Adjust the KTR code to use cache-aligned per-cpu structures.
[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
7d0bac62 76static int untimely_switch = 0;
0f7a3396
MD
77#ifdef INVARIANTS
78static int panic_on_cscount = 0;
79#endif
05220613
MD
80static __int64_t switch_count = 0;
81static __int64_t preempt_hit = 0;
82static __int64_t preempt_miss = 0;
83static __int64_t preempt_weird = 0;
38717797
HP
84static __int64_t token_contention_count = 0;
85static __int64_t mplock_contention_count = 0;
fb0f29c4 86static int lwkt_use_spin_port;
d2f86ad2 87#ifdef SMP
b9eb1c19 88static int chain_mplock = 0;
d2f86ad2 89#endif
40aaf5fc 90static struct objcache *thread_cache;
05220613 91
b9eb1c19
MD
92volatile cpumask_t mp_lock_contention_mask;
93
0855a2af
JG
94extern void cpu_heavy_restore(void);
95extern void cpu_lwkt_restore(void);
96extern void cpu_kthread_restore(void);
97extern void cpu_idle_restore(void);
98
99int
100jg_tos_ok(struct thread *td)
101{
102 if (td == NULL) {
103 return 1;
104 }
105 KKASSERT(td->td_sp != NULL);
106 unsigned long tos = ((unsigned long *)td->td_sp)[0];
107 int tos_ok = 0;
108 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore)
109 || (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
110 tos_ok = 1;
111 }
112 return tos_ok;
113}
114
fb0f29c4
MD
115/*
116 * We can make all thread ports use the spin backend instead of the thread
117 * backend. This should only be set to debug the spin backend.
118 */
119TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
120
05220613 121SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
122#ifdef INVARIANTS
123SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
124#endif
b9eb1c19
MD
125#ifdef SMP
126SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
127#endif
4b5f931b 128SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 129SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 130SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 131SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
132#ifdef INVARIANTS
133SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
134 &token_contention_count, 0, "spinning due to token contention");
135SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
136 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
137#endif
05220613 138
57aa743c
MD
139/*
140 * Kernel Trace
141 */
57aa743c
MD
142#if !defined(KTR_GIANT_CONTENTION)
143#define KTR_GIANT_CONTENTION KTR_ALL
144#endif
145
146KTR_INFO_MASTER(giant);
147KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
148KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
149
150#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
151
4b5f931b
MD
152/*
153 * These helper procedures handle the runq, they can only be called from
154 * within a critical section.
75cdbe6c
MD
155 *
156 * WARNING! Prior to SMP being brought up it is possible to enqueue and
157 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
158 * instead of 'mycpu' when referencing the globaldata structure. Once
159 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 160 */
f1d1c3fa
MD
161static __inline
162void
163_lwkt_dequeue(thread_t td)
164{
165 if (td->td_flags & TDF_RUNQ) {
4b5f931b 166 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 167 struct globaldata *gd = td->td_gd;
4b5f931b 168
f1d1c3fa 169 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
170 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
171 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
172 }
173}
174
175static __inline
176void
177_lwkt_enqueue(thread_t td)
178{
344ad853 179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 180 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 181 struct globaldata *gd = td->td_gd;
4b5f931b 182
f1d1c3fa 183 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
184 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
185 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
186 }
187}
8ad65e08 188
40aaf5fc
NT
189static __boolean_t
190_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
191{
192 struct thread *td = (struct thread *)obj;
193
194 td->td_kstack = NULL;
195 td->td_kstack_size = 0;
196 td->td_flags = TDF_ALLOCATED_THREAD;
197 return (1);
198}
199
200static void
201_lwkt_thread_dtor(void *obj, void *privdata)
202{
203 struct thread *td = (struct thread *)obj;
204
205 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
206 ("_lwkt_thread_dtor: not allocated from objcache"));
207 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
208 td->td_kstack_size > 0,
209 ("_lwkt_thread_dtor: corrupted stack"));
210 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
211}
212
213/*
214 * Initialize the lwkt s/system.
215 */
216void
217lwkt_init(void)
218{
219 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
220 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
221 NULL, CACHE_NTHREADS/2,
222 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
223}
224
37af14fe
MD
225/*
226 * Schedule a thread to run. As the current thread we can always safely
227 * schedule ourselves, and a shortcut procedure is provided for that
228 * function.
229 *
230 * (non-blocking, self contained on a per cpu basis)
231 */
232void
233lwkt_schedule_self(thread_t td)
234{
235 crit_enter_quick(td);
37af14fe 236 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 237 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 238 _lwkt_enqueue(td);
37af14fe
MD
239 crit_exit_quick(td);
240}
241
242/*
243 * Deschedule a thread.
244 *
245 * (non-blocking, self contained on a per cpu basis)
246 */
247void
248lwkt_deschedule_self(thread_t td)
249{
250 crit_enter_quick(td);
37af14fe
MD
251 _lwkt_dequeue(td);
252 crit_exit_quick(td);
253}
254
8ad65e08
MD
255/*
256 * LWKTs operate on a per-cpu basis
257 *
73e4f7b9 258 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
259 */
260void
261lwkt_gdinit(struct globaldata *gd)
262{
4b5f931b
MD
263 int i;
264
265 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
266 TAILQ_INIT(&gd->gd_tdrunq[i]);
267 gd->gd_runqmask = 0;
73e4f7b9 268 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
269}
270
7d0bac62
MD
271/*
272 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
273 * or LWKT start address before it can be scheduled. If the target cpu is
274 * -1 the thread will be created on the current cpu.
0cfcada1
MD
275 *
276 * If you intend to create a thread without a process context this function
277 * does everything except load the startup and switcher function.
7d0bac62
MD
278 */
279thread_t
d3d32139 280lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 281{
c070746a 282 globaldata_t gd = mycpu;
99df837e 283 void *stack;
7d0bac62 284
c070746a
MD
285 /*
286 * If static thread storage is not supplied allocate a thread. Reuse
287 * a cached free thread if possible. gd_freetd is used to keep an exiting
288 * thread intact through the exit.
