kernel - Make sys_ioctl() MPSAFE
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
b2b3ffcd 110#ifdef __x86_64__
85514115
MD
111
112static int
0855a2af
JG
113jg_tos_ok(struct thread *td)
114{
85514115
MD
115 void *tos;
116 int tos_ok;
117
0855a2af
JG
118 if (td == NULL) {
119 return 1;
120 }
121 KKASSERT(td->td_sp != NULL);
85514115
MD
122 tos = ((void **)td->td_sp)[0];
123 tos_ok = 0;
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
126 tos_ok = 1;
127 }
128 return tos_ok;
129}
130
85514115
MD
131#endif
132
fb0f29c4
MD
133/*
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
136 */
137TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138
0f7a3396
MD
139#ifdef INVARIANTS
140SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141#endif
4b5f931b 142SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
9733f757
VS
143SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
26a0694b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
148#ifdef INVARIANTS
149SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
38717797 151#endif
f9235b6d
MD
152static int fairq_enable = 1;
153SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
05220613 154
4b5f931b
MD
155/*
156 * These helper procedures handle the runq, they can only be called from
157 * within a critical section.
75cdbe6c
MD
158 *
159 * WARNING! Prior to SMP being brought up it is possible to enqueue and
160 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
161 * instead of 'mycpu' when referencing the globaldata structure. Once
162 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 163 */
f1d1c3fa
MD
164static __inline
165void
166_lwkt_dequeue(thread_t td)
167{
168 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 169 struct globaldata *gd = td->td_gd;
4b5f931b 170
f1d1c3fa 171 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
172 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
173 gd->gd_fairq_total_pri -= td->td_pri;
174 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
175 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
176 }
177}
178
f9235b6d
MD
179/*
180 * Priority enqueue.
181 *
182 * NOTE: There are a limited number of lwkt threads runnable since user
183 * processes only schedule one at a time per cpu.
184 */
f1d1c3fa
MD
185static __inline
186void
187_lwkt_enqueue(thread_t td)
188{
f9235b6d
MD
189 thread_t xtd;
190
7f5d7ed7 191 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 192 struct globaldata *gd = td->td_gd;
4b5f931b 193
f1d1c3fa 194 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
195 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
196 if (xtd == NULL) {
197 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
198 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
199 } else {
200 while (xtd && xtd->td_pri > td->td_pri)
201 xtd = TAILQ_NEXT(xtd, td_threadq);
202 if (xtd)
203 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
204 else
205 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 }
207 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
208 }
209}
8ad65e08 210
40aaf5fc
NT
211static __boolean_t
212_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
213{
214 struct thread *td = (struct thread *)obj;
215
216 td->td_kstack = NULL;
217 td->td_kstack_size = 0;
218 td->td_flags = TDF_ALLOCATED_THREAD;
219 return (1);
220}
221
222static void
223_lwkt_thread_dtor(void *obj, void *privdata)
224{
225 struct thread *td = (struct thread *)obj;
226
227 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
228 ("_lwkt_thread_dtor: not allocated from objcache"));
229 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
230 td->td_kstack_size > 0,
231 ("_lwkt_thread_dtor: corrupted stack"));
232 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
233}
234
235/*
236 * Initialize the lwkt s/system.
237 */
238void
239lwkt_init(void)
240{
241 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
242 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
243 NULL, CACHE_NTHREADS/2,
244 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
245}
246
37af14fe
MD
247/*
248 * Schedule a thread to run. As the current thread we can always safely
249 * schedule ourselves, and a shortcut procedure is provided for that
250 * function.
251 *
252 * (non-blocking, self contained on a per cpu basis)
253 */
254void
255lwkt_schedule_self(thread_t td)
256{
257 crit_enter_quick(td);
f9235b6d
MD
258 KASSERT(td != &td->td_gd->gd_idlethread,
259 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 260 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 261 _lwkt_enqueue(td);
37af14fe
MD
262 crit_exit_quick(td);
263}
264
265/*
266 * Deschedule a thread.
267 *
268 * (non-blocking, self contained on a per cpu basis)
269 */
270void
271lwkt_deschedule_self(thread_t td)
272{
273 crit_enter_quick(td);
37af14fe
MD
274 _lwkt_dequeue(td);
275 crit_exit_quick(td);
276}
277
8ad65e08
MD
278/*
279 * LWKTs operate on a per-cpu basis
280 *
73e4f7b9 281 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
282 */
283void
284lwkt_gdinit(struct globaldata *gd)
285{
f9235b6d 286 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 287 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
288}
289
7d0bac62
MD
290/*
291 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
292 * or LWKT start address before it can be scheduled. If the target cpu is
293 * -1 the thread will be created on the current cpu.
0cfcada1
MD
294 *
295 * If you intend to create a thread without a process context this function
296 * does everything except load the startup and switcher function.
7d0bac62
MD
297 */
298thread_t
d3d32139 299lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 300{
c070746a 301 globaldata_t gd = mycpu;
99df837e 302 void *stack;
7d0bac62 303
c070746a
MD
304 /*
305 * If static thread storage is not supplied allocate a thread. Reuse
306 * a cached free thread if possible. gd_freetd is used to keep an exiting
307 * thread intact through the exit.
308 */
ef0fdad1 309 if (td == NULL) {
c070746a
MD
310 if ((td = gd->gd_freetd) != NULL)
311 gd->gd_freetd = NULL;
312 else
313 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
314 KASSERT((td->td_flags &
315 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
316 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
317 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 318 }
c070746a
MD
319
320 /*
321 * Try to reuse cached stack.
322 */
f470d0c8
MD
323 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
324 if (flags & TDF_ALLOCATED_STACK) {
e4846942 325 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
326 stack = NULL;
327 }
328 }
329 if (stack == NULL) {
e4846942 330 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 331 flags |= TDF_ALLOCATED_STACK;
99df837e 332 }
75cdbe6c 333 if (cpu < 0)
c070746a 334 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 335 else
f470d0c8 336 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 337 return(td);
7d0bac62
MD
338}
339
340/*
341 * Initialize a preexisting thread structure. This function is used by
342 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
343 *
f8c3996b
MD
344 * All threads start out in a critical section at a priority of
345 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
346 * appropriate. This function may send an IPI message when the
347 * requested cpu is not the current cpu and consequently gd_tdallq may
348 * not be initialized synchronously from the point of view of the originating
349 * cpu.
350 *
351 * NOTE! we have to be careful in regards to creating threads for other cpus
352 * if SMP has not yet been activated.
