MPSAFE - tsleep_interlock, BUF/BIO, cluster, swap_pager.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
8c10bfcf 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
8c10bfcf 33 *
0aa16b5d 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/kernel.h>
47#include <sys/proc.h>
48#include <sys/rtprio.h>
49#include <sys/queue.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f6bf3af1 54#include <sys/caps.h>
9d265729 55#include <sys/spinlock.h>
57aa743c 56#include <sys/ktr.h>
9d265729
MD
57
58#include <sys/thread2.h>
59#include <sys/spinlock2.h>
f1d1c3fa 60
7d0bac62
MD
61#include <vm/vm.h>
62#include <vm/vm_param.h>
63#include <vm/vm_kern.h>
64#include <vm/vm_object.h>
65#include <vm/vm_page.h>
66#include <vm/vm_map.h>
67#include <vm/vm_pager.h>
68#include <vm/vm_extern.h>
7d0bac62 69
99df837e 70#include <machine/stdarg.h>
96728c05 71#include <machine/smp.h>
99df837e 72
1541028a 73
40aaf5fc
NT
74static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
75
3824f392
MD
76#ifdef SMP
77static int mplock_countx = 0;
78#endif
7d0bac62 79static int untimely_switch = 0;
0f7a3396
MD
80#ifdef INVARIANTS
81static int panic_on_cscount = 0;
82#endif
05220613
MD
83static __int64_t switch_count = 0;
84static __int64_t preempt_hit = 0;
85static __int64_t preempt_miss = 0;
86static __int64_t preempt_weird = 0;
38717797
HP
87static __int64_t token_contention_count = 0;
88static __int64_t mplock_contention_count = 0;
fb0f29c4 89static int lwkt_use_spin_port;
d2f86ad2 90#ifdef SMP
b9eb1c19 91static int chain_mplock = 0;
d2f86ad2 92#endif
40aaf5fc 93static struct objcache *thread_cache;
05220613 94
b9eb1c19
MD
95volatile cpumask_t mp_lock_contention_mask;
96
0855a2af
JG
97extern void cpu_heavy_restore(void);
98extern void cpu_lwkt_restore(void);
99extern void cpu_kthread_restore(void);
100extern void cpu_idle_restore(void);
101
85514115
MD
102#ifdef __amd64__
103
104static int
0855a2af
JG
105jg_tos_ok(struct thread *td)
106{
85514115
MD
107 void *tos;
108 int tos_ok;
109
0855a2af
JG
110 if (td == NULL) {
111 return 1;
112 }
113 KKASSERT(td->td_sp != NULL);
85514115
MD
114 tos = ((void **)td->td_sp)[0];
115 tos_ok = 0;
116 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
117 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
118 tos_ok = 1;
119 }
120 return tos_ok;
121}
122
85514115
MD
123#endif
124
fb0f29c4
MD
125/*
126 * We can make all thread ports use the spin backend instead of the thread
127 * backend. This should only be set to debug the spin backend.
128 */
129TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
130
05220613 131SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
132#ifdef INVARIANTS
133SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
134#endif
b9eb1c19
MD
135#ifdef SMP
136SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
137#endif
4b5f931b 138SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 139SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 140SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 141SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
142#ifdef INVARIANTS
143SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
144 &token_contention_count, 0, "spinning due to token contention");
145SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
146 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
147#endif
05220613 148
57aa743c
MD
149/*
150 * Kernel Trace
151 */
57aa743c
MD
152#if !defined(KTR_GIANT_CONTENTION)
153#define KTR_GIANT_CONTENTION KTR_ALL
154#endif
155
156KTR_INFO_MASTER(giant);
157KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
158KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
159
160#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
161
4b5f931b
MD
162/*
163 * These helper procedures handle the runq, they can only be called from
164 * within a critical section.
75cdbe6c
MD
165 *
166 * WARNING! Prior to SMP being brought up it is possible to enqueue and
167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
168 * instead of 'mycpu' when referencing the globaldata structure. Once
169 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 170 */
f1d1c3fa
MD
171static __inline
172void
173_lwkt_dequeue(thread_t td)
174{
175 if (td->td_flags & TDF_RUNQ) {
4b5f931b 176 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 177 struct globaldata *gd = td->td_gd;
4b5f931b 178
f1d1c3fa 179 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
180 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
181 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
182 }
183}
184
185static __inline
186void
187_lwkt_enqueue(thread_t td)
188{
344ad853 189 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 190 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 191 struct globaldata *gd = td->td_gd;
4b5f931b 192
f1d1c3fa 193 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
194 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
195 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
196 }
197}
8ad65e08 198
40aaf5fc
NT
199static __boolean_t
200_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
201{
202 struct thread *td = (struct thread *)obj;
203
204 td->td_kstack = NULL;
205 td->td_kstack_size = 0;
206 td->td_flags = TDF_ALLOCATED_THREAD;
207 return (1);
208}
209
210static void
211_lwkt_thread_dtor(void *obj, void *privdata)
212{
213 struct thread *td = (struct thread *)obj;
214
215 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
216 ("_lwkt_thread_dtor: not allocated from objcache"));
217 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
218 td->td_kstack_size > 0,
219 ("_lwkt_thread_dtor: corrupted stack"));
220 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
221}
222
223/*
224 * Initialize the lwkt s/system.
225 */
226void
227lwkt_init(void)
228{
229 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
230 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
231 NULL, CACHE_NTHREADS/2,
232 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
233}
234
37af14fe
MD
235/*
236 * Schedule a thread to run. As the current thread we can always safely
237 * schedule ourselves, and a shortcut procedure is provided for that
238 * function.
239 *
240 * (non-blocking, self contained on a per cpu basis)
241 */
242void
243lwkt_schedule_self(thread_t td)
244{
245 crit_enter_quick(td);
37af14fe 246 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 247 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 248 _lwkt_enqueue(td);
37af14fe
MD
249 crit_exit_quick(td);
250}
251
252/*
253 * Deschedule a thread.
254 *
255 * (non-blocking, self contained on a per cpu basis)
256 */
257void
258lwkt_deschedule_self(thread_t td)
259{
260 crit_enter_quick(td);
37af14fe
MD
261 _lwkt_dequeue(td);
262 crit_exit_quick(td);
263}
264
8ad65e08
MD
265/*
266 * LWKTs operate on a per-cpu basis
267 *
73e4f7b9 268 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
269 */
270void
271lwkt_gdinit(struct globaldata *gd)
272{
4b5f931b
MD
273 int i;
274
275 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
276 TAILQ_INIT(&gd->gd_tdrunq[i]);
277 gd->gd_runqmask = 0;
73e4f7b9 278 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
279}
280
7d0bac62
MD
281/*
282 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
283 * or LWKT start address before it can be scheduled. If the target cpu is
284 * -1 the thread will be created on the current cpu.