289 */
ef0fdad1 290 if (td == NULL) {
c070746a
MD
291 if ((td = gd->gd_freetd) != NULL)
292 gd->gd_freetd = NULL;
293 else
294 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
295 KASSERT((td->td_flags &
296 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
297 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
298 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 299 }
c070746a
MD
300
301 /*
302 * Try to reuse cached stack.
303 */
f470d0c8
MD
304 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
305 if (flags & TDF_ALLOCATED_STACK) {
e4846942 306 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
307 stack = NULL;
308 }
309 }
310 if (stack == NULL) {
e4846942 311 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 312 flags |= TDF_ALLOCATED_STACK;
99df837e 313 }
75cdbe6c 314 if (cpu < 0)
c070746a 315 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 316 else
f470d0c8 317 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 318 return(td);
7d0bac62
MD
319}
320
321/*
322 * Initialize a preexisting thread structure. This function is used by
323 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
324 *
f8c3996b
MD
325 * All threads start out in a critical section at a priority of
326 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
327 * appropriate. This function may send an IPI message when the
328 * requested cpu is not the current cpu and consequently gd_tdallq may
329 * not be initialized synchronously from the point of view of the originating
330 * cpu.
331 *
332 * NOTE! we have to be careful in regards to creating threads for other cpus
333 * if SMP has not yet been activated.
7d0bac62 334 */
41a01a4d
MD
335#ifdef SMP
336
75cdbe6c
MD
337static void
338lwkt_init_thread_remote(void *arg)
339{
340 thread_t td = arg;
341
52eedfb5
MD
342 /*
343 * Protected by critical section held by IPI dispatch
344 */
75cdbe6c
MD
345 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
346}
347
41a01a4d
MD
348#endif
349
7d0bac62 350void
f470d0c8
MD
351lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
352 struct globaldata *gd)
7d0bac62 353{
37af14fe
MD
354 globaldata_t mygd = mycpu;
355
99df837e
MD
356 bzero(td, sizeof(struct thread));
357 td->td_kstack = stack;
f470d0c8 358 td->td_kstack_size = stksize;
d3d32139 359 td->td_flags = flags;
26a0694b 360 td->td_gd = gd;
f8c3996b 361 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
362#ifdef SMP
363 if ((flags & TDF_MPSAFE) == 0)
364 td->td_mpcount = 1;
365#endif
fb0f29c4
MD
366 if (lwkt_use_spin_port)
367 lwkt_initport_spin(&td->td_msgport);
368 else
369 lwkt_initport_thread(&td->td_msgport, td);
99df837e 370 pmap_init_thread(td);
0f7a3396 371#ifdef SMP
5d21b981
MD
372 /*
373 * Normally initializing a thread for a remote cpu requires sending an
374 * IPI. However, the idlethread is setup before the other cpus are
375 * activated so we have to treat it as a special case. XXX manipulation
376 * of gd_tdallq requires the BGL.
377 */
378 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 379 crit_enter_gd(mygd);
75cdbe6c 380 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 381 crit_exit_gd(mygd);
75cdbe6c 382 } else {
2db3b277 383 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 384 }
0f7a3396 385#else
37af14fe 386 crit_enter_gd(mygd);
0f7a3396 387 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 388 crit_exit_gd(mygd);
0f7a3396 389#endif
73e4f7b9
MD
390}
391
392void
393lwkt_set_comm(thread_t td, const char *ctl, ...)
394{
e2565a42 395 __va_list va;
73e4f7b9 396
e2565a42 397 __va_start(va, ctl);
379210cb 398 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 399 __va_end(va);
7d0bac62
MD
400}
401
99df837e 402void
73e4f7b9 403lwkt_hold(thread_t td)
99df837e 404{
73e4f7b9
MD
405 ++td->td_refs;
406}
407
408void
409lwkt_rele(thread_t td)
410{
411 KKASSERT(td->td_refs > 0);
412 --td->td_refs;
413}
414
415void
416lwkt_wait_free(thread_t td)
417{
418 while (td->td_refs)
377d4740 419 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
420}
421
422void
423lwkt_free_thread(thread_t td)
424{
d9eea1a5 425 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
426 ("lwkt_free_thread: did not exit! %p", td));
427
40aaf5fc
NT
428 if (td->td_flags & TDF_ALLOCATED_THREAD) {
429 objcache_put(thread_cache, td);
430 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
431 /* client-allocated struct with internally allocated stack */
432 KASSERT(td->td_kstack && td->td_kstack_size > 0,
433 ("lwkt_free_thread: corrupted stack"));
434 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
435 td->td_kstack = NULL;
436 td->td_kstack_size = 0;
99df837e
MD
437 }
438}
439
440
8ad65e08
MD
441/*
442 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
443 * switch to the idlethread. Switching must occur within a critical
444 * section to avoid races with the scheduling queue.
445 *
446 * We always have full control over our cpu's run queue. Other cpus
447 * that wish to manipulate our queue must use the cpu_*msg() calls to
448 * talk to our cpu, so a critical section is all that is needed and
449 * the result is very, very fast thread switching.
450 *
96728c05
MD
451 * The LWKT scheduler uses a fixed priority model and round-robins at
452 * each priority level. User process scheduling is a totally
453 * different beast and LWKT priorities should not be confused with
454 * user process priorities.
f1d1c3fa 455 *
96728c05
MD
456 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
457 * cleans it up. Note that the td_switch() function cannot do anything that
458 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
459 * the target thread (not the current thread). It's nice to have a scheduler
460 * that does not need the MP lock to work because it allows us to do some
461 * really cool high-performance MP lock optimizations.