7d0bac62 353 */
41a01a4d
MD
354#ifdef SMP
355
75cdbe6c
MD
356static void
357lwkt_init_thread_remote(void *arg)
358{
359 thread_t td = arg;
360
52eedfb5
MD
361 /*
362 * Protected by critical section held by IPI dispatch
363 */
75cdbe6c
MD
364 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
365}
366
41a01a4d
MD
367#endif
368
7d0bac62 369void
f470d0c8
MD
370lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
371 struct globaldata *gd)
7d0bac62 372{
37af14fe
MD
373 globaldata_t mygd = mycpu;
374
99df837e
MD
375 bzero(td, sizeof(struct thread));
376 td->td_kstack = stack;
f470d0c8 377 td->td_kstack_size = stksize;
d3d32139 378 td->td_flags = flags;
26a0694b 379 td->td_gd = gd;
f9235b6d
MD
380 td->td_pri = TDPRI_KERN_DAEMON;
381 td->td_critcount = 1;
3b998fa9 382 td->td_toks_stop = &td->td_toks_base;
d3d32139
MD
383#ifdef SMP
384 if ((flags & TDF_MPSAFE) == 0)
385 td->td_mpcount = 1;
386#endif
fb0f29c4
MD
387 if (lwkt_use_spin_port)
388 lwkt_initport_spin(&td->td_msgport);
389 else
390 lwkt_initport_thread(&td->td_msgport, td);
99df837e 391 pmap_init_thread(td);
0f7a3396 392#ifdef SMP
5d21b981
MD
393 /*
394 * Normally initializing a thread for a remote cpu requires sending an
395 * IPI. However, the idlethread is setup before the other cpus are
396 * activated so we have to treat it as a special case. XXX manipulation
397 * of gd_tdallq requires the BGL.
398 */
399 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 400 crit_enter_gd(mygd);
75cdbe6c 401 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 402 crit_exit_gd(mygd);
75cdbe6c 403 } else {
2db3b277 404 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 405 }
0f7a3396 406#else
37af14fe 407 crit_enter_gd(mygd);
0f7a3396 408 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 409 crit_exit_gd(mygd);
0f7a3396 410#endif
8c72e3d5
AH
411
412 dsched_new_thread(td);
73e4f7b9
MD
413}
414
415void
416lwkt_set_comm(thread_t td, const char *ctl, ...)
417{
e2565a42 418 __va_list va;
73e4f7b9 419
e2565a42 420 __va_start(va, ctl);
379210cb 421 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 422 __va_end(va);
e7c0dbba 423 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
424}
425
99df837e 426void
73e4f7b9 427lwkt_hold(thread_t td)
99df837e 428{
73e4f7b9
MD
429 ++td->td_refs;
430}
431
432void
433lwkt_rele(thread_t td)
434{
435 KKASSERT(td->td_refs > 0);
436 --td->td_refs;
437}
438
439void
440lwkt_wait_free(thread_t td)
441{
442 while (td->td_refs)
377d4740 443 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
444}
445
446void
447lwkt_free_thread(thread_t td)
448{
d9eea1a5 449 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
450 ("lwkt_free_thread: did not exit! %p", td));
451
40aaf5fc
NT
452 if (td->td_flags & TDF_ALLOCATED_THREAD) {
453 objcache_put(thread_cache, td);
454 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
455 /* client-allocated struct with internally allocated stack */
456 KASSERT(td->td_kstack && td->td_kstack_size > 0,
457 ("lwkt_free_thread: corrupted stack"));
458 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
459 td->td_kstack = NULL;
460 td->td_kstack_size = 0;
99df837e 461 }
e7c0dbba 462 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
463}
464
465
8ad65e08
MD
466/*
467 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
468 * switch to the idlethread. Switching must occur within a critical
469 * section to avoid races with the scheduling queue.
470 *
471 * We always have full control over our cpu's run queue. Other cpus
472 * that wish to manipulate our queue must use the cpu_*msg() calls to
473 * talk to our cpu, so a critical section is all that is needed and
474 * the result is very, very fast thread switching.
475 *
96728c05
MD
476 * The LWKT scheduler uses a fixed priority model and round-robins at
477 * each priority level. User process scheduling is a totally
478 * different beast and LWKT priorities should not be confused with
479 * user process priorities.
f1d1c3fa 480 *
96728c05
MD
481 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
482 * cleans it up. Note that the td_switch() function cannot do anything that
483 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
484 * the target thread (not the current thread). It's nice to have a scheduler
485 * that does not need the MP lock to work because it allows us to do some
486 * really cool high-performance MP lock optimizations.
69d78e99
MD
487 *
488 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
489 * is not called by the current thread in the preemption case, only when
490 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
491 */
492void
493lwkt_switch(void)
494{
37af14fe
MD
495 globaldata_t gd = mycpu;
496 thread_t td = gd->gd_curthread;
8ad65e08 497 thread_t ntd;
f9235b6d
MD
498 thread_t xtd;
499 thread_t nlast;
f9235b6d 500 int nquserok;
6f207a2c 501#ifdef SMP
8a8d5d85
MD
502 int mpheld;
503#endif
f9235b6d 504 int didaccumulate;
b37f18d6
MD
505 const char *lmsg; /* diagnostic - 'systat -pv 1' */
506 const void *laddr;
8ad65e08 507
46a3f46d 508 /*
27e88a6e
MD
509 * Switching from within a 'fast' (non thread switched) interrupt or IPI
510 * is illegal. However, we may have to do it anyway if we hit a fatal
511 * kernel trap or we have paniced.
512 *
513 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 514 */
27e88a6e
MD
515 if (gd->gd_intr_nesting_level) {
516 int savegdnest;
517 int savegdtrap;
518
519 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
520 panic("lwkt_switch: cannot switch from within "
521 "a fast interrupt, yet, td %p\n", td);
522 } else {
523 savegdnest = gd->gd_intr_nesting_level;
524 savegdtrap = gd->gd_trap_nesting_level;
525 gd->gd_intr_nesting_level = 0;
526 gd->gd_trap_nesting_level = 0;
a7422615
MD
527 if ((td->td_flags & TDF_PANICWARN) == 0) {
528 td->td_flags |= TDF_PANICWARN;
6ea70f76 529 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 530 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 531 print_backtrace(-1);
a7422615 532 }
27e88a6e
MD
533 lwkt_switch();
534 gd->gd_intr_nesting_level = savegdnest;
535 gd->gd_trap_nesting_level = savegdtrap;
536 return;
537 }
96728c05 538 }
ef0fdad1 539
cb973d15
MD
540 /*
541 * Passive release (used to transition from user to kernel mode
542 * when we block or switch rather then when we enter the kernel).