0cfcada1
MD
285 *
286 * If you intend to create a thread without a process context this function
287 * does everything except load the startup and switcher function.
7d0bac62
MD
288 */
289thread_t
d3d32139 290lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 291{
c070746a 292 globaldata_t gd = mycpu;
99df837e 293 void *stack;
7d0bac62 294
c070746a
MD
295 /*
296 * If static thread storage is not supplied allocate a thread. Reuse
297 * a cached free thread if possible. gd_freetd is used to keep an exiting
298 * thread intact through the exit.
299 */
ef0fdad1 300 if (td == NULL) {
c070746a
MD
301 if ((td = gd->gd_freetd) != NULL)
302 gd->gd_freetd = NULL;
303 else
304 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
305 KASSERT((td->td_flags &
306 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
307 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
308 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 309 }
c070746a
MD
310
311 /*
312 * Try to reuse cached stack.
313 */
f470d0c8
MD
314 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
315 if (flags & TDF_ALLOCATED_STACK) {
e4846942 316 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
317 stack = NULL;
318 }
319 }
320 if (stack == NULL) {
e4846942 321 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 322 flags |= TDF_ALLOCATED_STACK;
99df837e 323 }
75cdbe6c 324 if (cpu < 0)
c070746a 325 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 326 else
f470d0c8 327 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 328 return(td);
7d0bac62
MD
329}
330
331/*
332 * Initialize a preexisting thread structure. This function is used by
333 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
334 *
f8c3996b
MD
335 * All threads start out in a critical section at a priority of
336 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
337 * appropriate. This function may send an IPI message when the
338 * requested cpu is not the current cpu and consequently gd_tdallq may
339 * not be initialized synchronously from the point of view of the originating
340 * cpu.
341 *
342 * NOTE! we have to be careful in regards to creating threads for other cpus
343 * if SMP has not yet been activated.
7d0bac62 344 */
41a01a4d
MD
345#ifdef SMP
346
75cdbe6c
MD
347static void
348lwkt_init_thread_remote(void *arg)
349{
350 thread_t td = arg;
351
52eedfb5
MD
352 /*
353 * Protected by critical section held by IPI dispatch
354 */
75cdbe6c
MD
355 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
356}
357
41a01a4d
MD
358#endif
359
7d0bac62 360void
f470d0c8
MD
361lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
362 struct globaldata *gd)
7d0bac62 363{
37af14fe
MD
364 globaldata_t mygd = mycpu;
365
99df837e
MD
366 bzero(td, sizeof(struct thread));
367 td->td_kstack = stack;
f470d0c8 368 td->td_kstack_size = stksize;
d3d32139 369 td->td_flags = flags;
26a0694b 370 td->td_gd = gd;
f8c3996b 371 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
372#ifdef SMP
373 if ((flags & TDF_MPSAFE) == 0)
374 td->td_mpcount = 1;
375#endif
fb0f29c4
MD
376 if (lwkt_use_spin_port)
377 lwkt_initport_spin(&td->td_msgport);
378 else
379 lwkt_initport_thread(&td->td_msgport, td);
99df837e 380 pmap_init_thread(td);
0f7a3396 381#ifdef SMP
5d21b981
MD
382 /*
383 * Normally initializing a thread for a remote cpu requires sending an
384 * IPI. However, the idlethread is setup before the other cpus are
385 * activated so we have to treat it as a special case. XXX manipulation
386 * of gd_tdallq requires the BGL.
387 */
388 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 389 crit_enter_gd(mygd);
75cdbe6c 390 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 391 crit_exit_gd(mygd);
75cdbe6c 392 } else {
2db3b277 393 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 394 }
0f7a3396 395#else
37af14fe 396 crit_enter_gd(mygd);
0f7a3396 397 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 398 crit_exit_gd(mygd);
0f7a3396 399#endif
73e4f7b9
MD
400}
401
402void
403lwkt_set_comm(thread_t td, const char *ctl, ...)
404{
e2565a42 405 __va_list va;
73e4f7b9 406
e2565a42 407 __va_start(va, ctl);
379210cb 408 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 409 __va_end(va);
7d0bac62
MD
410}
411
99df837e 412void
73e4f7b9 413lwkt_hold(thread_t td)
99df837e 414{
73e4f7b9
MD
415 ++td->td_refs;
416}
417
418void
419lwkt_rele(thread_t td)
420{
421 KKASSERT(td->td_refs > 0);
422 --td->td_refs;
423}
424
425void
426lwkt_wait_free(thread_t td)
427{
428 while (td->td_refs)
377d4740 429 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
430}
431
432void
433lwkt_free_thread(thread_t td)
434{
d9eea1a5 435 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
436 ("lwkt_free_thread: did not exit! %p", td));
437
40aaf5fc
NT
438 if (td->td_flags & TDF_ALLOCATED_THREAD) {
439 objcache_put(thread_cache, td);
440 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
441 /* client-allocated struct with internally allocated stack */
442 KASSERT(td->td_kstack && td->td_kstack_size > 0,
443 ("lwkt_free_thread: corrupted stack"));
444 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
445 td->td_kstack = NULL;
446 td->td_kstack_size = 0;
99df837e
MD
447 }
448}
449
450
8ad65e08
MD
451/*
452 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
453 * switch to the idlethread. Switching must occur within a critical
454 * section to avoid races with the scheduling queue.
455 *
456 * We always have full control over our cpu's run queue. Other cpus
457 * that wish to manipulate our queue must use the cpu_*msg() calls to
458 * talk to our cpu, so a critical section is all that is needed and
459 * the result is very, very fast thread switching.
460 *
96728c05
MD
461 * The LWKT scheduler uses a fixed priority model and round-robins at
462 * each priority level. User process scheduling is a totally
463 * different beast and LWKT priorities should not be confused with
464 * user process priorities.
f1d1c3fa 465 *
96728c05
MD
466 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
467 * cleans it up. Note that the td_switch() function cannot do anything that
468 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
469 * the target thread (not the current thread). It's nice to have a scheduler
470 * that does not need the MP lock to work because it allows us to do some
471 * really cool high-performance MP lock optimizations.