69d78e99
MD
462 *
463 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
464 * is not called by the current thread in the preemption case, only when
465 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
466 */
467void
468lwkt_switch(void)
469{
37af14fe
MD
470 globaldata_t gd = mycpu;
471 thread_t td = gd->gd_curthread;
8ad65e08 472 thread_t ntd;
8a8d5d85
MD
473#ifdef SMP
474 int mpheld;
475#endif
8ad65e08 476
46a3f46d 477 /*
27e88a6e
MD
478 * Switching from within a 'fast' (non thread switched) interrupt or IPI
479 * is illegal. However, we may have to do it anyway if we hit a fatal
480 * kernel trap or we have paniced.
481 *
482 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 483 */
27e88a6e
MD
484 if (gd->gd_intr_nesting_level) {
485 int savegdnest;
486 int savegdtrap;
487
488 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
489 panic("lwkt_switch: cannot switch from within "
490 "a fast interrupt, yet, td %p\n", td);
491 } else {
492 savegdnest = gd->gd_intr_nesting_level;
493 savegdtrap = gd->gd_trap_nesting_level;
494 gd->gd_intr_nesting_level = 0;
495 gd->gd_trap_nesting_level = 0;
a7422615
MD
496 if ((td->td_flags & TDF_PANICWARN) == 0) {
497 td->td_flags |= TDF_PANICWARN;
6ea70f76 498 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 499 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 500 print_backtrace();
a7422615 501 }
27e88a6e
MD
502 lwkt_switch();
503 gd->gd_intr_nesting_level = savegdnest;
504 gd->gd_trap_nesting_level = savegdtrap;
505 return;
506 }
96728c05 507 }
ef0fdad1 508
cb973d15
MD
509 /*
510 * Passive release (used to transition from user to kernel mode
511 * when we block or switch rather then when we enter the kernel).
512 * This function is NOT called if we are switching into a preemption
513 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
514 * our current process designation (if we have one) and become a true
515 * LWKT thread, and may also hand the current process designation to
516 * another process and schedule thread.
cb973d15
MD
517 */
518 if (td->td_release)
519 td->td_release(td);
520
37af14fe 521 crit_enter_gd(gd);
9d265729
MD
522 if (td->td_toks)
523 lwkt_relalltokens(td);
524
525 /*
b02926de
MD
526 * We had better not be holding any spin locks, but don't get into an
527 * endless panic loop.
9d265729 528 */
bbb31c5d
MD
529 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
530 ("lwkt_switch: still holding a shared spinlock %p!",
531 gd->gd_spinlock_rd));
d666840a
MD
532 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
533 ("lwkt_switch: still holding %d exclusive spinlocks!",
534 gd->gd_spinlocks_wr));
9d265729 535
8a8d5d85
MD
536
537#ifdef SMP
538 /*
539 * td_mpcount cannot be used to determine if we currently hold the
540 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
541 * to get the lock, and switch out if it can't. Our ownership of
542 * the actual lock will remain stable while we are in a critical section
543 * (but, of course, another cpu may own or release the lock so the
544 * actual value of mp_lock is not stable).
8a8d5d85
MD
545 */
546 mpheld = MP_LOCK_HELD();
0f7a3396
MD
547#ifdef INVARIANTS
548 if (td->td_cscount) {
6ea70f76 549 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
550 td);
551 if (panic_on_cscount)
552 panic("switching while mastering cpusync");
553 }
554#endif
8a8d5d85 555#endif
99df837e
MD
556 if ((ntd = td->td_preempted) != NULL) {
557 /*
558 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
559 * thread. This occurs transparently, whether the preempted thread
560 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
561 * itself).
562 *
563 * We have to setup the MP lock for the original thread after backing
564 * out the adjustment that was made to curthread when the original
565 * was preempted.
99df837e 566 */
26a0694b 567 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 568#ifdef SMP
96728c05 569 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 570 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
571 td, ntd, td->td_mpcount, ntd->td_mpcount);
572 }
8a8d5d85
MD
573 if (ntd->td_mpcount) {
574 td->td_mpcount -= ntd->td_mpcount;
575 KKASSERT(td->td_mpcount >= 0);
576 }
577#endif
26a0694b 578 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
579
580 /*
b9eb1c19
MD
581 * The interrupt may have woken a thread up, we need to properly
582 * set the reschedule flag if the originally interrupted thread is
583 * at a lower priority.
8ec60c3f
MD
584 */
585 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
586 need_lwkt_resched();
8a8d5d85 587 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 588 } else {
4b5f931b
MD
589 /*
590 * Priority queue / round-robin at each priority. Note that user
591 * processes run at a fixed, low priority and the user process
592 * scheduler deals with interactions between user processes
593 * by scheduling and descheduling them from the LWKT queue as
594 * necessary.
8a8d5d85
MD
595 *
596 * We have to adjust the MP lock for the target thread. If we
597 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
598 * thread that does not need the MP lock. If we cannot, we spin
599 * instead of HLT.
600 *
601 * A similar issue exists for the tokens held by the target thread.
602 * If we cannot obtain ownership of the tokens we cannot immediately
603 * schedule the thread.
604 */
605
8ec60c3f
MD
606 /*
607 * If an LWKT reschedule was requested, well that is what we are
608 * doing now so clear it.
609 */
610 clear_lwkt_resched();
4b5f931b
MD
611again:
612 if (gd->gd_runqmask) {
613 int nq = bsrl(gd->gd_runqmask);
614 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
615 gd->gd_runqmask &= ~(1 << nq);
616 goto again;
617 }
8a8d5d85 618#ifdef SMP
41a01a4d 619 /*
df6b8ba0
MD
620 * THREAD SELECTION FOR AN SMP MACHINE BUILD
621 *
41a01a4d
MD
622 * If the target needs the MP lock and we couldn't get it,
623 * or if the target is holding tokens and we could not
624 * gain ownership of the tokens, continue looking for a
625 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
626 *
627 * NOTE: the mpheld variable invalid after this conditional, it
628 * can change due to both cpu_try_mplock() returning success
9d265729 629 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
630 * we are trying to check the mpcount of a thread other then
631 * the current thread. Because of this, if the current thread
632 * is not holding td_mpcount, an IPI indirectly run via
9d265729 633 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 634 * cause the core MP lock to be released.