543 * This function is NOT called if we are switching into a preemption
544 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
545 * our current process designation (if we have one) and become a true
546 * LWKT thread, and may also hand the current process designation to
547 * another process and schedule thread.
cb973d15
MD
548 */
549 if (td->td_release)
550 td->td_release(td);
551
37af14fe 552 crit_enter_gd(gd);
3b998fa9 553 if (TD_TOKS_HELD(td))
9d265729
MD
554 lwkt_relalltokens(td);
555
556 /*
b02926de
MD
557 * We had better not be holding any spin locks, but don't get into an
558 * endless panic loop.
9d265729 559 */
bbb31c5d
MD
560 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
561 ("lwkt_switch: still holding a shared spinlock %p!",
562 gd->gd_spinlock_rd));
d666840a
MD
563 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
564 ("lwkt_switch: still holding %d exclusive spinlocks!",
565 gd->gd_spinlocks_wr));
9d265729 566
8a8d5d85
MD
567
568#ifdef SMP
569 /*
570 * td_mpcount cannot be used to determine if we currently hold the
571 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
572 * to get the lock, and switch out if it can't. Our ownership of
573 * the actual lock will remain stable while we are in a critical section
574 * (but, of course, another cpu may own or release the lock so the
575 * actual value of mp_lock is not stable).
8a8d5d85
MD
576 */
577 mpheld = MP_LOCK_HELD();
0f7a3396
MD
578#ifdef INVARIANTS
579 if (td->td_cscount) {
6ea70f76 580 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
581 td);
582 if (panic_on_cscount)
583 panic("switching while mastering cpusync");
584 }
585#endif
8a8d5d85 586#endif
f9235b6d
MD
587
588 /*
589 * If we had preempted another thread on this cpu, resume the preempted
590 * thread. This occurs transparently, whether the preempted thread
591 * was scheduled or not (it may have been preempted after descheduling
592 * itself).
593 *
594 * We have to setup the MP lock for the original thread after backing
595 * out the adjustment that was made to curthread when the original
596 * was preempted.
597 */
99df837e 598 if ((ntd = td->td_preempted) != NULL) {
26a0694b 599 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 600#ifdef SMP
96728c05 601 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 602 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
603 td, ntd, td->td_mpcount, ntd->td_mpcount);
604 }
8a8d5d85
MD
605 if (ntd->td_mpcount) {
606 td->td_mpcount -= ntd->td_mpcount;
607 KKASSERT(td->td_mpcount >= 0);
608 }
609#endif
26a0694b 610 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
611
612 /*
b9eb1c19
MD
613 * The interrupt may have woken a thread up, we need to properly
614 * set the reschedule flag if the originally interrupted thread is
615 * at a lower priority.
8ec60c3f 616 */
f9235b6d
MD
617 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
618 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 619 need_lwkt_resched();
f9235b6d 620 }
8a8d5d85 621 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
622 goto havethread_preempted;
623 }
624
625 /*
626 * Implement round-robin fairq with priority insertion. The priority
627 * insertion is handled by _lwkt_enqueue()
628 *
629 * We have to adjust the MP lock for the target thread. If we
630 * need the MP lock and cannot obtain it we try to locate a
631 * thread that does not need the MP lock. If we cannot, we spin
632 * instead of HLT.
633 *
634 * A similar issue exists for the tokens held by the target thread.
635 * If we cannot obtain ownership of the tokens we cannot immediately
636 * schedule the thread.
637 */
638 for (;;) {
639 clear_lwkt_resched();
640 didaccumulate = 0;
641 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
642
4b5f931b 643 /*
f9235b6d 644 * Hotpath if we can get all necessary resources.
41a01a4d 645 *
f9235b6d 646 * If nothing is runnable switch to the idle thread
41a01a4d 647 */
f9235b6d
MD
648 if (ntd == NULL) {
649 ntd = &gd->gd_idlethread;
650 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
651 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 652#ifdef SMP
f9235b6d
MD
653 if (ntd->td_mpcount) {
654 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
655 panic("Idle thread %p was holding the BGL!", ntd);
656 if (mpheld == 0) {
657 cpu_pause();
658 continue;
659 }
660 }
6f207a2c 661#endif
b37f18d6
MD
662 cpu_time.cp_msg[0] = 0;
663 cpu_time.cp_stallpc = 0;
f9235b6d
MD
664 goto haveidle;
665 }
41a01a4d 666
8ec60c3f 667 /*
f9235b6d 668 * Hotpath schedule
6f207a2c
MD
669 *
670 * NOTE: For UP there is no mplock and lwkt_getalltokens()
671 * always succeeds.
8ec60c3f 672 */
f9235b6d
MD
673 if (ntd->td_fairq_accum >= 0 &&
674#ifdef SMP
675 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
676#endif
b37f18d6 677 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 678 ) {
8a8d5d85 679#ifdef SMP
f9235b6d
MD
680 clr_mplock_contention_mask(gd);
681#endif
682 goto havethread;
683 }
684
b37f18d6
MD
685 lmsg = NULL;
686 laddr = NULL;
687
f9235b6d
MD
688#ifdef SMP
689 /* Reload mpheld (it become stale after mplock/token ops) */
690 mpheld = MP_LOCK_HELD();
b37f18d6
MD
691 if (ntd->td_mpcount && mpheld == 0) {
692 lmsg = "mplock";
693 laddr = ntd->td_mplock_stallpc;
694 }
f9235b6d
MD
695#endif
696
697 /*
698 * Coldpath - unable to schedule ntd, continue looking for threads
699 * to schedule. This is only allowed of the (presumably) kernel
700 * thread exhausted its fair share. A kernel thread stuck on
701 * resources does not currently allow a user thread to get in
702 * front of it.
703 */
704#ifdef SMP
705 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
706 (ntd->td_fairq_accum < 0));
6f207a2c
MD
707#else
708 nquserok = 1;
f9235b6d
MD
709#endif
710 nlast = NULL;
711
712 for (;;) {
41a01a4d 713 /*
f9235b6d
MD
714 * If the fair-share scheduler ran out ntd gets moved to the
715 * end and its accumulator will be bumped, if it didn't we
716 * maintain the same queue position.
df6b8ba0 717 *
f9235b6d 718 * nlast keeps track of the last element prior to any moves.