69d78e99
MD
472 *
473 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
474 * is not called by the current thread in the preemption case, only when
475 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
476 */
477void
478lwkt_switch(void)
479{
37af14fe
MD
480 globaldata_t gd = mycpu;
481 thread_t td = gd->gd_curthread;
8ad65e08 482 thread_t ntd;
8a8d5d85
MD
483#ifdef SMP
484 int mpheld;
485#endif
8ad65e08 486
46a3f46d 487 /*
27e88a6e
MD
488 * Switching from within a 'fast' (non thread switched) interrupt or IPI
489 * is illegal. However, we may have to do it anyway if we hit a fatal
490 * kernel trap or we have paniced.
491 *
492 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 493 */
27e88a6e
MD
494 if (gd->gd_intr_nesting_level) {
495 int savegdnest;
496 int savegdtrap;
497
498 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
499 panic("lwkt_switch: cannot switch from within "
500 "a fast interrupt, yet, td %p\n", td);
501 } else {
502 savegdnest = gd->gd_intr_nesting_level;
503 savegdtrap = gd->gd_trap_nesting_level;
504 gd->gd_intr_nesting_level = 0;
505 gd->gd_trap_nesting_level = 0;
a7422615
MD
506 if ((td->td_flags & TDF_PANICWARN) == 0) {
507 td->td_flags |= TDF_PANICWARN;
6ea70f76 508 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 509 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 510 print_backtrace();
a7422615 511 }
27e88a6e
MD
512 lwkt_switch();
513 gd->gd_intr_nesting_level = savegdnest;
514 gd->gd_trap_nesting_level = savegdtrap;
515 return;
516 }
96728c05 517 }
ef0fdad1 518
cb973d15
MD
519 /*
520 * Passive release (used to transition from user to kernel mode
521 * when we block or switch rather then when we enter the kernel).
522 * This function is NOT called if we are switching into a preemption
523 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
524 * our current process designation (if we have one) and become a true
525 * LWKT thread, and may also hand the current process designation to
526 * another process and schedule thread.
cb973d15
MD
527 */
528 if (td->td_release)
529 td->td_release(td);
530
37af14fe 531 crit_enter_gd(gd);
9d265729
MD
532 if (td->td_toks)
533 lwkt_relalltokens(td);
534
535 /*
b02926de
MD
536 * We had better not be holding any spin locks, but don't get into an
537 * endless panic loop.
9d265729 538 */
bbb31c5d
MD
539 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
540 ("lwkt_switch: still holding a shared spinlock %p!",
541 gd->gd_spinlock_rd));
d666840a
MD
542 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
543 ("lwkt_switch: still holding %d exclusive spinlocks!",
544 gd->gd_spinlocks_wr));
9d265729 545
8a8d5d85
MD
546
547#ifdef SMP
548 /*
549 * td_mpcount cannot be used to determine if we currently hold the
550 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
551 * to get the lock, and switch out if it can't. Our ownership of
552 * the actual lock will remain stable while we are in a critical section
553 * (but, of course, another cpu may own or release the lock so the
554 * actual value of mp_lock is not stable).
8a8d5d85
MD
555 */
556 mpheld = MP_LOCK_HELD();
0f7a3396
MD
557#ifdef INVARIANTS
558 if (td->td_cscount) {
6ea70f76 559 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
560 td);
561 if (panic_on_cscount)
562 panic("switching while mastering cpusync");
563 }
564#endif
8a8d5d85 565#endif
99df837e
MD
566 if ((ntd = td->td_preempted) != NULL) {
567 /*
568 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
569 * thread. This occurs transparently, whether the preempted thread
570 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
571 * itself).
572 *
573 * We have to setup the MP lock for the original thread after backing
574 * out the adjustment that was made to curthread when the original
575 * was preempted.
99df837e 576 */
26a0694b 577 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 578#ifdef SMP
96728c05 579 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 580 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
581 td, ntd, td->td_mpcount, ntd->td_mpcount);
582 }
8a8d5d85
MD
583 if (ntd->td_mpcount) {
584 td->td_mpcount -= ntd->td_mpcount;
585 KKASSERT(td->td_mpcount >= 0);
586 }
587#endif
26a0694b 588 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
589
590 /*
b9eb1c19
MD
591 * The interrupt may have woken a thread up, we need to properly
592 * set the reschedule flag if the originally interrupted thread is
593 * at a lower priority.
8ec60c3f
MD
594 */
595 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
596 need_lwkt_resched();
8a8d5d85 597 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 598 } else {
4b5f931b
MD
599 /*
600 * Priority queue / round-robin at each priority. Note that user
601 * processes run at a fixed, low priority and the user process
602 * scheduler deals with interactions between user processes
603 * by scheduling and descheduling them from the LWKT queue as
604 * necessary.
8a8d5d85
MD
605 *
606 * We have to adjust the MP lock for the target thread. If we
607 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
608 * thread that does not need the MP lock. If we cannot, we spin
609 * instead of HLT.
610 *
611 * A similar issue exists for the tokens held by the target thread.
612 * If we cannot obtain ownership of the tokens we cannot immediately
613 * schedule the thread.
614 */
615
8ec60c3f
MD
616 /*
617 * If an LWKT reschedule was requested, well that is what we are
618 * doing now so clear it.
619 */
620 clear_lwkt_resched();
4b5f931b
MD
621again:
622 if (gd->gd_runqmask) {
623 int nq = bsrl(gd->gd_runqmask);
624 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
625 gd->gd_runqmask &= ~(1 << nq);
626 goto again;
627 }
8a8d5d85 628#ifdef SMP
41a01a4d 629 /*
df6b8ba0
MD
630 * THREAD SELECTION FOR AN SMP MACHINE BUILD
631 *
41a01a4d
MD
632 * If the target needs the MP lock and we couldn't get it,
633 * or if the target is holding tokens and we could not
634 * gain ownership of the tokens, continue looking for a
635 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
636 *
637 * NOTE: the mpheld variable invalid after this conditional, it
638 * can change due to both cpu_try_mplock() returning success
9d265729 639 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
640 * we are trying to check the mpcount of a thread other then
641 * the current thread. Because of this, if the current thread
642 * is not holding td_mpcount, an IPI indirectly run via
9d265729 643 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 644 * cause the core MP lock to be released.