41a01a4d
MD
635 */
636 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 637 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 638 ) {
8a8d5d85 639 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
640
641 mpheld = MP_LOCK_HELD();
642 ntd = NULL;
8a8d5d85
MD
643 while (rqmask) {
644 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 645 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 646 /* spinning due to MP lock being held */
38717797 647#ifdef INVARIANTS
a453459d 648 ++mplock_contention_count;
38717797 649#endif
a453459d 650 /* mplock still not held, 'mpheld' still valid */
41a01a4d 651 continue;
38717797 652 }
a453459d
MD
653
654 /*
9d265729 655 * mpheld state invalid after getalltokens call returns
a453459d
MD
656 * failure, but the variable is only needed for
657 * the loop.
658 */
9d265729 659 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 660 /* spinning due to token contention */
38717797 661#ifdef INVARIANTS
a453459d 662 ++token_contention_count;
38717797 663#endif
a453459d 664 mpheld = MP_LOCK_HELD();
41a01a4d 665 continue;
38717797 666 }
41a01a4d 667 break;
8a8d5d85
MD
668 }
669 if (ntd)
670 break;
671 rqmask &= ~(1 << nq);
672 nq = bsrl(rqmask);
b9eb1c19
MD
673
674 /*
675 * We have two choices. We can either refuse to run a
676 * user thread when a kernel thread needs the MP lock
677 * but could not get it, or we can allow it to run but
678 * then expect an IPI (hopefully) later on to force a
679 * reschedule when the MP lock might become available.
680 */
681 if (nq < TDPRI_KERN_LPSCHED) {
682 if (chain_mplock == 0)
683 break;
684 atomic_set_int(&mp_lock_contention_mask,
685 gd->gd_cpumask);
686 /* continue loop, allow user threads to be scheduled */
687 }
8a8d5d85
MD
688 }
689 if (ntd == NULL) {
b402c633 690 cpu_mplock_contested();
a2a5ad0d
MD
691 ntd = &gd->gd_idlethread;
692 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 693 goto using_idle_thread;
8a8d5d85 694 } else {
344ad853 695 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
696 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
697 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
698 }
699 } else {
344ad853 700 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
701 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
702 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 }
704#else
df6b8ba0
MD
705 /*
706 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
707 * worry about tokens or the BGL. However, we still have
708 * to call lwkt_getalltokens() in order to properly detect
709 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 710 */
7eb611ef 711 lwkt_getalltokens(ntd);
344ad853 712 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
713 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
714 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 715#endif
4b5f931b 716 } else {
3c23a41a 717 /*
60f945af
MD
718 * We have nothing to run but only let the idle loop halt
719 * the cpu if there are no pending interrupts.
3c23a41a 720 */
a2a5ad0d 721 ntd = &gd->gd_idlethread;
60f945af 722 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 723 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 724#ifdef SMP
df6b8ba0
MD
725using_idle_thread:
726 /*
727 * The idle thread should not be holding the MP lock unless we
728 * are trapping in the kernel or in a panic. Since we select the
729 * idle thread unconditionally when no other thread is available,
730 * if the MP lock is desired during a panic or kernel trap, we
731 * have to loop in the scheduler until we get it.
732 */
733 if (ntd->td_mpcount) {
734 mpheld = MP_LOCK_HELD();
b402c633 735 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 736 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
737 } else if (mpheld == 0) {
738 cpu_mplock_contested();
df6b8ba0 739 goto again;
b402c633 740 }
df6b8ba0 741 }
a453459d 742#endif
4b5f931b 743 }
f1d1c3fa 744 }
26a0694b
MD
745 KASSERT(ntd->td_pri >= TDPRI_CRIT,
746 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
747
748 /*
749 * Do the actual switch. If the new target does not need the MP lock
750 * and we are holding it, release the MP lock. If the new target requires
751 * the MP lock we have already acquired it for the target.
752 */
753#ifdef SMP
754 if (ntd->td_mpcount == 0 ) {
755 if (MP_LOCK_HELD())
756 cpu_rel_mplock();
757 } else {
a453459d 758 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
759 }
760#endif
94f6d86e
MD
761 if (td != ntd) {
762 ++switch_count;
0855a2af 763 KKASSERT(jg_tos_ok(ntd));
f1d1c3fa 764 td->td_switch(ntd);
94f6d86e 765 }
37af14fe
MD
766 /* NOTE: current cpu may have changed after switch */
767 crit_exit_quick(td);
8ad65e08
MD
768}
769
b68b7282 770/*
96728c05
MD
771 * Request that the target thread preempt the current thread. Preemption
772 * only works under a specific set of conditions:
b68b7282 773 *
96728c05
MD
774 * - We are not preempting ourselves
775 * - The target thread is owned by the current cpu
776 * - We are not currently being preempted
777 * - The target is not currently being preempted
d3d1cbc8
MD
778 * - We are not holding any spin locks
779 * - The target thread is not holding any tokens
96728c05
MD
780 * - We are able to satisfy the target's MP lock requirements (if any).
781 *
782 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
783 * this is called via lwkt_schedule() through the td_preemptable callback.
784 * critpri is the managed critical priority that we should ignore in order
785 * to determine whether preemption is possible (aka usually just the crit
786 * priority of lwkt_schedule() itself).
b68b7282 787 *
26a0694b
MD
788 * XXX at the moment we run the target thread in a critical section during
789 * the preemption in order to prevent the target from taking interrupts
790 * that *WE* can't. Preemption is strictly limited to interrupt threads
791 * and interrupt-like threads, outside of a critical section, and the
792 * preempted source thread will be resumed the instant the target blocks
793 * whether or not the source is scheduled (i.e. preemption is supposed to
794 * be as transparent as possible).