41a01a4d 719 */
f9235b6d
MD
720 if (ntd->td_fairq_accum < 0) {
721 xtd = TAILQ_NEXT(ntd, td_threadq);
722 lwkt_fairq_accumulate(gd, ntd);
723 didaccumulate = 1;
724 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
725 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
726 if (nlast == NULL) {
727 nlast = ntd;
728 if (xtd == NULL)
729 xtd = ntd;
730 }
731 ntd = xtd;
732 } else {
733 ntd = TAILQ_NEXT(ntd, td_threadq);
734 }
a453459d 735
f9235b6d
MD
736 /*
737 * If we exhausted the run list switch to the idle thread.
738 * Since one or more threads had resource acquisition issues
739 * we do not allow the idle thread to halt.
740 *
741 * NOTE: nlast can be NULL.
742 */
743 if (ntd == nlast) {
e0a90d3b 744 cpu_pause();
f9235b6d
MD
745 ntd = &gd->gd_idlethread;
746 ntd->td_flags |= TDF_IDLE_NOHLT;
6f207a2c 747#ifdef SMP
f9235b6d
MD
748 set_mplock_contention_mask(gd);
749 cpu_mplock_contested();
750 if (ntd->td_mpcount) {
751 mpheld = MP_LOCK_HELD();
752 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
753 panic("Idle thread %p was holding the BGL!", ntd);
754 if (mpheld == 0) {
755 cpu_pause();
756 break; /* try again from the top, almost */
b9eb1c19 757 }
8a8d5d85 758 }
6f207a2c 759#endif
684a93c4
MD
760
761 /*
f9235b6d
MD
762 * If fairq accumulations occured we do not schedule the
763 * idle thread. This will cause us to try again from
764 * the (almost) top.
684a93c4 765 */
f9235b6d 766 if (didaccumulate)
b37f18d6
MD
767 break; /* try again from the top, almost */
768 if (lmsg)
769 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
770 cpu_time.cp_stallpc = (uintptr_t)laddr;
f9235b6d 771 goto haveidle;
8a8d5d85 772 }
f9235b6d 773
df6b8ba0 774 /*
f9235b6d 775 * Try to switch to this thread.
6f207a2c
MD
776 *
777 * NOTE: For UP there is no mplock and lwkt_getalltokens()
778 * always succeeds.
df6b8ba0 779 */
f9235b6d
MD
780 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok) &&
781 ntd->td_fairq_accum >= 0 &&
782#ifdef SMP
783 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
8a8d5d85 784#endif
b37f18d6 785 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
f9235b6d 786 ) {
a453459d 787#ifdef SMP
f9235b6d
MD
788 clr_mplock_contention_mask(gd);
789#endif
790 goto havethread;
df6b8ba0 791 }
f9235b6d
MD
792#ifdef SMP
793 /* Reload mpheld (it become stale after mplock/token ops) */
794 mpheld = MP_LOCK_HELD();
b37f18d6
MD
795 if (ntd->td_mpcount && mpheld == 0) {
796 lmsg = "mplock";
797 laddr = ntd->td_mplock_stallpc;
798 }
799
f9235b6d
MD
800 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
801 nquserok = 0;
a453459d 802#endif
4b5f931b 803 }
f1d1c3fa 804 }
8a8d5d85
MD
805
806 /*
f9235b6d
MD
807 * Do the actual switch. WARNING: mpheld is stale here.
808 *
809 * We must always decrement td_fairq_accum on non-idle threads just
810 * in case a thread never gets a tick due to being in a continuous
811 * critical section. The page-zeroing code does that.
812 *
813 * If the thread we came up with is a higher or equal priority verses
814 * the thread at the head of the queue we move our thread to the
815 * front. This way we can always check the front of the queue.
816 */
817havethread:
818 ++gd->gd_cnt.v_swtch;
819 --ntd->td_fairq_accum;
820 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
821 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
822 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
823 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
824 }
825havethread_preempted:
826 ;
827 /*
828 * If the new target does not need the MP lock and we are holding it,
829 * release the MP lock. If the new target requires the MP lock we have
830 * already acquired it for the target.
831 *
832 * WARNING: mpheld is stale here.
8a8d5d85 833 */
f9235b6d
MD
834haveidle:
835 KASSERT(ntd->td_critcount,
836 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
837#ifdef SMP
838 if (ntd->td_mpcount == 0 ) {
839 if (MP_LOCK_HELD())
840 cpu_rel_mplock();
841 } else {
a453459d 842 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
843 }
844#endif
94f6d86e
MD
845 if (td != ntd) {
846 ++switch_count;
b2b3ffcd 847#ifdef __x86_64__
f9235b6d
MD
848 {
849 int tos_ok __debugvar = jg_tos_ok(ntd);
850 KKASSERT(tos_ok);
851 }
85514115 852#endif
a1f0fb66 853 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 854 td->td_switch(ntd);
94f6d86e 855 }
37af14fe
MD
856 /* NOTE: current cpu may have changed after switch */
857 crit_exit_quick(td);
8ad65e08
MD
858}
859
b68b7282 860/*
96728c05
MD
861 * Request that the target thread preempt the current thread. Preemption
862 * only works under a specific set of conditions:
b68b7282 863 *
96728c05
MD
864 * - We are not preempting ourselves
865 * - The target thread is owned by the current cpu
866 * - We are not currently being preempted
867 * - The target is not currently being preempted
d3d1cbc8
MD
868 * - We are not holding any spin locks
869 * - The target thread is not holding any tokens
96728c05
MD
870 * - We are able to satisfy the target's MP lock requirements (if any).
871 *
872 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
873 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 874 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
875 * to determine whether preemption is possible (aka usually just the crit
876 * priority of lwkt_schedule() itself).
b68b7282 877 *
26a0694b
MD
878 * XXX at the moment we run the target thread in a critical section during
879 * the preemption in order to prevent the target from taking interrupts
880 * that *WE* can't. Preemption is strictly limited to interrupt threads
881 * and interrupt-like threads, outside of a critical section, and the
882 * preempted source thread will be resumed the instant the target blocks
883 * whether or not the source is scheduled (i.e. preemption is supposed to
884 * be as transparent as possible).