41a01a4d
MD
645 */
646 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 647 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 648 ) {
8a8d5d85 649 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
650
651 mpheld = MP_LOCK_HELD();
652 ntd = NULL;
8a8d5d85
MD
653 while (rqmask) {
654 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 655 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 656 /* spinning due to MP lock being held */
38717797 657#ifdef INVARIANTS
a453459d 658 ++mplock_contention_count;
38717797 659#endif
a453459d 660 /* mplock still not held, 'mpheld' still valid */
41a01a4d 661 continue;
38717797 662 }
a453459d
MD
663
664 /*
9d265729 665 * mpheld state invalid after getalltokens call returns
a453459d
MD
666 * failure, but the variable is only needed for
667 * the loop.
668 */
9d265729 669 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 670 /* spinning due to token contention */
38717797 671#ifdef INVARIANTS
a453459d 672 ++token_contention_count;
38717797 673#endif
a453459d 674 mpheld = MP_LOCK_HELD();
41a01a4d 675 continue;
38717797 676 }
41a01a4d 677 break;
8a8d5d85
MD
678 }
679 if (ntd)
680 break;
681 rqmask &= ~(1 << nq);
682 nq = bsrl(rqmask);
b9eb1c19
MD
683
684 /*
685 * We have two choices. We can either refuse to run a
686 * user thread when a kernel thread needs the MP lock
687 * but could not get it, or we can allow it to run but
688 * then expect an IPI (hopefully) later on to force a
689 * reschedule when the MP lock might become available.
690 */
691 if (nq < TDPRI_KERN_LPSCHED) {
692 if (chain_mplock == 0)
693 break;
694 atomic_set_int(&mp_lock_contention_mask,
695 gd->gd_cpumask);
696 /* continue loop, allow user threads to be scheduled */
697 }
8a8d5d85
MD
698 }
699 if (ntd == NULL) {
b402c633 700 cpu_mplock_contested();
a2a5ad0d
MD
701 ntd = &gd->gd_idlethread;
702 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 703 goto using_idle_thread;
8a8d5d85 704 } else {
344ad853 705 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
706 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
707 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
708 }
709 } else {
3824f392
MD
710 if (ntd->td_mpcount)
711 ++mplock_countx;
344ad853 712 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
713 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
714 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
715 }
716#else
df6b8ba0
MD
717 /*
718 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
719 * worry about tokens or the BGL. However, we still have
720 * to call lwkt_getalltokens() in order to properly detect
721 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 722 */
7eb611ef 723 lwkt_getalltokens(ntd);
344ad853 724 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
725 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
726 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 727#endif
4b5f931b 728 } else {
3c23a41a 729 /*
60f945af
MD
730 * We have nothing to run but only let the idle loop halt
731 * the cpu if there are no pending interrupts.
3c23a41a 732 */
a2a5ad0d 733 ntd = &gd->gd_idlethread;
60f945af 734 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 735 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 736#ifdef SMP
df6b8ba0
MD
737using_idle_thread:
738 /*
739 * The idle thread should not be holding the MP lock unless we
740 * are trapping in the kernel or in a panic. Since we select the
741 * idle thread unconditionally when no other thread is available,
742 * if the MP lock is desired during a panic or kernel trap, we
743 * have to loop in the scheduler until we get it.
744 */
745 if (ntd->td_mpcount) {
746 mpheld = MP_LOCK_HELD();
b402c633 747 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 748 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
749 } else if (mpheld == 0) {
750 cpu_mplock_contested();
df6b8ba0 751 goto again;
b402c633 752 }
df6b8ba0 753 }
a453459d 754#endif
4b5f931b 755 }
f1d1c3fa 756 }
26a0694b
MD
757 KASSERT(ntd->td_pri >= TDPRI_CRIT,
758 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
759
760 /*
761 * Do the actual switch. If the new target does not need the MP lock
762 * and we are holding it, release the MP lock. If the new target requires
763 * the MP lock we have already acquired it for the target.
764 */
765#ifdef SMP
766 if (ntd->td_mpcount == 0 ) {
767 if (MP_LOCK_HELD())
768 cpu_rel_mplock();
769 } else {
a453459d 770 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
771 }
772#endif
94f6d86e
MD
773 if (td != ntd) {
774 ++switch_count;
85514115 775#ifdef __amd64__
0855a2af 776 KKASSERT(jg_tos_ok(ntd));
85514115 777#endif
f1d1c3fa 778 td->td_switch(ntd);
94f6d86e 779 }
37af14fe
MD
780 /* NOTE: current cpu may have changed after switch */
781 crit_exit_quick(td);
8ad65e08
MD
782}
783
b68b7282 784/*
96728c05
MD
785 * Request that the target thread preempt the current thread. Preemption
786 * only works under a specific set of conditions:
b68b7282 787 *
96728c05
MD
788 * - We are not preempting ourselves
789 * - The target thread is owned by the current cpu
790 * - We are not currently being preempted
791 * - The target is not currently being preempted
d3d1cbc8
MD
792 * - We are not holding any spin locks
793 * - The target thread is not holding any tokens
96728c05
MD
794 * - We are able to satisfy the target's MP lock requirements (if any).
795 *
796 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
797 * this is called via lwkt_schedule() through the td_preemptable callback.
798 * critpri is the managed critical priority that we should ignore in order
799 * to determine whether preemption is possible (aka usually just the crit
800 * priority of lwkt_schedule() itself).
b68b7282 801 *
26a0694b
MD
802 * XXX at the moment we run the target thread in a critical section during
803 * the preemption in order to prevent the target from taking interrupts
804 * that *WE* can't. Preemption is strictly limited to interrupt threads
805 * and interrupt-like threads, outside of a critical section, and the
806 * preempted source thread will be resumed the instant the target blocks
807 * whether or not the source is scheduled (i.e. preemption is supposed to
808 * be as transparent as possible).
4b5f931b 809 *
8a8d5d85
MD
810 * The target thread inherits our MP count (added to its own) for the
811 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
812 * MP lock during the preemption. Therefore, any preempting targets must be
813 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
814 * out of sync with the physical mp_lock, but we do not have to preserve
815 * the original ownership of the lock if it was out of synch (that is, we
816 * can leave it synchronized on return).
b68b7282
MD
817 */
818void
96728c05 819lwkt_preempt(thread_t ntd, int critpri)
b68b7282 820{
46a3f46d 821 struct globaldata *gd = mycpu;
0a3f9b47 822 thread_t td;
8a8d5d85
MD
823#ifdef SMP
824 int mpheld;
57c254db 825 int savecnt;
8a8d5d85 826#endif
b68b7282 827
26a0694b 828 /*
96728c05
MD
829 * The caller has put us in a critical section. We can only preempt
830 * if the caller of the caller was not in a critical section (basically
d666840a 831 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 832 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 833 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
834 *
835 * YYY The target thread must be in a critical section (else it must
836 * inherit our critical section? I dunno yet).
