Fix issues with the scheduler that were causing unnecessary reschedules
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
8c10bfcf 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
8c10bfcf 33 *
c730be20 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.117 2008/09/09 04:06:13 dillon Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/kernel.h>
47#include <sys/proc.h>
48#include <sys/rtprio.h>
49#include <sys/queue.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f6bf3af1 54#include <sys/caps.h>
9d265729 55#include <sys/spinlock.h>
57aa743c 56#include <sys/ktr.h>
9d265729
MD
57
58#include <sys/thread2.h>
59#include <sys/spinlock2.h>
f1d1c3fa 60
7d0bac62
MD
61#include <vm/vm.h>
62#include <vm/vm_param.h>
63#include <vm/vm_kern.h>
64#include <vm/vm_object.h>
65#include <vm/vm_page.h>
66#include <vm/vm_map.h>
67#include <vm/vm_pager.h>
68#include <vm/vm_extern.h>
7d0bac62 69
99df837e 70#include <machine/stdarg.h>
96728c05 71#include <machine/smp.h>
99df837e 72
40aaf5fc
NT
73static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
74
7d0bac62 75static int untimely_switch = 0;
0f7a3396
MD
76#ifdef INVARIANTS
77static int panic_on_cscount = 0;
78#endif
05220613
MD
79static __int64_t switch_count = 0;
80static __int64_t preempt_hit = 0;
81static __int64_t preempt_miss = 0;
82static __int64_t preempt_weird = 0;
38717797
HP
83static __int64_t token_contention_count = 0;
84static __int64_t mplock_contention_count = 0;
fb0f29c4 85static int lwkt_use_spin_port;
40aaf5fc 86static struct objcache *thread_cache;
05220613 87
fb0f29c4
MD
88/*
89 * We can make all thread ports use the spin backend instead of the thread
90 * backend. This should only be set to debug the spin backend.
91 */
92TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
93
05220613 94SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
95#ifdef INVARIANTS
96SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
97#endif
4b5f931b 98SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 99SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 100SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 101SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
102#ifdef INVARIANTS
103SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
104 &token_contention_count, 0, "spinning due to token contention");
105SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
106 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
107#endif
05220613 108
57aa743c
MD
109/*
110 * Kernel Trace
111 */
57aa743c
MD
112#if !defined(KTR_GIANT_CONTENTION)
113#define KTR_GIANT_CONTENTION KTR_ALL
114#endif
115
116KTR_INFO_MASTER(giant);
117KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
118KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
119
120#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
121
4b5f931b
MD
122/*
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
75cdbe6c
MD
125 *
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 130 */
f1d1c3fa
MD
131static __inline
132void
133_lwkt_dequeue(thread_t td)
134{
135 if (td->td_flags & TDF_RUNQ) {
4b5f931b 136 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 137 struct globaldata *gd = td->td_gd;
4b5f931b 138
f1d1c3fa 139 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
142 }
143}
144
145static __inline
146void
147_lwkt_enqueue(thread_t td)
148{
344ad853 149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 150 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 151 struct globaldata *gd = td->td_gd;
4b5f931b 152
f1d1c3fa 153 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
156 }
157}
8ad65e08 158
40aaf5fc
NT
159static __boolean_t
160_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
161{
162 struct thread *td = (struct thread *)obj;
163
164 td->td_kstack = NULL;
165 td->td_kstack_size = 0;
166 td->td_flags = TDF_ALLOCATED_THREAD;
167 return (1);
168}
169
170static void
171_lwkt_thread_dtor(void *obj, void *privdata)
172{
173 struct thread *td = (struct thread *)obj;
174
175 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
176 ("_lwkt_thread_dtor: not allocated from objcache"));
177 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
178 td->td_kstack_size > 0,
179 ("_lwkt_thread_dtor: corrupted stack"));
180 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
181}
182
183/*
184 * Initialize the lwkt s/system.
185 */
186void
187lwkt_init(void)
188{
189 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
190 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 0,
191 CACHE_NTHREADS/2, _lwkt_thread_ctor, _lwkt_thread_dtor,
192 NULL);
193}
194
37af14fe
MD
195/*
196 * Schedule a thread to run. As the current thread we can always safely
197 * schedule ourselves, and a shortcut procedure is provided for that
198 * function.