4b5f931b 885 *
8a8d5d85
MD
886 * The target thread inherits our MP count (added to its own) for the
887 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
888 * MP lock during the preemption. Therefore, any preempting targets must be
889 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
890 * out of sync with the physical mp_lock, but we do not have to preserve
891 * the original ownership of the lock if it was out of synch (that is, we
892 * can leave it synchronized on return).
b68b7282
MD
893 */
894void
f9235b6d 895lwkt_preempt(thread_t ntd, int critcount)
b68b7282 896{
46a3f46d 897 struct globaldata *gd = mycpu;
0a3f9b47 898 thread_t td;
8a8d5d85
MD
899#ifdef SMP
900 int mpheld;
57c254db 901 int savecnt;
8a8d5d85 902#endif
b68b7282 903
26a0694b 904 /*
96728c05
MD
905 * The caller has put us in a critical section. We can only preempt
906 * if the caller of the caller was not in a critical section (basically
f9235b6d 907 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 908 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 909 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
910 *
911 * YYY The target thread must be in a critical section (else it must
912 * inherit our critical section? I dunno yet).
41a01a4d 913 *
0a3f9b47 914 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 915 */
f9235b6d 916 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 917
0a3f9b47 918 td = gd->gd_curthread;
f9235b6d 919 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
920 ++preempt_miss;
921 return;
922 }
f9235b6d 923 if (td->td_critcount > critcount) {
96728c05 924 ++preempt_miss;
8ec60c3f 925 need_lwkt_resched();
96728c05
MD
926 return;
927 }
928#ifdef SMP
46a3f46d 929 if (ntd->td_gd != gd) {
96728c05 930 ++preempt_miss;
8ec60c3f 931 need_lwkt_resched();
96728c05
MD
932 return;
933 }
934#endif
41a01a4d 935 /*
d3d1cbc8 936 * Take the easy way out and do not preempt if we are holding
d666840a 937 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
938 * preempted interlock against the target thread's tokens and whether
939 * we can get all the target thread's tokens, but this situation
940 * should not occur very often so its easier to simply not preempt.
d666840a
MD
941 * Also, plain spinlocks are impossible to figure out at this point so
942 * just don't preempt.
d3d1cbc8
MD
943 *
944 * Do not try to preempt if the target thread is holding any tokens.
945 * We could try to acquire the tokens but this case is so rare there
946 * is no need to support it.
41a01a4d 947 */
bbb31c5d 948 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 949 ++preempt_miss;
8ec60c3f 950 need_lwkt_resched();
41a01a4d
MD
951 return;
952 }
3b998fa9 953 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
954 ++preempt_miss;
955 need_lwkt_resched();
956 return;
957 }
26a0694b
MD
958 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
959 ++preempt_weird;
8ec60c3f 960 need_lwkt_resched();
26a0694b
MD
961 return;
962 }
963 if (ntd->td_preempted) {
4b5f931b 964 ++preempt_hit;
8ec60c3f 965 need_lwkt_resched();
26a0694b 966 return;
b68b7282 967 }
8a8d5d85 968#ifdef SMP
a2a5ad0d
MD
969 /*
970 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
971 * to or from the MP lock. In this case td_mpcount will be pre-disposed
972 * (non-zero) but not actually synchronized with the actual state of the
973 * lock. We can use it to imply an MP lock requirement for the
974 * preemption but we cannot use it to test whether we hold the MP lock
975 * or not.
a2a5ad0d 976 */
96728c05 977 savecnt = td->td_mpcount;
71ef2f5c 978 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
979 ntd->td_mpcount += td->td_mpcount;
980 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
981 ntd->td_mpcount -= td->td_mpcount;
982 ++preempt_miss;
8ec60c3f 983 need_lwkt_resched();
8a8d5d85
MD
984 return;
985 }
986#endif
26a0694b 987
8ec60c3f
MD
988 /*
989 * Since we are able to preempt the current thread, there is no need to
990 * call need_lwkt_resched().
991 */
26a0694b
MD
992 ++preempt_hit;
993 ntd->td_preempted = td;
994 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 995 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
26a0694b 996 td->td_switch(ntd);
b9eb1c19 997
26a0694b 998 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
999#ifdef SMP
1000 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
1001 mpheld = MP_LOCK_HELD();
1002 if (mpheld && td->td_mpcount == 0)
96728c05 1003 cpu_rel_mplock();
71ef2f5c 1004 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
1005 panic("lwkt_preempt(): MP lock was not held through");
1006#endif
26a0694b
MD
1007 ntd->td_preempted = NULL;
1008 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1009}
1010
f1d1c3fa 1011/*
faaeffac 1012 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 1013 *
faaeffac
MD
1014 * td_nest_count prevents deep nesting via splz() or doreti() which
1015 * might otherwise blow out the kernel stack. Note that except for
1016 * this special case, we MUST call splz() here to handle any
1017 * pending ints, particularly after we switch, or we might accidently
1018 * halt the cpu with interrupts pending.
ef0fdad1 1019 *
f1d1c3fa
MD
1020 * (self contained on a per cpu basis)
1021 */
1022void
faaeffac 1023splz_check(void)
f1d1c3fa 1024{
7966cb69
MD
1025 globaldata_t gd = mycpu;
1026 thread_t td = gd->gd_curthread;
ef0fdad1 1027
f9235b6d 1028 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
f1d1c3fa 1029 splz();
f1d1c3fa
MD
1030}
1031
8ad65e08 1032/*
f9235b6d
MD
1033 * This function is used to negotiate a passive release of the current
1034 * process/lwp designation with the user scheduler, allowing the user
1035 * scheduler to schedule another user thread. The related kernel thread
1036 * (curthread) continues running in the released state.
8ad65e08
MD
1037 */
1038void
f9235b6d 1039lwkt_passive_release(struct thread *td)
8ad65e08 1040{
f9235b6d
MD
1041 struct lwp *lp = td->td_lwp;
1042
1043 td->td_release = NULL;
1044 lwkt_setpri_self(TDPRI_KERN_USER);
1045 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1046}
1047
f9235b6d 1048
3824f392 1049/*
f9235b6d
MD
1050 * This implements a normal yield. This routine is virtually a nop if
1051 * there is nothing to yield to but it will always run any pending interrupts
1052 * if called from a critical section.
1053 *
1054 * This yield is designed for kernel threads without a user context.
1055 *
1056 * (self contained on a per cpu basis)
3824f392
MD
1057 */
1058void
f9235b6d 1059lwkt_yield(void)
3824f392 1060{
f9235b6d
MD
1061 globaldata_t gd = mycpu;
1062 thread_t td = gd->gd_curthread;
1063 thread_t xtd;
3824f392 1064
f9235b6d
MD
1065 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1066 splz();
1067 if (td->td_fairq_accum < 0) {
1068 lwkt_schedule_self(curthread);
1069 lwkt_switch();
1070 } else {
1071 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1072 if (xtd && xtd->td_pri > td->td_pri) {
1073 lwkt_schedule_self(curthread);
1074 lwkt_switch();
1075 }
1076 }
3824f392
MD
1077}
1078
1079/*
f9235b6d
MD
1080 * This yield is designed for kernel threads with a user context.