41a01a4d 837 *
0a3f9b47 838 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
839 */
840 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 841
0a3f9b47 842 td = gd->gd_curthread;
0a3f9b47 843 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
844 ++preempt_miss;
845 return;
846 }
96728c05
MD
847 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
848 ++preempt_miss;
8ec60c3f 849 need_lwkt_resched();
96728c05
MD
850 return;
851 }
852#ifdef SMP
46a3f46d 853 if (ntd->td_gd != gd) {
96728c05 854 ++preempt_miss;
8ec60c3f 855 need_lwkt_resched();
96728c05
MD
856 return;
857 }
858#endif
41a01a4d 859 /*
d3d1cbc8 860 * Take the easy way out and do not preempt if we are holding
d666840a 861 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
862 * preempted interlock against the target thread's tokens and whether
863 * we can get all the target thread's tokens, but this situation
864 * should not occur very often so its easier to simply not preempt.
d666840a
MD
865 * Also, plain spinlocks are impossible to figure out at this point so
866 * just don't preempt.
d3d1cbc8
MD
867 *
868 * Do not try to preempt if the target thread is holding any tokens.
869 * We could try to acquire the tokens but this case is so rare there
870 * is no need to support it.
41a01a4d 871 */
bbb31c5d 872 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 873 ++preempt_miss;
8ec60c3f 874 need_lwkt_resched();
41a01a4d
MD
875 return;
876 }
d3d1cbc8
MD
877 if (ntd->td_toks) {
878 ++preempt_miss;
879 need_lwkt_resched();
880 return;
881 }
26a0694b
MD
882 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
883 ++preempt_weird;
8ec60c3f 884 need_lwkt_resched();
26a0694b
MD
885 return;
886 }
887 if (ntd->td_preempted) {
4b5f931b 888 ++preempt_hit;
8ec60c3f 889 need_lwkt_resched();
26a0694b 890 return;
b68b7282 891 }
8a8d5d85 892#ifdef SMP
a2a5ad0d
MD
893 /*
894 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
895 * to or from the MP lock. In this case td_mpcount will be pre-disposed
896 * (non-zero) but not actually synchronized with the actual state of the
897 * lock. We can use it to imply an MP lock requirement for the
898 * preemption but we cannot use it to test whether we hold the MP lock
899 * or not.
a2a5ad0d 900 */
96728c05 901 savecnt = td->td_mpcount;
71ef2f5c 902 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
903 ntd->td_mpcount += td->td_mpcount;
904 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
905 ntd->td_mpcount -= td->td_mpcount;
906 ++preempt_miss;
8ec60c3f 907 need_lwkt_resched();
8a8d5d85
MD
908 return;
909 }
910#endif
26a0694b 911
8ec60c3f
MD
912 /*
913 * Since we are able to preempt the current thread, there is no need to
914 * call need_lwkt_resched().
915 */
26a0694b
MD
916 ++preempt_hit;
917 ntd->td_preempted = td;
918 td->td_flags |= TDF_PREEMPT_LOCK;
919 td->td_switch(ntd);
b9eb1c19 920
26a0694b 921 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
922#ifdef SMP
923 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
924 mpheld = MP_LOCK_HELD();
925 if (mpheld && td->td_mpcount == 0)
96728c05 926 cpu_rel_mplock();
71ef2f5c 927 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
928 panic("lwkt_preempt(): MP lock was not held through");
929#endif
26a0694b
MD
930 ntd->td_preempted = NULL;
931 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
932}
933
f1d1c3fa
MD
934/*
935 * Yield our thread while higher priority threads are pending. This is
936 * typically called when we leave a critical section but it can be safely
937 * called while we are in a critical section.
938 *
939 * This function will not generally yield to equal priority threads but it
940 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 941 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
942 * lwkt_yield_quick().
943 *
235957ed 944 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
945 * came along but was blocked and made pending.
946 *
f1d1c3fa
MD
947 * (self contained on a per cpu basis)
948 */
949void
950lwkt_yield_quick(void)
951{
7966cb69
MD
952 globaldata_t gd = mycpu;
953 thread_t td = gd->gd_curthread;
ef0fdad1 954
a2a5ad0d 955 /*
235957ed 956 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
957 * it with a non-zero cpl then we might not wind up calling splz after
958 * a task switch when the critical section is exited even though the
46a3f46d 959 * new task could accept the interrupt.
a2a5ad0d
MD
960 *
961 * XXX from crit_exit() only called after last crit section is released.
962 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
963 *
964 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
965 * except for this special case, we MUST call splz() here to handle any
966 * pending ints, particularly after we switch, or we might accidently
967 * halt the cpu with interrupts pending.
a2a5ad0d 968 */
46a3f46d 969 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 970 splz();
f1d1c3fa
MD
971
972 /*
973 * YYY enabling will cause wakeup() to task-switch, which really
974 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
975 * preemption and MP without actually doing preemption or MP, because a
976 * lot of code assumes that wakeup() does not block.
f1d1c3fa 977 */
46a3f46d
MD
978 if (untimely_switch && td->td_nest_count == 0 &&
979 gd->gd_intr_nesting_level == 0
980 ) {
37af14fe 981 crit_enter_quick(td);
f1d1c3fa
MD
982 /*
983 * YYY temporary hacks until we disassociate the userland scheduler
984 * from the LWKT scheduler.
985 */
986 if (td->td_flags & TDF_RUNQ) {
987 lwkt_switch(); /* will not reenter yield function */
988 } else {
37af14fe 989 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 990 lwkt_switch(); /* will not reenter yield function */
37af14fe 991 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 992 }
7966cb69 993 crit_exit_noyield(td);
f1d1c3fa 994 }
f1d1c3fa
MD
995}
996
8ad65e08 997/*
f1d1c3fa 998 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 999 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
1000 * the crit_exit() call in lwkt_switch().
1001 *
1002 * (self contained on a per cpu basis)
8ad65e08
MD
1003 */
1004void
f1d1c3fa 1005lwkt_yield(void)
8ad65e08 1006{
37af14fe 1007 lwkt_schedule_self(curthread);
f1d1c3fa
MD
1008 lwkt_switch();
1009}
1010
3824f392
MD
1011/*
1012 * This function is used along with the lwkt_passive_recover() inline
1013 * by the trap code to negotiate a passive release of the current
1014 * process/lwp designation with the user scheduler.