199 *
200 * (non-blocking, self contained on a per cpu basis)
201 */
202void
203lwkt_schedule_self(thread_t td)
204{
205 crit_enter_quick(td);
37af14fe 206 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 207 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 208 _lwkt_enqueue(td);
37af14fe
MD
209 crit_exit_quick(td);
210}
211
212/*
213 * Deschedule a thread.
214 *
215 * (non-blocking, self contained on a per cpu basis)
216 */
217void
218lwkt_deschedule_self(thread_t td)
219{
220 crit_enter_quick(td);
37af14fe
MD
221 _lwkt_dequeue(td);
222 crit_exit_quick(td);
223}
224
8ad65e08
MD
225/*
226 * LWKTs operate on a per-cpu basis
227 *
73e4f7b9 228 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
229 */
230void
231lwkt_gdinit(struct globaldata *gd)
232{
4b5f931b
MD
233 int i;
234
235 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
236 TAILQ_INIT(&gd->gd_tdrunq[i]);
237 gd->gd_runqmask = 0;
73e4f7b9 238 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
239}
240
7d0bac62
MD
241/*
242 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
243 * or LWKT start address before it can be scheduled. If the target cpu is
244 * -1 the thread will be created on the current cpu.
0cfcada1
MD
245 *
246 * If you intend to create a thread without a process context this function
247 * does everything except load the startup and switcher function.
7d0bac62
MD
248 */
249thread_t
d3d32139 250lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 251{
c070746a 252 globaldata_t gd = mycpu;
99df837e 253 void *stack;
7d0bac62 254
c070746a
MD
255 /*
256 * If static thread storage is not supplied allocate a thread. Reuse
257 * a cached free thread if possible. gd_freetd is used to keep an exiting
258 * thread intact through the exit.
259 */
ef0fdad1 260 if (td == NULL) {
c070746a
MD
261 if ((td = gd->gd_freetd) != NULL)
262 gd->gd_freetd = NULL;
263 else
264 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
265 KASSERT((td->td_flags &
266 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
267 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
268 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 269 }
c070746a
MD
270
271 /*
272 * Try to reuse cached stack.
273 */
f470d0c8
MD
274 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
275 if (flags & TDF_ALLOCATED_STACK) {
e4846942 276 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
277 stack = NULL;
278 }
279 }
280 if (stack == NULL) {
e4846942 281 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 282 flags |= TDF_ALLOCATED_STACK;
99df837e 283 }
75cdbe6c 284 if (cpu < 0)
c070746a 285 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 286 else
f470d0c8 287 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 288 return(td);
7d0bac62
MD
289}
290
291/*
292 * Initialize a preexisting thread structure. This function is used by
293 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
294 *
f8c3996b
MD
295 * All threads start out in a critical section at a priority of
296 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
297 * appropriate. This function may send an IPI message when the
298 * requested cpu is not the current cpu and consequently gd_tdallq may
299 * not be initialized synchronously from the point of view of the originating
300 * cpu.
301 *
302 * NOTE! we have to be careful in regards to creating threads for other cpus
303 * if SMP has not yet been activated.
7d0bac62 304 */
41a01a4d
MD
305#ifdef SMP
306
75cdbe6c
MD
307static void
308lwkt_init_thread_remote(void *arg)
309{
310 thread_t td = arg;
311
52eedfb5
MD
312 /*
313 * Protected by critical section held by IPI dispatch
314 */
75cdbe6c
MD
315 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
316}
317
41a01a4d
MD
318#endif
319
7d0bac62 320void
f470d0c8
MD
321lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
322 struct globaldata *gd)
7d0bac62 323{
37af14fe
MD
324 globaldata_t mygd = mycpu;
325
99df837e
MD
326 bzero(td, sizeof(struct thread));
327 td->td_kstack = stack;
f470d0c8 328 td->td_kstack_size = stksize;
d3d32139 329 td->td_flags = flags;
26a0694b 330 td->td_gd = gd;
f8c3996b 331 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
332#ifdef SMP
333 if ((flags & TDF_MPSAFE) == 0)
334 td->td_mpcount = 1;
335#endif
fb0f29c4
MD
336 if (lwkt_use_spin_port)
337 lwkt_initport_spin(&td->td_msgport);
338 else
339 lwkt_initport_thread(&td->td_msgport, td);
99df837e 340 pmap_init_thread(td);
0f7a3396 341#ifdef SMP
5d21b981
MD
342 /*
343 * Normally initializing a thread for a remote cpu requires sending an
344 * IPI. However, the idlethread is setup before the other cpus are
345 * activated so we have to treat it as a special case. XXX manipulation
346 * of gd_tdallq requires the BGL.
