Fix the SOEKRIS build (kernel without I686_CPU).
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf 2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
47#include <sys/queue.h>
7d0bac62 48#include <sys/sysctl.h>
99df837e 49#include <sys/kthread.h>
f1d1c3fa 50#include <machine/cpu.h>
99df837e 51#include <sys/lock.h>
f6bf3af1 52#include <sys/caps.h>
9d265729 53#include <sys/spinlock.h>
57aa743c 54#include <sys/ktr.h>
9d265729
MD
55
56#include <sys/thread2.h>
57#include <sys/spinlock2.h>
f1d1c3fa 58
7d0bac62
MD
59#include <vm/vm.h>
60#include <vm/vm_param.h>
61#include <vm/vm_kern.h>
62#include <vm/vm_object.h>
63#include <vm/vm_page.h>
64#include <vm/vm_map.h>
65#include <vm/vm_pager.h>
66#include <vm/vm_extern.h>
7d0bac62 67
99df837e 68#include <machine/stdarg.h>
96728c05 69#include <machine/smp.h>
99df837e 70
d850923c
AE
71#if !defined(KTR_CTXSW)
72#define KTR_CTXSW KTR_ALL
73#endif
74KTR_INFO_MASTER(ctxsw);
75KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "sw %p > %p", 2 * sizeof(struct thread *));
76KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
1541028a 77
40aaf5fc
NT
78static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
79
3824f392
MD
80#ifdef SMP
81static int mplock_countx = 0;
82#endif
0f7a3396
MD
83#ifdef INVARIANTS
84static int panic_on_cscount = 0;
85#endif
05220613
MD
86static __int64_t switch_count = 0;
87static __int64_t preempt_hit = 0;
88static __int64_t preempt_miss = 0;
89static __int64_t preempt_weird = 0;
38717797
HP
90static __int64_t token_contention_count = 0;
91static __int64_t mplock_contention_count = 0;
fb0f29c4 92static int lwkt_use_spin_port;
d2f86ad2 93#ifdef SMP
b9eb1c19 94static int chain_mplock = 0;
e381e77c 95static int bgl_yield = 10;
d2f86ad2 96#endif
40aaf5fc 97static struct objcache *thread_cache;
05220613 98
b9eb1c19
MD
99volatile cpumask_t mp_lock_contention_mask;
100
88ebb169 101#ifdef SMP
e381e77c 102static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 103#endif
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
b9eb1c19
MD
142#ifdef SMP
143SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
e381e77c 144SYSCTL_INT(_lwkt, OID_AUTO, bgl_yield_delay, CTLFLAG_RW, &bgl_yield, 0, "");
b9eb1c19 145#endif
4b5f931b 146SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 147SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 148SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 149SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
150#ifdef INVARIANTS
151SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
152 &token_contention_count, 0, "spinning due to token contention");
153SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
154 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
155#endif
05220613 156
57aa743c
MD
157/*
158 * Kernel Trace
159 */
57aa743c
MD
160#if !defined(KTR_GIANT_CONTENTION)
161#define KTR_GIANT_CONTENTION KTR_ALL
162#endif
163
164KTR_INFO_MASTER(giant);
165KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
166KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
167
168#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
169
4b5f931b
MD
170/*
171 * These helper procedures handle the runq, they can only be called from
172 * within a critical section.
75cdbe6c
MD
173 *
174 * WARNING! Prior to SMP being brought up it is possible to enqueue and
175 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
176 * instead of 'mycpu' when referencing the globaldata structure. Once
177 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 178 */
f1d1c3fa
MD
179static __inline
180void
181_lwkt_dequeue(thread_t td)
182{
183 if (td->td_flags & TDF_RUNQ) {
4b5f931b 184 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 185 struct globaldata *gd = td->td_gd;
4b5f931b 186
f1d1c3fa 187 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
188 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
189 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
190 }
191}
192
193static __inline
194void
195_lwkt_enqueue(thread_t td)
196{
7f5d7ed7 197 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
4b5f931b 198 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 199 struct globaldata *gd = td->td_gd;
4b5f931b 200
f1d1c3fa 201 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
202 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
203 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
204 }
205}
8ad65e08 206
40aaf5fc
NT
207static __boolean_t
208_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
209{
210 struct thread *td = (struct thread *)obj;
211
212 td->td_kstack = NULL;
213 td->td_kstack_size = 0;
214 td->td_flags = TDF_ALLOCATED_THREAD;
215 return (1);
216}
217
218static void
219_lwkt_thread_dtor(void *obj, void *privdata)
220{
221 struct thread *td = (struct thread *)obj;
222
223 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
224 ("_lwkt_thread_dtor: not allocated from objcache"));
225 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
226 td->td_kstack_size > 0,
227 ("_lwkt_thread_dtor: corrupted stack"));
228 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
229}
230
231/*
232 * Initialize the lwkt s/system.
233 */
234void
235lwkt_init(void)
236{
237 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
238 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
239 NULL, CACHE_NTHREADS/2,
240 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
241}
242
37af14fe
MD
243/*
244 * Schedule a thread to run. As the current thread we can always safely
245 * schedule ourselves, and a shortcut procedure is provided for that
246 * function.
247 *
248 * (non-blocking, self contained on a per cpu basis)
249 */
250void
251lwkt_schedule_self(thread_t td)
252{
253 crit_enter_quick(td);
37af14fe 254 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 255 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 256 _lwkt_enqueue(td);
37af14fe
MD
257 crit_exit_quick(td);
258}
259
260/*
261 * Deschedule a thread.
262 *
263 * (non-blocking, self contained on a per cpu basis)
264 */
265void
266lwkt_deschedule_self(thread_t td)
267{
268 crit_enter_quick(td);
37af14fe
MD
269 _lwkt_dequeue(td);
270 crit_exit_quick(td);
271}
272
8ad65e08
MD
273/*
274 * LWKTs operate on a per-cpu basis
275 *
73e4f7b9 276 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
277 */
278void
279lwkt_gdinit(struct globaldata *gd)
280{
4b5f931b
MD
281 int i;
282
283 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
284 TAILQ_INIT(&gd->gd_tdrunq[i]);
285 gd->gd_runqmask = 0;
73e4f7b9 286 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
287}
288
7d0bac62
MD
289/*
290 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
291 * or LWKT start address before it can be scheduled. If the target cpu is
292 * -1 the thread will be created on the current cpu.
0cfcada1
MD
293 *
294 * If you intend to create a thread without a process context this function
295 * does everything except load the startup and switcher function.