4b5f931b 795 *
8a8d5d85
MD
796 * The target thread inherits our MP count (added to its own) for the
797 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
798 * MP lock during the preemption. Therefore, any preempting targets must be
799 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
800 * out of sync with the physical mp_lock, but we do not have to preserve
801 * the original ownership of the lock if it was out of synch (that is, we
802 * can leave it synchronized on return).
b68b7282
MD
803 */
804void
96728c05 805lwkt_preempt(thread_t ntd, int critpri)
b68b7282 806{
46a3f46d 807 struct globaldata *gd = mycpu;
0a3f9b47 808 thread_t td;
8a8d5d85
MD
809#ifdef SMP
810 int mpheld;
57c254db 811 int savecnt;
8a8d5d85 812#endif
b68b7282 813
26a0694b 814 /*
96728c05
MD
815 * The caller has put us in a critical section. We can only preempt
816 * if the caller of the caller was not in a critical section (basically
d666840a 817 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 818 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 819 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
820 *
821 * YYY The target thread must be in a critical section (else it must
822 * inherit our critical section? I dunno yet).
41a01a4d 823 *
0a3f9b47 824 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
825 */
826 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 827
0a3f9b47 828 td = gd->gd_curthread;
0a3f9b47 829 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
830 ++preempt_miss;
831 return;
832 }
96728c05
MD
833 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
834 ++preempt_miss;
8ec60c3f 835 need_lwkt_resched();
96728c05
MD
836 return;
837 }
838#ifdef SMP
46a3f46d 839 if (ntd->td_gd != gd) {
96728c05 840 ++preempt_miss;
8ec60c3f 841 need_lwkt_resched();
96728c05
MD
842 return;
843 }
844#endif
41a01a4d 845 /*
d3d1cbc8 846 * Take the easy way out and do not preempt if we are holding
d666840a 847 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
848 * preempted interlock against the target thread's tokens and whether
849 * we can get all the target thread's tokens, but this situation
850 * should not occur very often so its easier to simply not preempt.
d666840a
MD
851 * Also, plain spinlocks are impossible to figure out at this point so
852 * just don't preempt.
d3d1cbc8
MD
853 *
854 * Do not try to preempt if the target thread is holding any tokens.
855 * We could try to acquire the tokens but this case is so rare there
856 * is no need to support it.
41a01a4d 857 */
bbb31c5d 858 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 859 ++preempt_miss;
8ec60c3f 860 need_lwkt_resched();
41a01a4d
MD
861 return;
862 }
d3d1cbc8
MD
863 if (ntd->td_toks) {
864 ++preempt_miss;
865 need_lwkt_resched();
866 return;
867 }
26a0694b
MD
868 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
869 ++preempt_weird;
8ec60c3f 870 need_lwkt_resched();
26a0694b
MD
871 return;
872 }
873 if (ntd->td_preempted) {
4b5f931b 874 ++preempt_hit;
8ec60c3f 875 need_lwkt_resched();
26a0694b 876 return;
b68b7282 877 }
8a8d5d85 878#ifdef SMP
a2a5ad0d
MD
879 /*
880 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
881 * to or from the MP lock. In this case td_mpcount will be pre-disposed
882 * (non-zero) but not actually synchronized with the actual state of the
883 * lock. We can use it to imply an MP lock requirement for the
884 * preemption but we cannot use it to test whether we hold the MP lock
885 * or not.
a2a5ad0d 886 */
96728c05 887 savecnt = td->td_mpcount;
71ef2f5c 888 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
889 ntd->td_mpcount += td->td_mpcount;
890 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
891 ntd->td_mpcount -= td->td_mpcount;
892 ++preempt_miss;
8ec60c3f 893 need_lwkt_resched();
8a8d5d85
MD
894 return;
895 }
896#endif
26a0694b 897
8ec60c3f
MD
898 /*
899 * Since we are able to preempt the current thread, there is no need to
900 * call need_lwkt_resched().
901 */
26a0694b
MD
902 ++preempt_hit;
903 ntd->td_preempted = td;
904 td->td_flags |= TDF_PREEMPT_LOCK;
905 td->td_switch(ntd);
b9eb1c19 906
26a0694b 907 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
908#ifdef SMP
909 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
910 mpheld = MP_LOCK_HELD();
911 if (mpheld && td->td_mpcount == 0)
96728c05 912 cpu_rel_mplock();
71ef2f5c 913 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
914 panic("lwkt_preempt(): MP lock was not held through");
915#endif
26a0694b
MD
916 ntd->td_preempted = NULL;
917 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
918}
919
f1d1c3fa
MD
920/*
921 * Yield our thread while higher priority threads are pending. This is
922 * typically called when we leave a critical section but it can be safely
923 * called while we are in a critical section.
924 *
925 * This function will not generally yield to equal priority threads but it
926 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 927 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
928 * lwkt_yield_quick().
929 *
235957ed 930 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
931 * came along but was blocked and made pending.
932 *
f1d1c3fa
MD
933 * (self contained on a per cpu basis)
934 */
935void
936lwkt_yield_quick(void)
937{
7966cb69
MD
938 globaldata_t gd = mycpu;
939 thread_t td = gd->gd_curthread;
ef0fdad1 940
a2a5ad0d 941 /*
235957ed 942 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
943 * it with a non-zero cpl then we might not wind up calling splz after
944 * a task switch when the critical section is exited even though the
46a3f46d 945 * new task could accept the interrupt.
a2a5ad0d
MD
946 *
947 * XXX from crit_exit() only called after last crit section is released.