347 */
348 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 349 crit_enter_gd(mygd);
75cdbe6c 350 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 351 crit_exit_gd(mygd);
75cdbe6c 352 } else {
2db3b277 353 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 354 }
0f7a3396 355#else
37af14fe 356 crit_enter_gd(mygd);
0f7a3396 357 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 358 crit_exit_gd(mygd);
0f7a3396 359#endif
73e4f7b9
MD
360}
361
362void
363lwkt_set_comm(thread_t td, const char *ctl, ...)
364{
e2565a42 365 __va_list va;
73e4f7b9 366
e2565a42 367 __va_start(va, ctl);
379210cb 368 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 369 __va_end(va);
7d0bac62
MD
370}
371
99df837e 372void
73e4f7b9 373lwkt_hold(thread_t td)
99df837e 374{
73e4f7b9
MD
375 ++td->td_refs;
376}
377
378void
379lwkt_rele(thread_t td)
380{
381 KKASSERT(td->td_refs > 0);
382 --td->td_refs;
383}
384
385void
386lwkt_wait_free(thread_t td)
387{
388 while (td->td_refs)
377d4740 389 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
390}
391
392void
393lwkt_free_thread(thread_t td)
394{
d9eea1a5 395 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
396 ("lwkt_free_thread: did not exit! %p", td));
397
40aaf5fc
NT
398 if (td->td_flags & TDF_ALLOCATED_THREAD) {
399 objcache_put(thread_cache, td);
400 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
401 /* client-allocated struct with internally allocated stack */
402 KASSERT(td->td_kstack && td->td_kstack_size > 0,
403 ("lwkt_free_thread: corrupted stack"));
404 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
405 td->td_kstack = NULL;
406 td->td_kstack_size = 0;
99df837e
MD
407 }
408}
409
410
8ad65e08
MD
411/*
412 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
413 * switch to the idlethread. Switching must occur within a critical
414 * section to avoid races with the scheduling queue.
415 *
416 * We always have full control over our cpu's run queue. Other cpus
417 * that wish to manipulate our queue must use the cpu_*msg() calls to
418 * talk to our cpu, so a critical section is all that is needed and
419 * the result is very, very fast thread switching.
420 *
96728c05
MD
421 * The LWKT scheduler uses a fixed priority model and round-robins at
422 * each priority level. User process scheduling is a totally
423 * different beast and LWKT priorities should not be confused with
424 * user process priorities.
f1d1c3fa 425 *
96728c05
MD
426 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
427 * cleans it up. Note that the td_switch() function cannot do anything that
428 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
429 * the target thread (not the current thread). It's nice to have a scheduler
430 * that does not need the MP lock to work because it allows us to do some
431 * really cool high-performance MP lock optimizations.
69d78e99
MD
432 *
433 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
434 * is not called by the current thread in the preemption case, only when
435 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
436 */
437void
438lwkt_switch(void)
439{
37af14fe
MD
440 globaldata_t gd = mycpu;
441 thread_t td = gd->gd_curthread;
8ad65e08 442 thread_t ntd;
8a8d5d85
MD
443#ifdef SMP
444 int mpheld;
445#endif
8ad65e08 446
46a3f46d 447 /*
27e88a6e
MD
448 * Switching from within a 'fast' (non thread switched) interrupt or IPI
449 * is illegal. However, we may have to do it anyway if we hit a fatal
450 * kernel trap or we have paniced.