7d0bac62
MD
296 */
297thread_t
d3d32139 298lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 299{
c070746a 300 globaldata_t gd = mycpu;
99df837e 301 void *stack;
7d0bac62 302
c070746a
MD
303 /*
304 * If static thread storage is not supplied allocate a thread. Reuse
305 * a cached free thread if possible. gd_freetd is used to keep an exiting
306 * thread intact through the exit.
307 */
ef0fdad1 308 if (td == NULL) {
c070746a
MD
309 if ((td = gd->gd_freetd) != NULL)
310 gd->gd_freetd = NULL;
311 else
312 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
313 KASSERT((td->td_flags &
314 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
315 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
316 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 317 }
c070746a
MD
318
319 /*
320 * Try to reuse cached stack.
321 */
f470d0c8
MD
322 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
323 if (flags & TDF_ALLOCATED_STACK) {
e4846942 324 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
325 stack = NULL;
326 }
327 }
328 if (stack == NULL) {
e4846942 329 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 330 flags |= TDF_ALLOCATED_STACK;
99df837e 331 }
75cdbe6c 332 if (cpu < 0)
c070746a 333 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 334 else
f470d0c8 335 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 336 return(td);
7d0bac62
MD
337}
338
339/*
340 * Initialize a preexisting thread structure. This function is used by
341 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
342 *
f8c3996b
MD
343 * All threads start out in a critical section at a priority of
344 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
345 * appropriate. This function may send an IPI message when the
346 * requested cpu is not the current cpu and consequently gd_tdallq may
347 * not be initialized synchronously from the point of view of the originating
348 * cpu.
349 *
350 * NOTE! we have to be careful in regards to creating threads for other cpus
351 * if SMP has not yet been activated.
7d0bac62 352 */
41a01a4d
MD
353#ifdef SMP
354
75cdbe6c
MD
355static void
356lwkt_init_thread_remote(void *arg)
357{
358 thread_t td = arg;
359
52eedfb5
MD
360 /*
361 * Protected by critical section held by IPI dispatch
362 */
75cdbe6c
MD
363 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
364}
365
41a01a4d
MD
366#endif
367
7d0bac62 368void
f470d0c8
MD
369lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
370 struct globaldata *gd)
7d0bac62 371{
37af14fe
MD
372 globaldata_t mygd = mycpu;
373
99df837e
MD
374 bzero(td, sizeof(struct thread));
375 td->td_kstack = stack;
f470d0c8 376 td->td_kstack_size = stksize;
d3d32139 377 td->td_flags = flags;
26a0694b 378 td->td_gd = gd;
f8c3996b 379 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
380#ifdef SMP
381 if ((flags & TDF_MPSAFE) == 0)
382 td->td_mpcount = 1;
383#endif
fb0f29c4
MD
384 if (lwkt_use_spin_port)
385 lwkt_initport_spin(&td->td_msgport);
386 else
387 lwkt_initport_thread(&td->td_msgport, td);
99df837e 388 pmap_init_thread(td);
0f7a3396 389#ifdef SMP
5d21b981
MD
390 /*
391 * Normally initializing a thread for a remote cpu requires sending an
392 * IPI. However, the idlethread is setup before the other cpus are
393 * activated so we have to treat it as a special case. XXX manipulation
394 * of gd_tdallq requires the BGL.
395 */
396 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 397 crit_enter_gd(mygd);
75cdbe6c 398 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 399 crit_exit_gd(mygd);
75cdbe6c 400 } else {
2db3b277 401 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 402 }
0f7a3396 403#else
37af14fe 404 crit_enter_gd(mygd);
0f7a3396 405 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 406 crit_exit_gd(mygd);
0f7a3396 407#endif
73e4f7b9
MD
408}
409
410void
411lwkt_set_comm(thread_t td, const char *ctl, ...)
412{
e2565a42 413 __va_list va;
73e4f7b9 414
e2565a42 415 __va_start(va, ctl);
379210cb 416 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 417 __va_end(va);
7d0bac62
MD
418}
419
99df837e 420void
73e4f7b9 421lwkt_hold(thread_t td)
99df837e 422{
73e4f7b9
MD
423 ++td->td_refs;
424}
425
426void
427lwkt_rele(thread_t td)
428{
429 KKASSERT(td->td_refs > 0);
430 --td->td_refs;
431}
432
433void
434lwkt_wait_free(thread_t td)
435{
436 while (td->td_refs)
377d4740 437 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
438}
439
440void
441lwkt_free_thread(thread_t td)
442{
d9eea1a5 443 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
444 ("lwkt_free_thread: did not exit! %p", td));
445
40aaf5fc
NT
446 if (td->td_flags & TDF_ALLOCATED_THREAD) {
447 objcache_put(thread_cache, td);
448 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
449 /* client-allocated struct with internally allocated stack */
450 KASSERT(td->td_kstack && td->td_kstack_size > 0,
451 ("lwkt_free_thread: corrupted stack"));
452 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
453 td->td_kstack = NULL;
454 td->td_kstack_size = 0;
99df837e
MD
455 }
456}
457
458
8ad65e08
MD
459/*
460 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
461 * switch to the idlethread. Switching must occur within a critical
462 * section to avoid races with the scheduling queue.
463 *
464 * We always have full control over our cpu's run queue. Other cpus
465 * that wish to manipulate our queue must use the cpu_*msg() calls to
466 * talk to our cpu, so a critical section is all that is needed and
467 * the result is very, very fast thread switching.
468 *
96728c05
MD
469 * The LWKT scheduler uses a fixed priority model and round-robins at
470 * each priority level. User process scheduling is a totally
471 * different beast and LWKT priorities should not be confused with
472 * user process priorities.
f1d1c3fa 473 *
96728c05
MD
474 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
475 * cleans it up. Note that the td_switch() function cannot do anything that
476 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
477 * the target thread (not the current thread). It's nice to have a scheduler
478 * that does not need the MP lock to work because it allows us to do some
479 * really cool high-performance MP lock optimizations.
69d78e99
MD
480 *
481 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
482 * is not called by the current thread in the preemption case, only when
483 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
484 */
485void
486lwkt_switch(void)
487{
37af14fe
MD
488 globaldata_t gd = mycpu;
489 thread_t td = gd->gd_curthread;
8ad65e08 490 thread_t ntd;
8a8d5d85
MD
491#ifdef SMP
492 int mpheld;
493#endif
8ad65e08 494
46a3f46d 495 /*
27e88a6e
MD
496 * Switching from within a 'fast' (non thread switched) interrupt or IPI
497 * is illegal. However, we may have to do it anyway if we hit a fatal
498 * kernel trap or we have paniced.