948 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
949 *
950 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
951 * except for this special case, we MUST call splz() here to handle any
952 * pending ints, particularly after we switch, or we might accidently
953 * halt the cpu with interrupts pending.
a2a5ad0d 954 */
46a3f46d 955 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 956 splz();
f1d1c3fa
MD
957
958 /*
959 * YYY enabling will cause wakeup() to task-switch, which really
960 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
961 * preemption and MP without actually doing preemption or MP, because a
962 * lot of code assumes that wakeup() does not block.
f1d1c3fa 963 */
46a3f46d
MD
964 if (untimely_switch && td->td_nest_count == 0 &&
965 gd->gd_intr_nesting_level == 0
966 ) {
37af14fe 967 crit_enter_quick(td);
f1d1c3fa
MD
968 /*
969 * YYY temporary hacks until we disassociate the userland scheduler
970 * from the LWKT scheduler.
971 */
972 if (td->td_flags & TDF_RUNQ) {
973 lwkt_switch(); /* will not reenter yield function */
974 } else {
37af14fe 975 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 976 lwkt_switch(); /* will not reenter yield function */
37af14fe 977 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 978 }
7966cb69 979 crit_exit_noyield(td);
f1d1c3fa 980 }
f1d1c3fa
MD
981}
982
8ad65e08 983/*
f1d1c3fa 984 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 985 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
986 * the crit_exit() call in lwkt_switch().
987 *
988 * (self contained on a per cpu basis)
8ad65e08
MD
989 */
990void
f1d1c3fa 991lwkt_yield(void)
8ad65e08 992{
37af14fe 993 lwkt_schedule_self(curthread);
f1d1c3fa
MD
994 lwkt_switch();
995}
996
b9eb1c19
MD
997/*
998 * Return 0 if no runnable threads are pending at the same or higher
999 * priority as the passed thread.
1000 *
1001 * Return 1 if runnable threads are pending at the same priority.
1002 *
1003 * Return 2 if runnable threads are pending at a higher priority.
1004 */
1005int
1006lwkt_check_resched(thread_t td)
1007{
1008 int pri = td->td_pri & TDPRI_MASK;
1009
1010 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1011 return(2);
1012 if (TAILQ_NEXT(td, td_threadq))
1013 return(1);
1014 return(0);
1015}
1016
8ad65e08 1017/*
f1d1c3fa
MD
1018 * Generic schedule. Possibly schedule threads belonging to other cpus and
1019 * deal with threads that might be blocked on a wait queue.
1020 *
0a3f9b47
MD
1021 * We have a little helper inline function which does additional work after
1022 * the thread has been enqueued, including dealing with preemption and
1023 * setting need_lwkt_resched() (which prevents the kernel from returning
1024 * to userland until it has processed higher priority threads).
6330a558
MD
1025 *
1026 * It is possible for this routine to be called after a failed _enqueue
1027 * (due to the target thread migrating, sleeping, or otherwise blocked).
1028 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1029 *
1030 * reschedok is an optimized constant propagated from lwkt_schedule() or
1031 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1032 * reschedule to be requested if the target thread has a higher priority.
1033 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1034 * be 0, prevented undesired reschedules.
8ad65e08 1035 */
0a3f9b47
MD
1036static __inline
1037void
361d01dd 1038_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1039{
b9eb1c19 1040 thread_t otd;
c730be20 1041
6330a558 1042 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1043 if (ntd->td_preemptable && reschedok) {
6330a558 1044 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1045 } else if (reschedok) {
b9eb1c19
MD
1046 otd = curthread;
1047 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1048 need_lwkt_resched();
6330a558 1049 }
0a3f9b47
MD
1050 }
1051}
1052
361d01dd 1053static __inline
8ad65e08 1054void
361d01dd 1055_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1056{
37af14fe
MD
1057 globaldata_t mygd = mycpu;
1058
41a01a4d 1059 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1060 crit_enter_gd(mygd);
9388413d 1061 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1062 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1063 _lwkt_enqueue(td);
1064 } else {
f1d1c3fa 1065 /*
7cd8d145
MD
1066 * If we own the thread, there is no race (since we are in a
1067 * critical section). If we do not own the thread there might
1068 * be a race but the target cpu will deal with it.
f1d1c3fa 1069 */
0f7a3396 1070#ifdef SMP
7cd8d145 1071 if (td->td_gd == mygd) {
9d265729 1072 _lwkt_enqueue(td);
361d01dd 1073 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1074 } else {
7cd8d145
MD
1075 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1076 }
0f7a3396 1077#else
7cd8d145 1078 _lwkt_enqueue(td);
361d01dd 1079 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1080#endif
8ad65e08 1081 }
37af14fe 1082 crit_exit_gd(mygd);
8ad65e08
MD
1083}
1084
361d01dd
MD
1085void
1086lwkt_schedule(thread_t td)
1087{
1088 _lwkt_schedule(td, 1);
1089}
1090
1091void
1092lwkt_schedule_noresched(thread_t td)
1093{
1094 _lwkt_schedule(td, 0);
1095}
1096
52eedfb5
MD
1097#ifdef SMP
1098
d9eea1a5 1099/*
52eedfb5
MD
1100 * Thread migration using a 'Pull' method. The thread may or may not be
1101 * the current thread. It MUST be descheduled and in a stable state.
1102 * lwkt_giveaway() must be called on the cpu owning the thread.
1103 *
1104 * At any point after lwkt_giveaway() is called, the target cpu may
1105 * 'pull' the thread by calling lwkt_acquire().