1015 */
1016void
1017lwkt_passive_release(struct thread *td)
1018{
1019 struct lwp *lp = td->td_lwp;
1020
1021 td->td_release = NULL;
1022 lwkt_setpri_self(TDPRI_KERN_USER);
1023 lp->lwp_proc->p_usched->release_curproc(lp);
1024}
1025
1026/*
1027 * Make a kernel thread act as if it were in user mode with regards
1028 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
1029 * loops which may be potentially cpu-bound can call lwkt_user_yield().
1030 *
1031 * The lwkt_user_yield() function is designed to have very low overhead
1032 * if no yield is determined to be needed.
1033 */
1034void
1035lwkt_user_yield(void)
1036{
1037 thread_t td = curthread;
1038 struct lwp *lp = td->td_lwp;
1039
1040#ifdef SMP
1041 /*
1042 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1043 * kernel can prevent other cpus from servicing interrupt threads
1044 * which still require the MP lock (which is a lot of them). This
1045 * has a chaining effect since if the interrupt is blocked, so is
1046 * the event, so normal scheduling will not pick up on the problem.
1047 */
1048 if (mplock_countx && td->td_mpcount) {
1049 int savecnt = td->td_mpcount;
1050
1051 td->td_mpcount = 1;
1052 rel_mplock();
1053 DELAY(10);
1054 get_mplock();
1055 td->td_mpcount = savecnt;
1056 mplock_countx = 0;
1057 }
1058#endif
1059
1060 /*
1061 * Another kernel thread wants the cpu
1062 */
1063 if (lwkt_resched_wanted())
1064 lwkt_switch();
1065
1066 /*
1067 * If the user scheduler has asynchronously determined that the current
1068 * process (when running in user mode) needs to lose the cpu then make
1069 * sure we are released.
1070 */
1071 if (user_resched_wanted()) {
1072 if (td->td_release)
1073 td->td_release(td);
1074 }
1075
1076 /*
1077 * If we are released reduce our priority
1078 */
1079 if (td->td_release == NULL) {
1080 if (lwkt_check_resched(td) > 0)
1081 lwkt_switch();
1082 lp->lwp_proc->p_usched->acquire_curproc(lp);
1083 td->td_release = lwkt_passive_release;
1084 lwkt_setpri_self(TDPRI_USER_NORM);
1085 }
1086}
1087
b9eb1c19
MD
1088/*
1089 * Return 0 if no runnable threads are pending at the same or higher
1090 * priority as the passed thread.
1091 *
1092 * Return 1 if runnable threads are pending at the same priority.
1093 *
1094 * Return 2 if runnable threads are pending at a higher priority.
1095 */
1096int
1097lwkt_check_resched(thread_t td)
1098{
1099 int pri = td->td_pri & TDPRI_MASK;
1100
1101 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1102 return(2);
1103 if (TAILQ_NEXT(td, td_threadq))
1104 return(1);
1105 return(0);
1106}
1107
8ad65e08 1108/*
f1d1c3fa
MD
1109 * Generic schedule. Possibly schedule threads belonging to other cpus and
1110 * deal with threads that might be blocked on a wait queue.
1111 *
0a3f9b47
MD
1112 * We have a little helper inline function which does additional work after
1113 * the thread has been enqueued, including dealing with preemption and
1114 * setting need_lwkt_resched() (which prevents the kernel from returning
1115 * to userland until it has processed higher priority threads).
6330a558
MD
1116 *
1117 * It is possible for this routine to be called after a failed _enqueue
1118 * (due to the target thread migrating, sleeping, or otherwise blocked).
1119 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1120 *
1121 * reschedok is an optimized constant propagated from lwkt_schedule() or
1122 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1123 * reschedule to be requested if the target thread has a higher priority.
1124 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1125 * be 0, prevented undesired reschedules.
8ad65e08 1126 */
0a3f9b47
MD
1127static __inline
1128void
361d01dd 1129_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1130{
b9eb1c19 1131 thread_t otd;
c730be20 1132
6330a558 1133 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1134 if (ntd->td_preemptable && reschedok) {
6330a558 1135 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1136 } else if (reschedok) {
b9eb1c19
MD
1137 otd = curthread;
1138 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1139 need_lwkt_resched();
6330a558 1140 }
0a3f9b47
MD
1141 }
1142}
1143
361d01dd 1144static __inline
8ad65e08 1145void
361d01dd 1146_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1147{
37af14fe
MD
1148 globaldata_t mygd = mycpu;
1149
41a01a4d 1150 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1151 crit_enter_gd(mygd);
9388413d 1152 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1153 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1154 _lwkt_enqueue(td);
1155 } else {
f1d1c3fa 1156 /*
7cd8d145
MD
1157 * If we own the thread, there is no race (since we are in a
1158 * critical section). If we do not own the thread there might
1159 * be a race but the target cpu will deal with it.
f1d1c3fa 1160 */
0f7a3396 1161#ifdef SMP
7cd8d145 1162 if (td->td_gd == mygd) {
9d265729 1163 _lwkt_enqueue(td);
361d01dd 1164 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1165 } else {
7cd8d145
MD
1166 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1167 }
0f7a3396 1168#else
7cd8d145 1169 _lwkt_enqueue(td);
361d01dd 1170 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1171#endif
8ad65e08 1172 }
37af14fe 1173 crit_exit_gd(mygd);
8ad65e08
MD
1174}
1175
361d01dd
MD
1176void
1177lwkt_schedule(thread_t td)
1178{
1179 _lwkt_schedule(td, 1);
1180}
1181
1182void
1183lwkt_schedule_noresched(thread_t td)
1184{
1185 _lwkt_schedule(td, 0);
1186}
1187
52eedfb5
MD
1188#ifdef SMP
1189
d9eea1a5 1190/*
52eedfb5
MD
1191 * Thread migration using a 'Pull' method. The thread may or may not be
1192 * the current thread. It MUST be descheduled and in a stable state.
1193 * lwkt_giveaway() must be called on the cpu owning the thread.
1194 *
1195 * At any point after lwkt_giveaway() is called, the target cpu may
1196 * 'pull' the thread by calling lwkt_acquire().
1197 *
ae8e83e6
MD
1198 * We have to make sure the thread is not sitting on a per-cpu tsleep
1199 * queue or it will blow up when it moves to another cpu.