499 *
500 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 501 */
27e88a6e
MD
502 if (gd->gd_intr_nesting_level) {
503 int savegdnest;
504 int savegdtrap;
505
506 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
507 panic("lwkt_switch: cannot switch from within "
508 "a fast interrupt, yet, td %p\n", td);
509 } else {
510 savegdnest = gd->gd_intr_nesting_level;
511 savegdtrap = gd->gd_trap_nesting_level;
512 gd->gd_intr_nesting_level = 0;
513 gd->gd_trap_nesting_level = 0;
a7422615
MD
514 if ((td->td_flags & TDF_PANICWARN) == 0) {
515 td->td_flags |= TDF_PANICWARN;
6ea70f76 516 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 517 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 518 print_backtrace();
a7422615 519 }
27e88a6e
MD
520 lwkt_switch();
521 gd->gd_intr_nesting_level = savegdnest;
522 gd->gd_trap_nesting_level = savegdtrap;
523 return;
524 }
96728c05 525 }
ef0fdad1 526
cb973d15
MD
527 /*
528 * Passive release (used to transition from user to kernel mode
529 * when we block or switch rather then when we enter the kernel).
530 * This function is NOT called if we are switching into a preemption
531 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
532 * our current process designation (if we have one) and become a true
533 * LWKT thread, and may also hand the current process designation to
534 * another process and schedule thread.
cb973d15
MD
535 */
536 if (td->td_release)
537 td->td_release(td);
538
37af14fe 539 crit_enter_gd(gd);
9d265729
MD
540 if (td->td_toks)
541 lwkt_relalltokens(td);
542
543 /*
b02926de
MD
544 * We had better not be holding any spin locks, but don't get into an
545 * endless panic loop.
9d265729 546 */
bbb31c5d
MD
547 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
548 ("lwkt_switch: still holding a shared spinlock %p!",
549 gd->gd_spinlock_rd));
d666840a
MD
550 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
551 ("lwkt_switch: still holding %d exclusive spinlocks!",
552 gd->gd_spinlocks_wr));
9d265729 553
8a8d5d85
MD
554
555#ifdef SMP
556 /*
557 * td_mpcount cannot be used to determine if we currently hold the
558 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
559 * to get the lock, and switch out if it can't. Our ownership of
560 * the actual lock will remain stable while we are in a critical section
561 * (but, of course, another cpu may own or release the lock so the
562 * actual value of mp_lock is not stable).
8a8d5d85
MD
563 */
564 mpheld = MP_LOCK_HELD();
0f7a3396
MD
565#ifdef INVARIANTS
566 if (td->td_cscount) {
6ea70f76 567 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
568 td);
569 if (panic_on_cscount)
570 panic("switching while mastering cpusync");
571 }
572#endif
8a8d5d85 573#endif
99df837e
MD
574 if ((ntd = td->td_preempted) != NULL) {
575 /*
576 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
577 * thread. This occurs transparently, whether the preempted thread
578 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
579 * itself).
580 *
581 * We have to setup the MP lock for the original thread after backing
582 * out the adjustment that was made to curthread when the original
583 * was preempted.
99df837e 584 */
26a0694b 585 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 586#ifdef SMP
96728c05 587 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 588 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
589 td, ntd, td->td_mpcount, ntd->td_mpcount);
590 }
8a8d5d85
MD
591 if (ntd->td_mpcount) {
592 td->td_mpcount -= ntd->td_mpcount;
593 KKASSERT(td->td_mpcount >= 0);
594 }
595#endif
26a0694b 596 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
597
598 /*
b9eb1c19
MD
599 * The interrupt may have woken a thread up, we need to properly
600 * set the reschedule flag if the originally interrupted thread is
601 * at a lower priority.
8ec60c3f
MD
602 */
603 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
604 need_lwkt_resched();
8a8d5d85 605 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 606 } else {
4b5f931b
MD
607 /*
608 * Priority queue / round-robin at each priority. Note that user
609 * processes run at a fixed, low priority and the user process
610 * scheduler deals with interactions between user processes
611 * by scheduling and descheduling them from the LWKT queue as
612 * necessary.
8a8d5d85
MD
613 *
614 * We have to adjust the MP lock for the target thread. If we
615 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
616 * thread that does not need the MP lock. If we cannot, we spin
617 * instead of HLT.
618 *
619 * A similar issue exists for the tokens held by the target thread.
620 * If we cannot obtain ownership of the tokens we cannot immediately
621 * schedule the thread.
622 */
623
8ec60c3f
MD
624 /*
625 * If an LWKT reschedule was requested, well that is what we are
626 * doing now so clear it.
627 */
628 clear_lwkt_resched();
4b5f931b
MD
629again:
630 if (gd->gd_runqmask) {
631 int nq = bsrl(gd->gd_runqmask);
632 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
633 gd->gd_runqmask &= ~(1 << nq);
634 goto again;
635 }
8a8d5d85 636#ifdef SMP
41a01a4d 637 /*
df6b8ba0
MD
638 * THREAD SELECTION FOR AN SMP MACHINE BUILD
639 *
41a01a4d
MD
640 * If the target needs the MP lock and we couldn't get it,
641 * or if the target is holding tokens and we could not
642 * gain ownership of the tokens, continue looking for a
643 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
644 *
645 * NOTE: the mpheld variable invalid after this conditional, it
646 * can change due to both cpu_try_mplock() returning success
9d265729 647 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
648 * we are trying to check the mpcount of a thread other then
649 * the current thread. Because of this, if the current thread
650 * is not holding td_mpcount, an IPI indirectly run via
9d265729 651 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 652 * cause the core MP lock to be released.
41a01a4d
MD
653 */
654 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 655 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 656 ) {
8a8d5d85 657 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
658
659 mpheld = MP_LOCK_HELD();
660 ntd = NULL;
8a8d5d85
MD
661 while (rqmask) {
662 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 663 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 664 /* spinning due to MP lock being held */
38717797 665#ifdef INVARIANTS
a453459d 666 ++mplock_contention_count;
38717797 667#endif
a453459d 668 /* mplock still not held, 'mpheld' still valid */
41a01a4d 669 continue;
38717797 670 }
a453459d
MD
671
672 /*
9d265729 673 * mpheld state invalid after getalltokens call returns
a453459d
MD
674 * failure, but the variable is only needed for
675 * the loop.