1106 *
1107 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1108 */
52eedfb5
MD
1109void
1110lwkt_giveaway(thread_t td)
1111{
1112 globaldata_t gd = mycpu;
1113
1114 crit_enter_gd(gd);
1115 KKASSERT(td->td_gd == gd);
1116 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1117 td->td_flags |= TDF_MIGRATING;
1118 crit_exit_gd(gd);
1119}
1120
a2a5ad0d
MD
1121void
1122lwkt_acquire(thread_t td)
1123{
37af14fe
MD
1124 globaldata_t gd;
1125 globaldata_t mygd;
a2a5ad0d 1126
52eedfb5 1127 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1128 gd = td->td_gd;
37af14fe 1129 mygd = mycpu;
52eedfb5 1130 if (gd != mycpu) {
35238fa5 1131 cpu_lfence();
52eedfb5 1132 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1133 crit_enter_gd(mygd);
df910c23
MD
1134 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1135#ifdef SMP
1136 lwkt_process_ipiq();
1137#endif
52eedfb5 1138 cpu_lfence();
df910c23 1139 }
37af14fe 1140 td->td_gd = mygd;
52eedfb5
MD
1141 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1142 td->td_flags &= ~TDF_MIGRATING;
1143 crit_exit_gd(mygd);
1144 } else {
1145 crit_enter_gd(mygd);
1146 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1147 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1148 crit_exit_gd(mygd);
a2a5ad0d
MD
1149 }
1150}
1151
52eedfb5
MD
1152#endif
1153
f1d1c3fa
MD
1154/*
1155 * Generic deschedule. Descheduling threads other then your own should be
1156 * done only in carefully controlled circumstances. Descheduling is
1157 * asynchronous.
1158 *
1159 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1160 */
1161void
1162lwkt_deschedule(thread_t td)
1163{
f1d1c3fa 1164 crit_enter();
b8a98473 1165#ifdef SMP
f1d1c3fa
MD
1166 if (td == curthread) {
1167 _lwkt_dequeue(td);
1168 } else {
a72187e9 1169 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1170 _lwkt_dequeue(td);
1171 } else {
b8a98473 1172 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1173 }
1174 }
b8a98473
MD
1175#else
1176 _lwkt_dequeue(td);
1177#endif
f1d1c3fa
MD
1178 crit_exit();
1179}
1180
4b5f931b
MD
1181/*
1182 * Set the target thread's priority. This routine does not automatically
1183 * switch to a higher priority thread, LWKT threads are not designed for
1184 * continuous priority changes. Yield if you want to switch.
1185 *
1186 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1187 * of the td_pri field. The specified priority may also indicate zero or
1188 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1189 *
1190 * Note that we requeue the thread whether it winds up on a different runq
1191 * or not. uio_yield() depends on this and the routine is not normally
1192 * called with the same priority otherwise.
4b5f931b
MD
1193 */
1194void
1195lwkt_setpri(thread_t td, int pri)
1196{
26a0694b 1197 KKASSERT(pri >= 0);
a72187e9 1198 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1199 crit_enter();
1200 if (td->td_flags & TDF_RUNQ) {
1201 _lwkt_dequeue(td);
1202 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1203 _lwkt_enqueue(td);
1204 } else {
1205 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1206 }
1207 crit_exit();
1208}
1209
1210void
1211lwkt_setpri_self(int pri)
1212{
1213 thread_t td = curthread;
1214
4b5f931b
MD
1215 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1216 crit_enter();
1217 if (td->td_flags & TDF_RUNQ) {
1218 _lwkt_dequeue(td);
1219 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1220 _lwkt_enqueue(td);
1221 } else {
1222 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1223 }
1224 crit_exit();
1225}
1226
5d21b981 1227/*
52eedfb5
MD
1228 * Migrate the current thread to the specified cpu.
1229 *
1230 * This is accomplished by descheduling ourselves from the current cpu,
1231 * moving our thread to the tdallq of the target cpu, IPI messaging the
1232 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1233 * races while the thread is being migrated.
5d21b981 1234 */
3d28ff59 1235#ifdef SMP
5d21b981 1236static void lwkt_setcpu_remote(void *arg);
3d28ff59 1237#endif
5d21b981
MD
1238
1239void
1240lwkt_setcpu_self(globaldata_t rgd)
1241{
1242#ifdef SMP
1243 thread_t td = curthread;
1244
1245 if (td->td_gd != rgd) {
1246 crit_enter_quick(td);
1247 td->td_flags |= TDF_MIGRATING;
1248 lwkt_deschedule_self(td);
52eedfb5 1249 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1250 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1251 lwkt_switch();
1252 /* we are now on the target cpu */
52eedfb5 1253 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1254 crit_exit_quick(td);
1255 }
1256#endif
1257}
1258
ecdefdda
MD
1259void
1260lwkt_migratecpu(int cpuid)
1261{
1262#ifdef SMP
1263 globaldata_t rgd;
1264
1265 rgd = globaldata_find(cpuid);
1266 lwkt_setcpu_self(rgd);
1267#endif
1268}
1269
5d21b981
MD
1270/*
1271 * Remote IPI for cpu migration (called while in a critical section so we
1272 * do not have to enter another one). The thread has already been moved to
1273 * our cpu's allq, but we must wait for the thread to be completely switched
1274 * out on the originating cpu before we schedule it on ours or the stack
1275 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1276 * change to main memory.
1277 *
1278 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1279 * against wakeups. It is best if this interface is used only when there
1280 * are no pending events that might try to schedule the thread.
1281 */
3d28ff59 1282#ifdef SMP
5d21b981
MD
1283static void
1284lwkt_setcpu_remote(void *arg)
1285{
1286 thread_t td = arg;
1287 globaldata_t gd = mycpu;
1288
df910c23
MD
1289 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1290#ifdef SMP
1291 lwkt_process_ipiq();
1292#endif
35238fa5 1293 cpu_lfence();
df910c23 1294 }
5d21b981 1295 td->td_gd = gd;
35238fa5 1296 cpu_sfence();
5d21b981 1297 td->td_flags &= ~TDF_MIGRATING;
9388413d 1298 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1299 _lwkt_enqueue(td);
1300}
3d28ff59 1301#endif
5d21b981 1302
553ea3c8 1303struct lwp *
4b5f931b
MD
1304lwkt_preempted_proc(void)
1305{
73e4f7b9 1306 thread_t td = curthread;
4b5f931b
MD
1307 while (td->td_preempted)
1308 td = td->td_preempted;
553ea3c8 1309 return(td->td_lwp);
4b5f931b
MD
1310}
1311
99df837e
MD
1312/*
1313 * Create a kernel process/thread/whatever. It shares it's address space
1314 * with proc0 - ie: kernel only.