1200 *
52eedfb5 1201 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1202 */
52eedfb5
MD
1203void
1204lwkt_giveaway(thread_t td)
1205{
1206 globaldata_t gd = mycpu;
1207
1208 crit_enter_gd(gd);
ae8e83e6
MD
1209 if (td->td_flags & TDF_TSLEEPQ)
1210 tsleep_remove(td);
52eedfb5
MD
1211 KKASSERT(td->td_gd == gd);
1212 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1213 td->td_flags |= TDF_MIGRATING;
1214 crit_exit_gd(gd);
1215}
1216
a2a5ad0d
MD
1217void
1218lwkt_acquire(thread_t td)
1219{
37af14fe
MD
1220 globaldata_t gd;
1221 globaldata_t mygd;
a2a5ad0d 1222
52eedfb5 1223 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1224 gd = td->td_gd;
37af14fe 1225 mygd = mycpu;
52eedfb5 1226 if (gd != mycpu) {
35238fa5 1227 cpu_lfence();
52eedfb5 1228 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1229 crit_enter_gd(mygd);
df910c23
MD
1230 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1231#ifdef SMP
1232 lwkt_process_ipiq();
1233#endif
52eedfb5 1234 cpu_lfence();
df910c23 1235 }
37af14fe 1236 td->td_gd = mygd;
52eedfb5
MD
1237 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1238 td->td_flags &= ~TDF_MIGRATING;
1239 crit_exit_gd(mygd);
1240 } else {
1241 crit_enter_gd(mygd);
1242 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1243 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1244 crit_exit_gd(mygd);
a2a5ad0d
MD
1245 }
1246}
1247
52eedfb5
MD
1248#endif
1249
f1d1c3fa
MD
1250/*
1251 * Generic deschedule. Descheduling threads other then your own should be
1252 * done only in carefully controlled circumstances. Descheduling is
1253 * asynchronous.
1254 *
1255 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1256 */
1257void
1258lwkt_deschedule(thread_t td)
1259{
f1d1c3fa 1260 crit_enter();
b8a98473 1261#ifdef SMP
f1d1c3fa
MD
1262 if (td == curthread) {
1263 _lwkt_dequeue(td);
1264 } else {
a72187e9 1265 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1266 _lwkt_dequeue(td);
1267 } else {
b8a98473 1268 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1269 }
1270 }
b8a98473
MD
1271#else
1272 _lwkt_dequeue(td);
1273#endif
f1d1c3fa
MD
1274 crit_exit();
1275}
1276
4b5f931b
MD
1277/*
1278 * Set the target thread's priority. This routine does not automatically
1279 * switch to a higher priority thread, LWKT threads are not designed for
1280 * continuous priority changes. Yield if you want to switch.
1281 *
1282 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1283 * of the td_pri field. The specified priority may also indicate zero or
1284 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1285 *
1286 * Note that we requeue the thread whether it winds up on a different runq
1287 * or not. uio_yield() depends on this and the routine is not normally
1288 * called with the same priority otherwise.
4b5f931b
MD
1289 */
1290void
1291lwkt_setpri(thread_t td, int pri)
1292{
26a0694b 1293 KKASSERT(pri >= 0);
a72187e9 1294 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1295 crit_enter();
1296 if (td->td_flags & TDF_RUNQ) {
1297 _lwkt_dequeue(td);
1298 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1299 _lwkt_enqueue(td);
1300 } else {
1301 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1302 }
1303 crit_exit();
1304}
1305
1306void
1307lwkt_setpri_self(int pri)
1308{
1309 thread_t td = curthread;
1310
4b5f931b
MD
1311 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1312 crit_enter();
1313 if (td->td_flags & TDF_RUNQ) {
1314 _lwkt_dequeue(td);
1315 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1316 _lwkt_enqueue(td);
1317 } else {
1318 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1319 }
1320 crit_exit();
1321}
1322
5d21b981 1323/*
52eedfb5
MD
1324 * Migrate the current thread to the specified cpu.
1325 *
1326 * This is accomplished by descheduling ourselves from the current cpu,
1327 * moving our thread to the tdallq of the target cpu, IPI messaging the
1328 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1329 * races while the thread is being migrated.
ae8e83e6
MD
1330 *
1331 * We must be sure to remove ourselves from the current cpu's tsleepq
1332 * before potentially moving to another queue. The thread can be on
1333 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1334 */
3d28ff59 1335#ifdef SMP
5d21b981 1336static void lwkt_setcpu_remote(void *arg);
3d28ff59 1337#endif
5d21b981
MD
1338
1339void
1340lwkt_setcpu_self(globaldata_t rgd)
1341{
1342#ifdef SMP
1343 thread_t td = curthread;
1344
1345 if (td->td_gd != rgd) {
1346 crit_enter_quick(td);
ae8e83e6
MD
1347 if (td->td_flags & TDF_TSLEEPQ)
1348 tsleep_remove(td);
5d21b981
MD
1349 td->td_flags |= TDF_MIGRATING;
1350 lwkt_deschedule_self(td);
52eedfb5 1351 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1352 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1353 lwkt_switch();
1354 /* we are now on the target cpu */
52eedfb5 1355 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1356 crit_exit_quick(td);
1357 }
1358#endif
1359}
1360
ecdefdda
MD
1361void
1362lwkt_migratecpu(int cpuid)
1363{
1364#ifdef SMP
1365 globaldata_t rgd;
1366
1367 rgd = globaldata_find(cpuid);
1368 lwkt_setcpu_self(rgd);
1369#endif
1370}
1371
5d21b981
MD
1372/*
1373 * Remote IPI for cpu migration (called while in a critical section so we
1374 * do not have to enter another one). The thread has already been moved to
1375 * our cpu's allq, but we must wait for the thread to be completely switched
1376 * out on the originating cpu before we schedule it on ours or the stack
1377 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1378 * change to main memory.
1379 *
1380 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1381 * against wakeups. It is best if this interface is used only when there
1382 * are no pending events that might try to schedule the thread.