676 */
9d265729 677 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 678 /* spinning due to token contention */
38717797 679#ifdef INVARIANTS
a453459d 680 ++token_contention_count;
38717797 681#endif
a453459d 682 mpheld = MP_LOCK_HELD();
41a01a4d 683 continue;
38717797 684 }
41a01a4d 685 break;
8a8d5d85
MD
686 }
687 if (ntd)
688 break;
689 rqmask &= ~(1 << nq);
690 nq = bsrl(rqmask);
b9eb1c19
MD
691
692 /*
693 * We have two choices. We can either refuse to run a
694 * user thread when a kernel thread needs the MP lock
695 * but could not get it, or we can allow it to run but
696 * then expect an IPI (hopefully) later on to force a
697 * reschedule when the MP lock might become available.
698 */
699 if (nq < TDPRI_KERN_LPSCHED) {
700 if (chain_mplock == 0)
701 break;
702 atomic_set_int(&mp_lock_contention_mask,
703 gd->gd_cpumask);
704 /* continue loop, allow user threads to be scheduled */
705 }
8a8d5d85
MD
706 }
707 if (ntd == NULL) {
b402c633 708 cpu_mplock_contested();
a2a5ad0d
MD
709 ntd = &gd->gd_idlethread;
710 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 711 goto using_idle_thread;
8a8d5d85 712 } else {
344ad853 713 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
714 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
715 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
716 }
717 } else {
3824f392
MD
718 if (ntd->td_mpcount)
719 ++mplock_countx;
344ad853 720 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
721 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
722 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
723 }
724#else
df6b8ba0
MD
725 /*
726 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
727 * worry about tokens or the BGL. However, we still have
728 * to call lwkt_getalltokens() in order to properly detect
729 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 730 */
7eb611ef 731 lwkt_getalltokens(ntd);
344ad853 732 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
733 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
734 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 735#endif
4b5f931b 736 } else {
3c23a41a 737 /*
60f945af
MD
738 * We have nothing to run but only let the idle loop halt
739 * the cpu if there are no pending interrupts.
3c23a41a 740 */
a2a5ad0d 741 ntd = &gd->gd_idlethread;
60f945af 742 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 743 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 744#ifdef SMP
df6b8ba0
MD
745using_idle_thread:
746 /*
747 * The idle thread should not be holding the MP lock unless we
748 * are trapping in the kernel or in a panic. Since we select the
749 * idle thread unconditionally when no other thread is available,
750 * if the MP lock is desired during a panic or kernel trap, we
751 * have to loop in the scheduler until we get it.
752 */
753 if (ntd->td_mpcount) {
754 mpheld = MP_LOCK_HELD();
b402c633 755 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 756 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
757 } else if (mpheld == 0) {
758 cpu_mplock_contested();
df6b8ba0 759 goto again;
b402c633 760 }
df6b8ba0 761 }
a453459d 762#endif
4b5f931b 763 }
f1d1c3fa 764 }
26a0694b
MD
765 KASSERT(ntd->td_pri >= TDPRI_CRIT,
766 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
767
768 /*
769 * Do the actual switch. If the new target does not need the MP lock
770 * and we are holding it, release the MP lock. If the new target requires
771 * the MP lock we have already acquired it for the target.
772 */
773#ifdef SMP
774 if (ntd->td_mpcount == 0 ) {
775 if (MP_LOCK_HELD())
776 cpu_rel_mplock();
777 } else {
a453459d 778 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
779 }
780#endif
94f6d86e
MD
781 if (td != ntd) {
782 ++switch_count;
b2b3ffcd 783#ifdef __x86_64__
0855a2af 784 KKASSERT(jg_tos_ok(ntd));
85514115 785#endif
d850923c 786 KTR_LOG(ctxsw_sw, td, ntd);
f1d1c3fa 787 td->td_switch(ntd);
94f6d86e 788 }
37af14fe
MD
789 /* NOTE: current cpu may have changed after switch */
790 crit_exit_quick(td);
8ad65e08
MD
791}
792
b68b7282 793/*
96728c05
MD
794 * Request that the target thread preempt the current thread. Preemption
795 * only works under a specific set of conditions:
b68b7282 796 *
96728c05
MD
797 * - We are not preempting ourselves
798 * - The target thread is owned by the current cpu
799 * - We are not currently being preempted
800 * - The target is not currently being preempted
d3d1cbc8
MD
801 * - We are not holding any spin locks
802 * - The target thread is not holding any tokens
96728c05
MD
803 * - We are able to satisfy the target's MP lock requirements (if any).
804 *
805 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
806 * this is called via lwkt_schedule() through the td_preemptable callback.
807 * critpri is the managed critical priority that we should ignore in order
808 * to determine whether preemption is possible (aka usually just the crit
809 * priority of lwkt_schedule() itself).
b68b7282 810 *
26a0694b
MD
811 * XXX at the moment we run the target thread in a critical section during
812 * the preemption in order to prevent the target from taking interrupts
813 * that *WE* can't. Preemption is strictly limited to interrupt threads
814 * and interrupt-like threads, outside of a critical section, and the
815 * preempted source thread will be resumed the instant the target blocks
816 * whether or not the source is scheduled (i.e. preemption is supposed to
817 * be as transparent as possible).
4b5f931b 818 *
8a8d5d85
MD
819 * The target thread inherits our MP count (added to its own) for the
820 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
821 * MP lock during the preemption. Therefore, any preempting targets must be
822 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
823 * out of sync with the physical mp_lock, but we do not have to preserve
824 * the original ownership of the lock if it was out of synch (that is, we
825 * can leave it synchronized on return).
b68b7282
MD
826 */
827void
96728c05 828lwkt_preempt(thread_t ntd, int critpri)
b68b7282 829{
46a3f46d 830 struct globaldata *gd = mycpu;
0a3f9b47 831 thread_t td;
8a8d5d85
MD
832#ifdef SMP
833 int mpheld;
57c254db 834 int savecnt;
8a8d5d85 835#endif
b68b7282 836
26a0694b 837 /*
96728c05
MD
838 * The caller has put us in a critical section. We can only preempt
839 * if the caller of the caller was not in a critical section (basically
d666840a 840 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 841 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 842 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
843 *
844 * YYY The target thread must be in a critical section (else it must
845 * inherit our critical section? I dunno yet).
