to interrupt a potentially lower priority
1586 * thread, such as a user thread, to allow the scheduler to reschedule
1587 * a higher priority kernel thread that needs the MP lock.
1588 *
1589 * For now we set the LWKT reschedule flag which generates an AST in
1590 * doreti, though theoretically it is also possible to possibly preempt
1591 * here if the underlying thread was operating in user mode. Nah.
1592 */
1593static void
1594lwkt_mp_lock_uncontested_remote(void *arg __unused)
1595{
1596 need_lwkt_resched();
1597}
1598
d165e668 1599#endif