1315 *
365fa13f
MD
1316 * NOTE! By default new threads are created with the MP lock held. A
1317 * thread which does not require the MP lock should release it by calling
1318 * rel_mplock() at the start of the new thread.
99df837e
MD
1319 */
1320int
1321lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1322 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1323 const char *fmt, ...)
99df837e 1324{
73e4f7b9 1325 thread_t td;
e2565a42 1326 __va_list ap;
99df837e 1327
d3d32139 1328 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1329 tdflags);
a2a5ad0d
MD
1330 if (tdp)
1331 *tdp = td;
709799ea 1332 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1333
1334 /*
1335 * Set up arg0 for 'ps' etc
1336 */
e2565a42 1337 __va_start(ap, fmt);
379210cb 1338 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1339 __va_end(ap);
99df837e
MD
1340
1341 /*
1342 * Schedule the thread to run
1343 */
ef0fdad1
MD
1344 if ((td->td_flags & TDF_STOPREQ) == 0)
1345 lwkt_schedule(td);
1346 else
1347 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1348 return 0;
1349}
1350
1351/*
1352 * Destroy an LWKT thread. Warning! This function is not called when
1353 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1354 * uses a different reaping mechanism.
1355 */
1356void
1357lwkt_exit(void)
1358{
1359 thread_t td = curthread;
c070746a 1360 thread_t std;
8826f33a 1361 globaldata_t gd;
99df837e
MD
1362
1363 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1364 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1365 caps_exit(td);
c070746a
MD
1366
1367 /*
1368 * Get us into a critical section to interlock gd_freetd and loop
1369 * until we can get it freed.
1370 *
1371 * We have to cache the current td in gd_freetd because objcache_put()ing
1372 * it would rip it out from under us while our thread is still active.
1373 */
1374 gd = mycpu;
37af14fe 1375 crit_enter_quick(td);
c070746a
MD
1376 while ((std = gd->gd_freetd) != NULL) {
1377 gd->gd_freetd = NULL;
1378 objcache_put(thread_cache, std);
1379 }
37af14fe 1380 lwkt_deschedule_self(td);
e56e4dea 1381 lwkt_remove_tdallq(td);
c070746a
MD
1382 if (td->td_flags & TDF_ALLOCATED_THREAD)
1383 gd->gd_freetd = td;
99df837e
MD
1384 cpu_thread_exit();
1385}
1386
e56e4dea
MD
1387void
1388lwkt_remove_tdallq(thread_t td)
1389{
1390 KKASSERT(td->td_gd == mycpu);
1391 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1392}
1393
2d93b37a
MD
1394void
1395crit_panic(void)
1396{
1397 thread_t td = curthread;
1398 int lpri = td->td_pri;
1399
1400 td->td_pri = 0;
1401 panic("td_pri is/would-go negative! %p %d", td, lpri);
1402}
1403
d165e668
MD
1404#ifdef SMP
1405
bd8015ca
MD
1406/*
1407 * Called from debugger/panic on cpus which have been stopped. We must still
1408 * process the IPIQ while stopped, even if we were stopped while in a critical
1409 * section (XXX).
1410 *
1411 * If we are dumping also try to process any pending interrupts. This may
1412 * or may not work depending on the state of the cpu at the point it was
1413 * stopped.
1414 */
1415void
1416lwkt_smp_stopped(void)
1417{
1418 globaldata_t gd = mycpu;
1419
1420 crit_enter_gd(gd);
1421 if (dumping) {
1422 lwkt_process_ipiq();
1423 splz();
1424 } else {
1425 lwkt_process_ipiq();
1426 }
1427 crit_exit_gd(gd);
1428}
1429
57aa743c
MD
1430/*
1431 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1432 * get_mplock() has already incremented td_mpcount. We must block and
1433 * not return until giant is held.
1434 *
1435 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1436 * reschedule the thread until it can obtain the giant lock for it.
1437 */
1438void
1439lwkt_mp_lock_contested(void)
1440{
57aa743c 1441 loggiant(beg);
57aa743c 1442 lwkt_switch();
57aa743c 1443 loggiant(end);
57aa743c
MD
1444}
1445
b9eb1c19
MD
1446/*
1447 * The rel_mplock() code will call this function after releasing the
1448 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1449 *
1450 * We then chain an IPI to a single other cpu potentially needing the
1451 * lock. This is a bit heuristical and we can wind up with IPIs flying
1452 * all over the place.
1453 */
1454static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1455
1456void
1457lwkt_mp_lock_uncontested(void)
1458{
1459 globaldata_t gd;
1460 globaldata_t dgd;
1461 cpumask_t mask;
1462 cpumask_t tmpmask;
1463 int cpuid;
1464
1465 if (chain_mplock) {
1466 gd = mycpu;
1467 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1468 mask = mp_lock_contention_mask;
1469 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1470
1471 if (mask) {
1472 if (mask & tmpmask)
1473 cpuid = bsfl(mask & tmpmask);
1474 else
1475 cpuid = bsfl(mask);
1476 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1477 dgd = globaldata_find(cpuid);
1478 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1479 }
1480 }
1481}
1482
1483/*
1484 * The idea is for this IPI to interrupt a potentially lower priority
1485 * thread, such as a user thread, to allow the scheduler to reschedule
1486 * a higher priority kernel thread that needs the MP lock.
1487 *
1488 * For now we set the LWKT reschedule flag which generates an AST in
1489 * doreti, though theoretically it is also possible to possibly preempt
1490 * here if the underlying thread was operating in user mode. Nah.
1491 */
1492static void
1493lwkt_mp_lock_uncontested_remote(void *arg __unused)
1494{
1495 need_lwkt_resched();
1496}
1497
d165e668 1498#endif