1383 */
3d28ff59 1384#ifdef SMP
5d21b981
MD
1385static void
1386lwkt_setcpu_remote(void *arg)
1387{
1388 thread_t td = arg;
1389 globaldata_t gd = mycpu;
1390
df910c23
MD
1391 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1392#ifdef SMP
1393 lwkt_process_ipiq();
1394#endif
35238fa5 1395 cpu_lfence();
df910c23 1396 }
5d21b981 1397 td->td_gd = gd;
35238fa5 1398 cpu_sfence();
5d21b981 1399 td->td_flags &= ~TDF_MIGRATING;
9388413d 1400 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1401 _lwkt_enqueue(td);
1402}
3d28ff59 1403#endif
5d21b981 1404
553ea3c8 1405struct lwp *
4b5f931b
MD
1406lwkt_preempted_proc(void)
1407{
73e4f7b9 1408 thread_t td = curthread;
4b5f931b
MD
1409 while (td->td_preempted)
1410 td = td->td_preempted;
553ea3c8 1411 return(td->td_lwp);
4b5f931b
MD
1412}
1413
99df837e
MD
1414/*
1415 * Create a kernel process/thread/whatever. It shares it's address space
1416 * with proc0 - ie: kernel only.
1417 *
365fa13f
MD
1418 * NOTE! By default new threads are created with the MP lock held. A
1419 * thread which does not require the MP lock should release it by calling
1420 * rel_mplock() at the start of the new thread.
99df837e
MD
1421 */
1422int
1423lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1424 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1425 const char *fmt, ...)
99df837e 1426{
73e4f7b9 1427 thread_t td;
e2565a42 1428 __va_list ap;
99df837e 1429
d3d32139 1430 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1431 tdflags);
a2a5ad0d
MD
1432 if (tdp)
1433 *tdp = td;
709799ea 1434 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1435
1436 /*
1437 * Set up arg0 for 'ps' etc
1438 */
e2565a42 1439 __va_start(ap, fmt);
379210cb 1440 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1441 __va_end(ap);
99df837e
MD
1442
1443 /*
1444 * Schedule the thread to run
1445 */
ef0fdad1
MD
1446 if ((td->td_flags & TDF_STOPREQ) == 0)
1447 lwkt_schedule(td);
1448 else
1449 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1450 return 0;
1451}
1452
1453/*
1454 * Destroy an LWKT thread. Warning! This function is not called when
1455 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1456 * uses a different reaping mechanism.
1457 */
1458void
1459lwkt_exit(void)
1460{
1461 thread_t td = curthread;
c070746a 1462 thread_t std;
8826f33a 1463 globaldata_t gd;
99df837e
MD
1464
1465 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1466 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1467 caps_exit(td);
c070746a
MD
1468
1469 /*
1470 * Get us into a critical section to interlock gd_freetd and loop
1471 * until we can get it freed.
1472 *
1473 * We have to cache the current td in gd_freetd because objcache_put()ing
1474 * it would rip it out from under us while our thread is still active.
1475 */
1476 gd = mycpu;
37af14fe 1477 crit_enter_quick(td);
c070746a
MD
1478 while ((std = gd->gd_freetd) != NULL) {
1479 gd->gd_freetd = NULL;
1480 objcache_put(thread_cache, std);
1481 }
37af14fe 1482 lwkt_deschedule_self(td);
e56e4dea 1483 lwkt_remove_tdallq(td);
c070746a
MD
1484 if (td->td_flags & TDF_ALLOCATED_THREAD)
1485 gd->gd_freetd = td;
99df837e
MD
1486 cpu_thread_exit();
1487}
1488
e56e4dea
MD
1489void
1490lwkt_remove_tdallq(thread_t td)
1491{
1492 KKASSERT(td->td_gd == mycpu);
1493 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1494}
1495
2d93b37a
MD
1496void
1497crit_panic(void)
1498{
1499 thread_t td = curthread;
1500 int lpri = td->td_pri;
1501
1502 td->td_pri = 0;
1503 panic("td_pri is/would-go negative! %p %d", td, lpri);
1504}
1505
d165e668
MD
1506#ifdef SMP
1507
bd8015ca
MD
1508/*
1509 * Called from debugger/panic on cpus which have been stopped. We must still
1510 * process the IPIQ while stopped, even if we were stopped while in a critical
1511 * section (XXX).
1512 *
1513 * If we are dumping also try to process any pending interrupts. This may
1514 * or may not work depending on the state of the cpu at the point it was
1515 * stopped.
1516 */
1517void
1518lwkt_smp_stopped(void)
1519{
1520 globaldata_t gd = mycpu;
1521
1522 crit_enter_gd(gd);
1523 if (dumping) {
1524 lwkt_process_ipiq();
1525 splz();
1526 } else {
1527 lwkt_process_ipiq();
1528 }
1529 crit_exit_gd(gd);
1530}
1531
57aa743c
MD
1532/*
1533 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1534 * get_mplock() has already incremented td_mpcount. We must block and
1535 * not return until giant is held.
1536 *
1537 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1538 * reschedule the thread until it can obtain the giant lock for it.
1539 */
1540void
1541lwkt_mp_lock_contested(void)
1542{
3824f392 1543 ++mplock_countx;
57aa743c 1544 loggiant(beg);
57aa743c 1545 lwkt_switch();
57aa743c 1546 loggiant(end);
57aa743c
MD
1547}
1548
b9eb1c19
MD
1549/*
1550 * The rel_mplock() code will call this function after releasing the
1551 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1552 *
1553 * We then chain an IPI to a single other cpu potentially needing the
1554 * lock. This is a bit heuristical and we can wind up with IPIs flying
1555 * all over the place.
1556 */
1557static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1558
1559void
1560lwkt_mp_lock_uncontested(void)
1561{
1562 globaldata_t gd;
1563 globaldata_t dgd;
1564 cpumask_t mask;
1565 cpumask_t tmpmask;
1566 int cpuid;
1567
1568 if (chain_mplock) {
1569 gd = mycpu;
1570 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1571 mask = mp_lock_contention_mask;
1572 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1573
1574 if (mask) {
1575 if (mask & tmpmask)
1576 cpuid = bsfl(mask & tmpmask);
1577 else
1578 cpuid = bsfl(mask);
1579 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1580 dgd = globaldata_find(cpuid);
1581 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1582 }
1583 }
1584}
1585
1586/*
1587 * The idea is for this IPI to interrupt a potentially lower priority
1588 * thread, such as a user thread, to allow the scheduler to reschedule
1589 * a higher priority kernel thread that needs the MP lock.
1590 *
1591 * For now we set the LWKT reschedule flag which generates an AST in
1592 * doreti, though theoretically it is also possible to possibly preempt
1593 * here if the underlying thread was operating in user mode. Nah.
1594 */
1595static void
1596lwkt_mp_lock_uncontested_remote(void *arg __unused)
1597{
1598 need_lwkt_resched();
1599}
1600
d165e668 1601#endif