41a01a4d 846 *
0a3f9b47 847 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
848 */
849 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 850
0a3f9b47 851 td = gd->gd_curthread;
0a3f9b47 852 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
853 ++preempt_miss;
854 return;
855 }
96728c05
MD
856 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
857 ++preempt_miss;
8ec60c3f 858 need_lwkt_resched();
96728c05
MD
859 return;
860 }
861#ifdef SMP
46a3f46d 862 if (ntd->td_gd != gd) {
96728c05 863 ++preempt_miss;
8ec60c3f 864 need_lwkt_resched();
96728c05
MD
865 return;
866 }
867#endif
41a01a4d 868 /*
d3d1cbc8 869 * Take the easy way out and do not preempt if we are holding
d666840a 870 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
871 * preempted interlock against the target thread's tokens and whether
872 * we can get all the target thread's tokens, but this situation
873 * should not occur very often so its easier to simply not preempt.
d666840a
MD
874 * Also, plain spinlocks are impossible to figure out at this point so
875 * just don't preempt.
d3d1cbc8
MD
876 *
877 * Do not try to preempt if the target thread is holding any tokens.
878 * We could try to acquire the tokens but this case is so rare there
879 * is no need to support it.
41a01a4d 880 */
bbb31c5d 881 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 882 ++preempt_miss;
8ec60c3f 883 need_lwkt_resched();
41a01a4d
MD
884 return;
885 }
d3d1cbc8
MD
886 if (ntd->td_toks) {
887 ++preempt_miss;
888 need_lwkt_resched();
889 return;
890 }
26a0694b
MD
891 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
892 ++preempt_weird;
8ec60c3f 893 need_lwkt_resched();
26a0694b
MD
894 return;
895 }
896 if (ntd->td_preempted) {
4b5f931b 897 ++preempt_hit;
8ec60c3f 898 need_lwkt_resched();
26a0694b 899 return;
b68b7282 900 }
8a8d5d85 901#ifdef SMP
a2a5ad0d
MD
902 /*
903 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
904 * to or from the MP lock. In this case td_mpcount will be pre-disposed
905 * (non-zero) but not actually synchronized with the actual state of the
906 * lock. We can use it to imply an MP lock requirement for the
907 * preemption but we cannot use it to test whether we hold the MP lock
908 * or not.
a2a5ad0d 909 */
96728c05 910 savecnt = td->td_mpcount;
71ef2f5c 911 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
912 ntd->td_mpcount += td->td_mpcount;
913 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
914 ntd->td_mpcount -= td->td_mpcount;
915 ++preempt_miss;
8ec60c3f 916 need_lwkt_resched();
8a8d5d85
MD
917 return;
918 }
919#endif
26a0694b 920
8ec60c3f
MD
921 /*
922 * Since we are able to preempt the current thread, there is no need to
923 * call need_lwkt_resched().
924 */
26a0694b
MD
925 ++preempt_hit;
926 ntd->td_preempted = td;
927 td->td_flags |= TDF_PREEMPT_LOCK;
d850923c 928 KTR_LOG(ctxsw_pre, td, ntd);
26a0694b 929 td->td_switch(ntd);
b9eb1c19 930
26a0694b 931 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
932#ifdef SMP
933 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
934 mpheld = MP_LOCK_HELD();
935 if (mpheld && td->td_mpcount == 0)
96728c05 936 cpu_rel_mplock();
71ef2f5c 937 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
938 panic("lwkt_preempt(): MP lock was not held through");
939#endif
26a0694b
MD
940 ntd->td_preempted = NULL;
941 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
942}
943
f1d1c3fa 944/*
faaeffac 945 * Conditionally call splz() if gd_reqflags indicates work is pending.
f1d1c3fa 946 *
faaeffac
MD
947 * td_nest_count prevents deep nesting via splz() or doreti() which
948 * might otherwise blow out the kernel stack. Note that except for
949 * this special case, we MUST call splz() here to handle any
950 * pending ints, particularly after we switch, or we might accidently
951 * halt the cpu with interrupts pending.
ef0fdad1 952 *
f1d1c3fa
MD
953 * (self contained on a per cpu basis)
954 */
955void
faaeffac 956splz_check(void)
f1d1c3fa 957{
7966cb69
MD
958 globaldata_t gd = mycpu;
959 thread_t td = gd->gd_curthread;
ef0fdad1 960
46a3f46d 961 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 962 splz();
f1d1c3fa
MD
963}
964
8ad65e08 965/*
faaeffac
MD
966 * This implements a normal yield which will yield to equal priority
967 * threads as well as higher priority threads. Note that gd_reqflags
968 * tests will be handled by the crit_exit() call in lwkt_switch().
f1d1c3fa
MD
969 *
970 * (self contained on a per cpu basis)
8ad65e08
MD
971 */
972void
f1d1c3fa 973lwkt_yield(void)
8ad65e08 974{
37af14fe 975 lwkt_schedule_self(curthread);
f1d1c3fa
MD
976 lwkt_switch();
977}
978
3824f392
MD
979/*
980 * This function is used along with the lwkt_passive_recover() inline
981 * by the trap code to negotiate a passive release of the current
982 * process/lwp designation with the user scheduler.
983 */
984void
985lwkt_passive_release(struct thread *td)
986{
987 struct lwp *lp = td->td_lwp;
988
989 td->td_release = NULL;
990 lwkt_setpri_self(TDPRI_KERN_USER);
991 lp->lwp_proc->p_usched->release_curproc(lp);
992}
993
994/*
995 * Make a kernel thread act as if it were in user mode with regards
996 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
997 * loops which may be potentially cpu-bound can call lwkt_user_yield().
998 *
999 * The lwkt_user_yield() function is designed to have very low overhead
1000 * if no yield is determined to be needed.
1001 */
1002void
1003lwkt_user_yield(void)
1004{
1005 thread_t td = curthread;
1006 struct lwp *lp = td->td_lwp;
1007
1008#ifdef SMP
1009 /*
1010 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1011 * kernel can prevent other cpus from servicing interrupt threads
1012 * which still require the MP lock (which is a lot of them). This
1013 * has a chaining effect since if the interrupt is blocked, so is
1014 * the event, so normal scheduling will not pick up on the problem.
1015 */
1016 if (mplock_countx && td->td_mpcount) {
1017 int savecnt = td->td_mpcount;
1018
1019 td->td_mpcount = 1;
e381e77c 1020 mplock_countx = 0;
3824f392 1021 rel_mplock();
e381e77c 1022 DELAY(bgl_yield);
3824f392
MD
1023 get_mplock();
1024 td->td_mpcount = savecnt;
3824f392
MD
1025 }
1026#endif
1027
1028 /*
1029 * Another kernel thread wants the cpu
1030 */
1031 if (lwkt_resched_wanted())
1032 lwkt_switch();
1033
1034 /*
1035 * If the user scheduler has asynchronously determined that the current
1036 * process (when running in user mode) needs to lose the cpu then make
1037 * sure we are released.