1081 *
1082 * The kernel acting on behalf of the user is potentially cpu-bound,
1083 * this function will efficiently allow other threads to run and also
1084 * switch to other processes by releasing.
3824f392
MD
1085 *
1086 * The lwkt_user_yield() function is designed to have very low overhead
1087 * if no yield is determined to be needed.
1088 */
1089void
1090lwkt_user_yield(void)
1091{
f9235b6d
MD
1092 globaldata_t gd = mycpu;
1093 thread_t td = gd->gd_curthread;
1094
1095 /*
1096 * Always run any pending interrupts in case we are in a critical
1097 * section.
1098 */
1099 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1100 splz();
3824f392
MD
1101
1102#ifdef SMP
1103 /*
1104 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1105 * kernel can prevent other cpus from servicing interrupt threads
1106 * which still require the MP lock (which is a lot of them). This
1107 * has a chaining effect since if the interrupt is blocked, so is
1108 * the event, so normal scheduling will not pick up on the problem.
1109 */
684a93c4
MD
1110 if (mp_lock_contention_mask && td->td_mpcount) {
1111 yield_mplock(td);
3824f392
MD
1112 }
1113#endif
1114
1115 /*
f9235b6d
MD
1116 * Switch (which forces a release) if another kernel thread needs
1117 * the cpu, if userland wants us to resched, or if our kernel
1118 * quantum has run out.
3824f392 1119 */
f9235b6d
MD
1120 if (lwkt_resched_wanted() ||
1121 user_resched_wanted() ||
1122 td->td_fairq_accum < 0)
1123 {
3824f392 1124 lwkt_switch();
3824f392
MD
1125 }
1126
f9235b6d 1127#if 0
3824f392 1128 /*
f9235b6d
MD
1129 * Reacquire the current process if we are released.
1130 *
1131 * XXX not implemented atm. The kernel may be holding locks and such,
1132 * so we want the thread to continue to receive cpu.
3824f392 1133 */
f9235b6d
MD
1134 if (td->td_release == NULL && lp) {
1135 lp->lwp_proc->p_usched->acquire_curproc(lp);
1136 td->td_release = lwkt_passive_release;
1137 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1138 }
f9235b6d 1139#endif
b9eb1c19
MD
1140}
1141
8ad65e08 1142/*
f1d1c3fa
MD
1143 * Generic schedule. Possibly schedule threads belonging to other cpus and
1144 * deal with threads that might be blocked on a wait queue.
1145 *
0a3f9b47
MD
1146 * We have a little helper inline function which does additional work after
1147 * the thread has been enqueued, including dealing with preemption and
1148 * setting need_lwkt_resched() (which prevents the kernel from returning
1149 * to userland until it has processed higher priority threads).
6330a558
MD
1150 *
1151 * It is possible for this routine to be called after a failed _enqueue
1152 * (due to the target thread migrating, sleeping, or otherwise blocked).
1153 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1154 *
1155 * reschedok is an optimized constant propagated from lwkt_schedule() or
1156 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1157 * reschedule to be requested if the target thread has a higher priority.
1158 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1159 * be 0, prevented undesired reschedules.
8ad65e08 1160 */
0a3f9b47
MD
1161static __inline
1162void
f9235b6d 1163_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1164{
b9eb1c19 1165 thread_t otd;
c730be20 1166
6330a558 1167 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1168 if (ntd->td_preemptable && reschedok) {
f9235b6d 1169 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1170 } else if (reschedok) {
b9eb1c19 1171 otd = curthread;
f9235b6d 1172 if (ntd->td_pri > otd->td_pri)
c730be20 1173 need_lwkt_resched();
6330a558 1174 }
f9235b6d
MD
1175
1176 /*
1177 * Give the thread a little fair share scheduler bump if it
1178 * has been asleep for a while. This is primarily to avoid
1179 * a degenerate case for interrupt threads where accumulator
1180 * crosses into negative territory unnecessarily.
1181 */
1182 if (ntd->td_fairq_lticks != ticks) {
1183 ntd->td_fairq_lticks = ticks;
1184 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1185 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1186 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1187 }
0a3f9b47
MD
1188 }
1189}
1190
361d01dd 1191static __inline
8ad65e08 1192void
361d01dd 1193_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1194{
37af14fe
MD
1195 globaldata_t mygd = mycpu;
1196
41a01a4d 1197 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1198 crit_enter_gd(mygd);
9388413d 1199 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1200 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1201 _lwkt_enqueue(td);
1202 } else {
f1d1c3fa 1203 /*
7cd8d145
MD
1204 * If we own the thread, there is no race (since we are in a
1205 * critical section). If we do not own the thread there might
1206 * be a race but the target cpu will deal with it.
f1d1c3fa 1207 */
0f7a3396 1208#ifdef SMP
7cd8d145 1209 if (td->td_gd == mygd) {
9d265729 1210 _lwkt_enqueue(td);
f9235b6d 1211 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1212 } else {
e381e77c 1213 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1214 }
0f7a3396 1215#else
7cd8d145 1216 _lwkt_enqueue(td);
f9235b6d 1217 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1218#endif
8ad65e08 1219 }
37af14fe 1220 crit_exit_gd(mygd);
8ad65e08
MD
1221}
1222
361d01dd
MD
1223void
1224lwkt_schedule(thread_t td)
1225{
1226 _lwkt_schedule(td, 1);
1227}
1228
1229void
1230lwkt_schedule_noresched(thread_t td)
1231{
1232 _lwkt_schedule(td, 0);
1233}
1234
88ebb169
SW
1235#ifdef SMP
1236
e381e77c
MD
1237/*
1238 * When scheduled remotely if frame != NULL the IPIQ is being
1239 * run via doreti or an interrupt then preemption can be allowed.
1240 *
1241 * To allow preemption we have to drop the critical section so only
1242 * one is present in _lwkt_schedule_post.
1243 */
1244static void
1245lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1246{
1247 thread_t td = curthread;
1248 thread_t ntd = arg;
1249
1250 if (frame && ntd->td_preemptable) {
1251 crit_exit_noyield(td);
1252 _lwkt_schedule(ntd, 1);
1253 crit_enter_quick(td);
1254 } else {
1255 _lwkt_schedule(ntd, 1);
1256 }
1257}
1258
d9eea1a5 1259/*
52eedfb5
MD
1260 * Thread migration using a 'Pull' method. The thread may or may not be
1261 * the current thread. It MUST be descheduled and in a stable state.