451 *
452 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 453 */
27e88a6e
MD
454 if (gd->gd_intr_nesting_level) {
455 int savegdnest;
456 int savegdtrap;
457
458 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
459 panic("lwkt_switch: cannot switch from within "
460 "a fast interrupt, yet, td %p\n", td);
461 } else {
462 savegdnest = gd->gd_intr_nesting_level;
463 savegdtrap = gd->gd_trap_nesting_level;
464 gd->gd_intr_nesting_level = 0;
465 gd->gd_trap_nesting_level = 0;
a7422615
MD
466 if ((td->td_flags & TDF_PANICWARN) == 0) {
467 td->td_flags |= TDF_PANICWARN;
6ea70f76 468 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615
MD
469 "thread %p (%s)\n", td, td->td_comm);
470#ifdef DDB
471 db_print_backtrace();
472#endif
473 }
27e88a6e
MD
474 lwkt_switch();
475 gd->gd_intr_nesting_level = savegdnest;
476 gd->gd_trap_nesting_level = savegdtrap;
477 return;
478 }
96728c05 479 }
ef0fdad1 480
cb973d15
MD
481 /*
482 * Passive release (used to transition from user to kernel mode
483 * when we block or switch rather then when we enter the kernel).
484 * This function is NOT called if we are switching into a preemption
485 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
486 * our current process designation (if we have one) and become a true
487 * LWKT thread, and may also hand the current process designation to
488 * another process and schedule thread.
cb973d15
MD
489 */
490 if (td->td_release)
491 td->td_release(td);
492
37af14fe 493 crit_enter_gd(gd);
9d265729
MD
494 if (td->td_toks)
495 lwkt_relalltokens(td);
496
497 /*
b02926de
MD
498 * We had better not be holding any spin locks, but don't get into an
499 * endless panic loop.
9d265729 500 */
bbb31c5d
MD
501 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
502 ("lwkt_switch: still holding a shared spinlock %p!",
503 gd->gd_spinlock_rd));
d666840a
MD
504 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
505 ("lwkt_switch: still holding %d exclusive spinlocks!",
506 gd->gd_spinlocks_wr));
9d265729 507
8a8d5d85
MD
508
509#ifdef SMP
510 /*
511 * td_mpcount cannot be used to determine if we currently hold the
512 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
513 * to get the lock, and switch out if it can't. Our ownership of
514 * the actual lock will remain stable while we are in a critical section
515 * (but, of course, another cpu may own or release the lock so the
516 * actual value of mp_lock is not stable).
8a8d5d85
MD
517 */
518 mpheld = MP_LOCK_HELD();
0f7a3396
MD
519#ifdef INVARIANTS
520 if (td->td_cscount) {
6ea70f76 521 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
522 td);
523 if (panic_on_cscount)
524 panic("switching while mastering cpusync");
525 }
526#endif
8a8d5d85 527#endif
99df837e
MD
528 if ((ntd = td->td_preempted) != NULL) {
529 /*
530 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
531 * thread. This occurs transparently, whether the preempted thread
532 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
533 * itself).
534 *
535 * We have to setup the MP lock for the original thread after backing
536 * out the adjustment that was made to curthread when the original
537 * was preempted.
99df837e 538 */
26a0694b 539 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 540#ifdef SMP
96728c05 541 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 542 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
543 td, ntd, td->td_mpcount, ntd->td_mpcount);
544 }
8a8d5d85
MD
545 if (ntd->td_mpcount) {
546 td->td_mpcount -= ntd->td_mpcount;
547 KKASSERT(td->td_mpcount >= 0);
548 }
549#endif
26a0694b 550 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
551
552 /*
553 * XXX. The interrupt may have woken a thread up, we need to properly
554 * set the reschedule flag if the originally interrupted thread is at
555 * a lower priority.
556 */
557 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
558 need_lwkt_resched();
8a8d5d85 559 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 560 } else {
4b5f931b
MD
561 /*
562 * Priority queue / round-robin at each priority. Note that user
563 * processes run at a fixed, low priority and the user process
564 * scheduler deals with interactions between user processes
565 * by scheduling and descheduling them from the LWKT queue as
566 * necessary.