1038 */
1039 if (user_resched_wanted()) {
1040 if (td->td_release)
1041 td->td_release(td);
1042 }
1043
1044 /*
1045 * If we are released reduce our priority
1046 */
1047 if (td->td_release == NULL) {
1048 if (lwkt_check_resched(td) > 0)
1049 lwkt_switch();
e381e77c
MD
1050 if (lp) {
1051 lp->lwp_proc->p_usched->acquire_curproc(lp);
1052 td->td_release = lwkt_passive_release;
1053 lwkt_setpri_self(TDPRI_USER_NORM);
1054 }
3824f392
MD
1055 }
1056}
1057
b9eb1c19
MD
1058/*
1059 * Return 0 if no runnable threads are pending at the same or higher
1060 * priority as the passed thread.
1061 *
1062 * Return 1 if runnable threads are pending at the same priority.
1063 *
1064 * Return 2 if runnable threads are pending at a higher priority.
1065 */
1066int
1067lwkt_check_resched(thread_t td)
1068{
1069 int pri = td->td_pri & TDPRI_MASK;
1070
1071 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1072 return(2);
1073 if (TAILQ_NEXT(td, td_threadq))
1074 return(1);
1075 return(0);
1076}
1077
8ad65e08 1078/*
f1d1c3fa
MD
1079 * Generic schedule. Possibly schedule threads belonging to other cpus and
1080 * deal with threads that might be blocked on a wait queue.
1081 *
0a3f9b47
MD
1082 * We have a little helper inline function which does additional work after
1083 * the thread has been enqueued, including dealing with preemption and
1084 * setting need_lwkt_resched() (which prevents the kernel from returning
1085 * to userland until it has processed higher priority threads).
6330a558
MD
1086 *
1087 * It is possible for this routine to be called after a failed _enqueue
1088 * (due to the target thread migrating, sleeping, or otherwise blocked).
1089 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1090 *
1091 * reschedok is an optimized constant propagated from lwkt_schedule() or
1092 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1093 * reschedule to be requested if the target thread has a higher priority.
1094 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1095 * be 0, prevented undesired reschedules.
8ad65e08 1096 */
0a3f9b47
MD
1097static __inline
1098void
361d01dd 1099_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1100{
b9eb1c19 1101 thread_t otd;
c730be20 1102
6330a558 1103 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1104 if (ntd->td_preemptable && reschedok) {
6330a558 1105 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1106 } else if (reschedok) {
b9eb1c19
MD
1107 otd = curthread;
1108 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1109 need_lwkt_resched();
6330a558 1110 }
0a3f9b47
MD
1111 }
1112}
1113
361d01dd 1114static __inline
8ad65e08 1115void
361d01dd 1116_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1117{
37af14fe
MD
1118 globaldata_t mygd = mycpu;
1119
41a01a4d 1120 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1121 crit_enter_gd(mygd);
9388413d 1122 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1123 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1124 _lwkt_enqueue(td);
1125 } else {
f1d1c3fa 1126 /*
7cd8d145
MD
1127 * If we own the thread, there is no race (since we are in a
1128 * critical section). If we do not own the thread there might
1129 * be a race but the target cpu will deal with it.
f1d1c3fa 1130 */
0f7a3396 1131#ifdef SMP
7cd8d145 1132 if (td->td_gd == mygd) {
9d265729 1133 _lwkt_enqueue(td);
361d01dd 1134 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1135 } else {
e381e77c 1136 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1137 }
0f7a3396 1138#else
7cd8d145 1139 _lwkt_enqueue(td);
361d01dd 1140 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1141#endif
8ad65e08 1142 }
37af14fe 1143 crit_exit_gd(mygd);
8ad65e08
MD
1144}
1145
361d01dd
MD
1146void
1147lwkt_schedule(thread_t td)
1148{
1149 _lwkt_schedule(td, 1);
1150}
1151
1152void
1153lwkt_schedule_noresched(thread_t td)
1154{
1155 _lwkt_schedule(td, 0);
1156}
1157
88ebb169
SW
1158#ifdef SMP
1159
e381e77c
MD
1160/*
1161 * When scheduled remotely if frame != NULL the IPIQ is being
1162 * run via doreti or an interrupt then preemption can be allowed.
1163 *
1164 * To allow preemption we have to drop the critical section so only
1165 * one is present in _lwkt_schedule_post.
1166 */
1167static void
1168lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1169{
1170 thread_t td = curthread;
1171 thread_t ntd = arg;
1172
1173 if (frame && ntd->td_preemptable) {
1174 crit_exit_noyield(td);
1175 _lwkt_schedule(ntd, 1);
1176 crit_enter_quick(td);
1177 } else {
1178 _lwkt_schedule(ntd, 1);
1179 }
1180}
1181
d9eea1a5 1182/*
52eedfb5
MD
1183 * Thread migration using a 'Pull' method. The thread may or may not be
1184 * the current thread. It MUST be descheduled and in a stable state.
1185 * lwkt_giveaway() must be called on the cpu owning the thread.
1186 *
1187 * At any point after lwkt_giveaway() is called, the target cpu may
1188 * 'pull' the thread by calling lwkt_acquire().
1189 *
ae8e83e6
MD
1190 * We have to make sure the thread is not sitting on a per-cpu tsleep
1191 * queue or it will blow up when it moves to another cpu.