1262 * lwkt_giveaway() must be called on the cpu owning the thread.
1263 *
1264 * At any point after lwkt_giveaway() is called, the target cpu may
1265 * 'pull' the thread by calling lwkt_acquire().
1266 *
ae8e83e6
MD
1267 * We have to make sure the thread is not sitting on a per-cpu tsleep
1268 * queue or it will blow up when it moves to another cpu.
1269 *
52eedfb5 1270 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1271 */
52eedfb5
MD
1272void
1273lwkt_giveaway(thread_t td)
1274{
3b4192fb 1275 globaldata_t gd = mycpu;
52eedfb5 1276
3b4192fb
MD
1277 crit_enter_gd(gd);
1278 if (td->td_flags & TDF_TSLEEPQ)
1279 tsleep_remove(td);
1280 KKASSERT(td->td_gd == gd);
1281 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1282 td->td_flags |= TDF_MIGRATING;
1283 crit_exit_gd(gd);
52eedfb5
MD
1284}
1285
a2a5ad0d
MD
1286void
1287lwkt_acquire(thread_t td)
1288{
37af14fe
MD
1289 globaldata_t gd;
1290 globaldata_t mygd;
a2a5ad0d 1291
52eedfb5 1292 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1293 gd = td->td_gd;
37af14fe 1294 mygd = mycpu;
52eedfb5 1295 if (gd != mycpu) {
35238fa5 1296 cpu_lfence();
52eedfb5 1297 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1298 crit_enter_gd(mygd);
df910c23
MD
1299 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1300#ifdef SMP
1301 lwkt_process_ipiq();
1302#endif
52eedfb5 1303 cpu_lfence();
df910c23 1304 }
37af14fe 1305 td->td_gd = mygd;
52eedfb5
MD
1306 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1307 td->td_flags &= ~TDF_MIGRATING;
1308 crit_exit_gd(mygd);
1309 } else {
1310 crit_enter_gd(mygd);
1311 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1312 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1313 crit_exit_gd(mygd);
a2a5ad0d
MD
1314 }
1315}
1316
52eedfb5
MD
1317#endif
1318
f1d1c3fa
MD
1319/*
1320 * Generic deschedule. Descheduling threads other then your own should be
1321 * done only in carefully controlled circumstances. Descheduling is
1322 * asynchronous.
1323 *
1324 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1325 */
1326void
1327lwkt_deschedule(thread_t td)
1328{
f1d1c3fa 1329 crit_enter();
b8a98473 1330#ifdef SMP
f1d1c3fa
MD
1331 if (td == curthread) {
1332 _lwkt_dequeue(td);
1333 } else {
a72187e9 1334 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1335 _lwkt_dequeue(td);
1336 } else {
b8a98473 1337 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1338 }
1339 }
b8a98473
MD
1340#else
1341 _lwkt_dequeue(td);
1342#endif
f1d1c3fa
MD
1343 crit_exit();
1344}
1345
4b5f931b
MD
1346/*
1347 * Set the target thread's priority. This routine does not automatically
1348 * switch to a higher priority thread, LWKT threads are not designed for
1349 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1350 */
1351void
1352lwkt_setpri(thread_t td, int pri)
1353{
a72187e9 1354 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1355 if (td->td_pri != pri) {
1356 KKASSERT(pri >= 0);
1357 crit_enter();
1358 if (td->td_flags & TDF_RUNQ) {
1359 _lwkt_dequeue(td);
1360 td->td_pri = pri;
1361 _lwkt_enqueue(td);
1362 } else {
1363 td->td_pri = pri;
1364 }
1365 crit_exit();
26a0694b 1366 }
26a0694b
MD
1367}
1368
03bd0a5e
MD
1369/*
1370 * Set the initial priority for a thread prior to it being scheduled for
1371 * the first time. The thread MUST NOT be scheduled before or during
1372 * this call. The thread may be assigned to a cpu other then the current
1373 * cpu.
1374 *
1375 * Typically used after a thread has been created with TDF_STOPPREQ,
1376 * and before the thread is initially scheduled.
1377 */
1378void
1379lwkt_setpri_initial(thread_t td, int pri)
1380{
1381 KKASSERT(pri >= 0);
1382 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1383 td->td_pri = pri;
03bd0a5e
MD
1384}
1385
26a0694b
MD
1386void
1387lwkt_setpri_self(int pri)
1388{
1389 thread_t td = curthread;
1390
4b5f931b
MD
1391 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1392 crit_enter();
1393 if (td->td_flags & TDF_RUNQ) {
1394 _lwkt_dequeue(td);
f9235b6d 1395 td->td_pri = pri;
4b5f931b
MD
1396 _lwkt_enqueue(td);
1397 } else {
f9235b6d 1398 td->td_pri = pri;
4b5f931b
MD
1399 }
1400 crit_exit();
1401}
1402
f9235b6d
MD
1403/*
1404 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1405 *
1406 * Example: two competing threads, same priority N. decrement by (2*N)
1407 * increment by N*8, each thread will get 4 ticks.
1408 */
1409void
1410lwkt_fairq_schedulerclock(thread_t td)
1411{
1412 if (fairq_enable) {
1413 while (td) {
1414 if (td != &td->td_gd->gd_idlethread) {
1415 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1416 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1417 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1418 if (td->td_fairq_accum < 0)
1419 need_lwkt_resched();
1420 td->td_fairq_lticks = ticks;
1421 }
1422 td = td->td_preempted;
1423 }
1424 }
1425}
1426
1427static void
1428lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1429{
1430 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1431 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1432 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1433}
1434
5d21b981 1435/*
52eedfb5
MD
1436 * Migrate the current thread to the specified cpu.