1192 *
52eedfb5 1193 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1194 */
52eedfb5
MD
1195void
1196lwkt_giveaway(thread_t td)
1197{
3b4192fb 1198 globaldata_t gd = mycpu;
52eedfb5 1199
3b4192fb
MD
1200 crit_enter_gd(gd);
1201 if (td->td_flags & TDF_TSLEEPQ)
1202 tsleep_remove(td);
1203 KKASSERT(td->td_gd == gd);
1204 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1205 td->td_flags |= TDF_MIGRATING;
1206 crit_exit_gd(gd);
52eedfb5
MD
1207}
1208
a2a5ad0d
MD
1209void
1210lwkt_acquire(thread_t td)
1211{
37af14fe
MD
1212 globaldata_t gd;
1213 globaldata_t mygd;
a2a5ad0d 1214
52eedfb5 1215 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1216 gd = td->td_gd;
37af14fe 1217 mygd = mycpu;
52eedfb5 1218 if (gd != mycpu) {
35238fa5 1219 cpu_lfence();
52eedfb5 1220 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1221 crit_enter_gd(mygd);
df910c23
MD
1222 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1223#ifdef SMP
1224 lwkt_process_ipiq();
1225#endif
52eedfb5 1226 cpu_lfence();
df910c23 1227 }
37af14fe 1228 td->td_gd = mygd;
52eedfb5
MD
1229 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1230 td->td_flags &= ~TDF_MIGRATING;
1231 crit_exit_gd(mygd);
1232 } else {
1233 crit_enter_gd(mygd);
1234 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1235 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1236 crit_exit_gd(mygd);
a2a5ad0d
MD
1237 }
1238}
1239
52eedfb5
MD
1240#endif
1241
f1d1c3fa
MD
1242/*
1243 * Generic deschedule. Descheduling threads other then your own should be
1244 * done only in carefully controlled circumstances. Descheduling is
1245 * asynchronous.
1246 *
1247 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1248 */
1249void
1250lwkt_deschedule(thread_t td)
1251{
f1d1c3fa 1252 crit_enter();
b8a98473 1253#ifdef SMP
f1d1c3fa
MD
1254 if (td == curthread) {
1255 _lwkt_dequeue(td);
1256 } else {
a72187e9 1257 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1258 _lwkt_dequeue(td);
1259 } else {
b8a98473 1260 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1261 }
1262 }
b8a98473
MD
1263#else
1264 _lwkt_dequeue(td);
1265#endif
f1d1c3fa
MD
1266 crit_exit();
1267}
1268
4b5f931b
MD
1269/*
1270 * Set the target thread's priority. This routine does not automatically
1271 * switch to a higher priority thread, LWKT threads are not designed for
1272 * continuous priority changes. Yield if you want to switch.
1273 *
1274 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1275 * of the td_pri field. The specified priority may also indicate zero or
1276 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1277 *
1278 * Note that we requeue the thread whether it winds up on a different runq
1279 * or not. uio_yield() depends on this and the routine is not normally
1280 * called with the same priority otherwise.
4b5f931b
MD
1281 */
1282void
1283lwkt_setpri(thread_t td, int pri)
1284{
26a0694b 1285 KKASSERT(pri >= 0);
a72187e9 1286 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1287 crit_enter();
1288 if (td->td_flags & TDF_RUNQ) {
1289 _lwkt_dequeue(td);
1290 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1291 _lwkt_enqueue(td);
1292 } else {
1293 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1294 }
1295 crit_exit();
1296}
1297
03bd0a5e
MD
1298/*
1299 * Set the initial priority for a thread prior to it being scheduled for
1300 * the first time. The thread MUST NOT be scheduled before or during
1301 * this call. The thread may be assigned to a cpu other then the current
1302 * cpu.
1303 *
1304 * Typically used after a thread has been created with TDF_STOPPREQ,
1305 * and before the thread is initially scheduled.
1306 */
1307void
1308lwkt_setpri_initial(thread_t td, int pri)
1309{
1310 KKASSERT(pri >= 0);
1311 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1312 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1313}
1314
26a0694b
MD
1315void
1316lwkt_setpri_self(int pri)
1317{
1318 thread_t td = curthread;
1319
4b5f931b
MD
1320 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1321 crit_enter();
1322 if (td->td_flags & TDF_RUNQ) {
1323 _lwkt_dequeue(td);
1324 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1325 _lwkt_enqueue(td);
1326 } else {
1327 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1328 }
1329 crit_exit();
1330}
1331
5d21b981 1332/*
52eedfb5
MD
1333 * Migrate the current thread to the specified cpu.
1334 *
1335 * This is accomplished by descheduling ourselves from the current cpu,
1336 * moving our thread to the tdallq of the target cpu, IPI messaging the
1337 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1338 * races while the thread is being migrated.
ae8e83e6
MD
1339 *
1340 * We must be sure to remove ourselves from the current cpu's tsleepq
1341 * before potentially moving to another queue. The thread can be on
1342 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1343 */
3d28ff59 1344#ifdef SMP
5d21b981 1345static void lwkt_setcpu_remote(void *arg);
3d28ff59 1346#endif
5d21b981
MD
1347
1348void
1349lwkt_setcpu_self(globaldata_t rgd)
1350{
1351#ifdef SMP
1352 thread_t td = curthread;
1353
1354 if (td->td_gd != rgd) {
1355 crit_enter_quick(td);
ae8e83e6 1356 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1357 tsleep_remove(td);
5d21b981
MD
1358 td->td_flags |= TDF_MIGRATING;
1359 lwkt_deschedule_self(td);
52eedfb5 1360 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1361 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1362 lwkt_switch();
1363 /* we are now on the target cpu */
52eedfb5 1364 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1365 crit_exit_quick(td);
1366 }
1367#endif
1368}
1369
ecdefdda
MD
1370void
1371lwkt_migratecpu(int cpuid)
1372{
1373#ifdef SMP
1374 globaldata_t rgd;
1375
1376 rgd = globaldata_find(cpuid);
1377 lwkt_setcpu_self(rgd);
1378#endif
1379}
1380
5d21b981
MD
1381/*
1382 * Remote IPI for cpu migration (called while in a critical section so we
1383 * do not have to enter another one). The thread has already been moved to
1384 * our cpu's allq, but we must wait for the thread to be completely switched
1385 * out on the originating cpu before we schedule it on ours or the stack
1386 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1387 * change to main memory.
1388 *
1389 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1390 * against wakeups. It is best if this interface is used only when there
1391 * are no pending events that might try to schedule the thread.
1392 */
3d28ff59 1393#ifdef SMP
5d21b981
MD
1394static void
1395lwkt_setcpu_remote(void *arg)
1396{
1397 thread_t td = arg;
1398 globaldata_t gd = mycpu;
1399
df910c23
MD
1400 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1401#ifdef SMP
1402 lwkt_process_ipiq();
1403#endif
35238fa5 1404 cpu_lfence();
df910c23 1405 }
5d21b981 1406 td->td_gd = gd;
35238fa5 1407 cpu_sfence();
5d21b981 1408 td->td_flags &= ~TDF_MIGRATING;
9388413d 1409 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1410 _lwkt_enqueue(td);
1411}
3d28ff59 1412#endif
5d21b981 1413
553ea3c8 1414struct lwp *
4b5f931b
MD
1415lwkt_preempted_proc(void)
1416{
73e4f7b9 1417 thread_t td = curthread;
4b5f931b
MD
1418 while (td->td_preempted)
1419 td = td->td_preempted;
553ea3c8 1420 return(td->td_lwp);
4b5f931b
MD
1421}
1422
99df837e
MD
1423/*
1424 * Create a kernel process/thread/whatever. It shares it's address space
1425 * with proc0 - ie: kernel only.
1426 *
365fa13f
MD
1427 * NOTE! By default new threads are created with the MP lock held. A
1428 * thread which does not require the MP lock should release it by calling
1429 * rel_mplock() at the start of the new thread.
99df837e
MD
1430 */
1431int
1432lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1433 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1434 const char *fmt, ...)
99df837e 1435{
73e4f7b9 1436 thread_t td;
e2565a42 1437 __va_list ap;
99df837e 1438
d3d32139 1439 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1440 tdflags);
a2a5ad0d
MD
1441 if (tdp)
1442 *tdp = td;
709799ea 1443 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1444
1445 /*
1446 * Set up arg0 for 'ps' etc
1447 */
e2565a42 1448 __va_start(ap, fmt);
379210cb 1449 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1450 __va_end(ap);
99df837e
MD
1451
1452 /*
1453 * Schedule the thread to run
1454 */
ef0fdad1
MD
1455 if ((td->td_flags & TDF_STOPREQ) == 0)
1456 lwkt_schedule(td);
1457 else
1458 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1459 return 0;
1460}
1461
1462/*
1463 * Destroy an LWKT thread. Warning! This function is not called when
1464 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1465 * uses a different reaping mechanism.
1466 */
1467void
1468lwkt_exit(void)
1469{
1470 thread_t td = curthread;
c070746a 1471 thread_t std;
8826f33a 1472 globaldata_t gd;
99df837e
MD
1473
1474 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1475 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1476 caps_exit(td);
c070746a
MD
1477
1478 /*
1479 * Get us into a critical section to interlock gd_freetd and loop
1480 * until we can get it freed.
1481 *
1482 * We have to cache the current td in gd_freetd because objcache_put()ing
1483 * it would rip it out from under us while our thread is still active.
1484 */
1485 gd = mycpu;
37af14fe 1486 crit_enter_quick(td);
c070746a
MD
1487 while ((std = gd->gd_freetd) != NULL) {
1488 gd->gd_freetd = NULL;
1489 objcache_put(thread_cache, std);
1490 }
3b4192fb
MD
1491
1492 /*
1493 * Remove thread resources from kernel lists and deschedule us for
1494 * the last time.
1495 */
1496 if (td->td_flags & TDF_TSLEEPQ)
1497 tsleep_remove(td);
79eae878 1498 biosched_done(td);
37af14fe 1499 lwkt_deschedule_self(td);
e56e4dea 1500 lwkt_remove_tdallq(td);
c070746a
MD
1501 if (td->td_flags & TDF_ALLOCATED_THREAD)
1502 gd->gd_freetd = td;
99df837e
MD
1503 cpu_thread_exit();
1504}
1505
e56e4dea
MD
1506void
1507lwkt_remove_tdallq(thread_t td)
1508{
1509 KKASSERT(td->td_gd == mycpu);
1510 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1511}
1512
2d93b37a
MD
1513void
1514crit_panic(void)
1515{
1516 thread_t td = curthread;
1517 int lpri = td->td_pri;
1518
1519 td->td_pri = 0;
1520 panic("td_pri is/would-go negative! %p %d", td, lpri);
1521}
1522
d165e668
MD
1523#ifdef SMP
1524
bd8015ca
MD
1525/*
1526 * Called from debugger/panic on cpus which have been stopped. We must still
1527 * process the IPIQ while stopped, even if we were stopped while in a critical
1528 * section (XXX).
1529 *
1530 * If we are dumping also try to process any pending interrupts. This may
1531 * or may not work depending on the state of the cpu at the point it was
1532 * stopped.
1533 */
1534void
1535lwkt_smp_stopped(void)
1536{
1537 globaldata_t gd = mycpu;
1538
1539 crit_enter_gd(gd);
1540 if (dumping) {
1541 lwkt_process_ipiq();
1542 splz();
1543 } else {
1544 lwkt_process_ipiq();
1545 }
1546 crit_exit_gd(gd);
1547}
1548
57aa743c
MD
1549/*
1550 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1551 * get_mplock() has already incremented td_mpcount. We must block and
1552 * not return until giant is held.
1553 *
1554 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1555 * reschedule the thread until it can obtain the giant lock for it.
1556 */
1557void
1558lwkt_mp_lock_contested(void)
1559{
3824f392 1560 ++mplock_countx;
57aa743c 1561 loggiant(beg);
57aa743c 1562 lwkt_switch();
57aa743c 1563 loggiant(end);
57aa743c
MD
1564}
1565
b9eb1c19
MD
1566/*
1567 * The rel_mplock() code will call this function after releasing the
1568 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1569 *
1570 * We then chain an IPI to a single other cpu potentially needing the
1571 * lock. This is a bit heuristical and we can wind up with IPIs flying
1572 * all over the place.
1573 */
1574static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1575
1576void
1577lwkt_mp_lock_uncontested(void)
1578{
1579 globaldata_t gd;
1580 globaldata_t dgd;
1581 cpumask_t mask;
1582 cpumask_t tmpmask;
1583 int cpuid;
1584
1585 if (chain_mplock) {
1586 gd = mycpu;
1587 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1588 mask = mp_lock_contention_mask;
1589 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1590
1591 if (mask) {
1592 if (mask & tmpmask)
1593 cpuid = bsfl(mask & tmpmask);
1594 else
1595 cpuid = bsfl(mask);
1596 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1597 dgd = globaldata_find(cpuid);
1598 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1599 }
1600 }
1601}
1602
1603/*
1604 * The idea is for this IPI to interrupt a potentially lower priority
1605 * thread, such as a user thread, to allow the scheduler to reschedule
1606 * a higher priority kernel thread that needs the MP lock.
1607 *
1608 * For now we set the LWKT reschedule flag which generates an AST in
1609 * doreti, though theoretically it is also possible to possibly preempt
1610 * here if the underlying thread was operating in user mode. Nah.
1611 */
1612static void
1613lwkt_mp_lock_uncontested_remote(void *arg __unused)
1614{
1615 need_lwkt_resched();
1616}
1617
d165e668 1618#endif