1437 *
1438 * This is accomplished by descheduling ourselves from the current cpu,
1439 * moving our thread to the tdallq of the target cpu, IPI messaging the
1440 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1441 * races while the thread is being migrated.
ae8e83e6
MD
1442 *
1443 * We must be sure to remove ourselves from the current cpu's tsleepq
1444 * before potentially moving to another queue. The thread can be on
1445 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1446 */
3d28ff59 1447#ifdef SMP
5d21b981 1448static void lwkt_setcpu_remote(void *arg);
3d28ff59 1449#endif
5d21b981
MD
1450
1451void
1452lwkt_setcpu_self(globaldata_t rgd)
1453{
1454#ifdef SMP
1455 thread_t td = curthread;
1456
1457 if (td->td_gd != rgd) {
1458 crit_enter_quick(td);
ae8e83e6 1459 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1460 tsleep_remove(td);
5d21b981
MD
1461 td->td_flags |= TDF_MIGRATING;
1462 lwkt_deschedule_self(td);
52eedfb5 1463 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1464 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1465 lwkt_switch();
1466 /* we are now on the target cpu */
52eedfb5 1467 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1468 crit_exit_quick(td);
1469 }
1470#endif
1471}
1472
ecdefdda
MD
1473void
1474lwkt_migratecpu(int cpuid)
1475{
1476#ifdef SMP
1477 globaldata_t rgd;
1478
1479 rgd = globaldata_find(cpuid);
1480 lwkt_setcpu_self(rgd);
1481#endif
1482}
1483
5d21b981
MD
1484/*
1485 * Remote IPI for cpu migration (called while in a critical section so we
1486 * do not have to enter another one). The thread has already been moved to
1487 * our cpu's allq, but we must wait for the thread to be completely switched
1488 * out on the originating cpu before we schedule it on ours or the stack
1489 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1490 * change to main memory.
1491 *
1492 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1493 * against wakeups. It is best if this interface is used only when there
1494 * are no pending events that might try to schedule the thread.
1495 */
3d28ff59 1496#ifdef SMP
5d21b981
MD
1497static void
1498lwkt_setcpu_remote(void *arg)
1499{
1500 thread_t td = arg;
1501 globaldata_t gd = mycpu;
1502
df910c23
MD
1503 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1504#ifdef SMP
1505 lwkt_process_ipiq();
1506#endif
35238fa5 1507 cpu_lfence();
df910c23 1508 }
5d21b981 1509 td->td_gd = gd;
35238fa5 1510 cpu_sfence();
5d21b981 1511 td->td_flags &= ~TDF_MIGRATING;
9388413d 1512 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1513 _lwkt_enqueue(td);
1514}
3d28ff59 1515#endif
5d21b981 1516
553ea3c8 1517struct lwp *
4b5f931b
MD
1518lwkt_preempted_proc(void)
1519{
73e4f7b9 1520 thread_t td = curthread;
4b5f931b
MD
1521 while (td->td_preempted)
1522 td = td->td_preempted;
553ea3c8 1523 return(td->td_lwp);
4b5f931b
MD
1524}
1525
99df837e
MD
1526/*
1527 * Create a kernel process/thread/whatever. It shares it's address space
1528 * with proc0 - ie: kernel only.
1529 *
365fa13f
MD
1530 * NOTE! By default new threads are created with the MP lock held. A
1531 * thread which does not require the MP lock should release it by calling
1532 * rel_mplock() at the start of the new thread.
99df837e
MD
1533 */
1534int
1535lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1536 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1537 const char *fmt, ...)
99df837e 1538{
73e4f7b9 1539 thread_t td;
e2565a42 1540 __va_list ap;
99df837e 1541
d3d32139 1542 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1543 tdflags);
a2a5ad0d
MD
1544 if (tdp)
1545 *tdp = td;
709799ea 1546 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1547
1548 /*
1549 * Set up arg0 for 'ps' etc
1550 */
e2565a42 1551 __va_start(ap, fmt);
379210cb 1552 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1553 __va_end(ap);
99df837e
MD
1554
1555 /*
1556 * Schedule the thread to run
1557 */
ef0fdad1
MD
1558 if ((td->td_flags & TDF_STOPREQ) == 0)
1559 lwkt_schedule(td);
1560 else
1561 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1562 return 0;
1563}
1564
1565/*
1566 * Destroy an LWKT thread. Warning! This function is not called when
1567 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1568 * uses a different reaping mechanism.
1569 */
1570void
1571lwkt_exit(void)
1572{
1573 thread_t td = curthread;
c070746a 1574 thread_t std;
8826f33a 1575 globaldata_t gd;
99df837e
MD
1576
1577 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1578 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1579 caps_exit(td);
c070746a
MD
1580
1581 /*
1582 * Get us into a critical section to interlock gd_freetd and loop
1583 * until we can get it freed.
1584 *
1585 * We have to cache the current td in gd_freetd because objcache_put()ing
1586 * it would rip it out from under us while our thread is still active.
1587 */
1588 gd = mycpu;
37af14fe 1589 crit_enter_quick(td);
c070746a
MD
1590 while ((std = gd->gd_freetd) != NULL) {
1591 gd->gd_freetd = NULL;
1592 objcache_put(thread_cache, std);
1593 }
3b4192fb
MD
1594
1595 /*
1596 * Remove thread resources from kernel lists and deschedule us for
1597 * the last time.
1598 */
1599 if (td->td_flags & TDF_TSLEEPQ)
1600 tsleep_remove(td);
79eae878 1601 biosched_done(td);
f8abf63c 1602 dsched_exit_thread(td);
37af14fe 1603 lwkt_deschedule_self(td);
e56e4dea 1604 lwkt_remove_tdallq(td);
c070746a
MD
1605 if (td->td_flags & TDF_ALLOCATED_THREAD)
1606 gd->gd_freetd = td;
99df837e
MD
1607 cpu_thread_exit();
1608}
1609
e56e4dea
MD
1610void
1611lwkt_remove_tdallq(thread_t td)
1612{
1613 KKASSERT(td->td_gd == mycpu);
1614 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1615}
1616
2d93b37a
MD
1617void
1618crit_panic(void)
1619{
1620 thread_t td = curthread;
1621 int lpri = td->td_pri;
1622
1623 td->td_pri = 0;
1624 panic("td_pri is/would-go negative! %p %d", td, lpri);
1625}
1626
d165e668
MD
1627#ifdef SMP
1628
bd8015ca
MD
1629/*
1630 * Called from debugger/panic on cpus which have been stopped. We must still
1631 * process the IPIQ while stopped, even if we were stopped while in a critical
1632 * section (XXX).
1633 *
1634 * If we are dumping also try to process any pending interrupts. This may
1635 * or may not work depending on the state of the cpu at the point it was
1636 * stopped.
1637 */
1638void
1639lwkt_smp_stopped(void)
1640{
1641 globaldata_t gd = mycpu;
1642
1643 crit_enter_gd(gd);
1644 if (dumping) {
1645 lwkt_process_ipiq();
1646 splz();
1647 } else {
1648 lwkt_process_ipiq();
1649 }
1650 crit_exit_gd(gd);
1651}
1652
d165e668 1653#endif