8a8d5d85
MD
567 *
568 * We have to adjust the MP lock for the target thread. If we
569 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
570 * thread that does not need the MP lock. If we cannot, we spin
571 * instead of HLT.
572 *
573 * A similar issue exists for the tokens held by the target thread.
574 * If we cannot obtain ownership of the tokens we cannot immediately
575 * schedule the thread.
576 */
577
8ec60c3f
MD
578 /*
579 * If an LWKT reschedule was requested, well that is what we are
580 * doing now so clear it.
581 */
582 clear_lwkt_resched();
4b5f931b
MD
583again:
584 if (gd->gd_runqmask) {
585 int nq = bsrl(gd->gd_runqmask);
586 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
587 gd->gd_runqmask &= ~(1 << nq);
588 goto again;
589 }
8a8d5d85 590#ifdef SMP
41a01a4d 591 /*
df6b8ba0
MD
592 * THREAD SELECTION FOR AN SMP MACHINE BUILD
593 *
41a01a4d
MD
594 * If the target needs the MP lock and we couldn't get it,
595 * or if the target is holding tokens and we could not
596 * gain ownership of the tokens, continue looking for a
597 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
598 *
599 * NOTE: the mpheld variable invalid after this conditional, it
600 * can change due to both cpu_try_mplock() returning success
9d265729 601 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
602 * we are trying to check the mpcount of a thread other then
603 * the current thread. Because of this, if the current thread
604 * is not holding td_mpcount, an IPI indirectly run via
9d265729 605 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 606 * cause the core MP lock to be released.
41a01a4d
MD
607 */
608 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 609 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 610 ) {
8a8d5d85 611 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
612
613 mpheld = MP_LOCK_HELD();
614 ntd = NULL;
8a8d5d85
MD
615 while (rqmask) {
616 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 617 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 618 /* spinning due to MP lock being held */
38717797 619#ifdef INVARIANTS
a453459d 620 ++mplock_contention_count;
38717797 621#endif
a453459d 622 /* mplock still not held, 'mpheld' still valid */
41a01a4d 623 continue;
38717797 624 }
a453459d
MD
625
626 /*
9d265729 627 * mpheld state invalid after getalltokens call returns
a453459d
MD
628 * failure, but the variable is only needed for
629 * the loop.
630 */
9d265729 631 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 632 /* spinning due to token contention */
38717797 633#ifdef INVARIANTS
a453459d 634 ++token_contention_count;
38717797 635#endif
a453459d 636 mpheld = MP_LOCK_HELD();
41a01a4d 637 continue;
38717797 638 }
41a01a4d 639 break;
8a8d5d85
MD
640 }
641 if (ntd)
642 break;
643 rqmask &= ~(1 << nq);
644 nq = bsrl(rqmask);
645 }
646 if (ntd == NULL) {
b402c633 647 cpu_mplock_contested();
a2a5ad0d
MD
648 ntd = &gd->gd_idlethread;
649 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 650 goto using_idle_thread;
8a8d5d85 651 } else {
344ad853 652 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
653 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
654 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
655 }
656 } else {
344ad853 657 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
658 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
659 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
660 }
661#else
df6b8ba0
MD
662 /*
663 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
664 * worry about tokens or the BGL. However, we still have
665 * to call lwkt_getalltokens() in order to properly detect
666 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 667 */
7eb611ef 668 lwkt_getalltokens(ntd);
344ad853 669 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
670 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
671 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 672#endif
4b5f931b 673 } else {
3c23a41a 674 /*
60f945af
MD
675 * We have nothing to run but only let the idle loop halt
676 * the cpu if there are no pending interrupts.
3c23a41a 677 */
a2a5ad0d 678 ntd = &gd->gd_idlethread;
60f945af 679 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 680 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 681#ifdef SMP
df6b8ba0
MD
682using_idle_thread:
683 /*
684 * The idle thread should not be holding the MP lock unless we
685 * are trapping in the kernel or in a panic. Since we select the
686 * idle thread unconditionally when no other thread is available,
687 * if the MP lock is desired during a panic or kernel trap, we
688 * have to loop in the scheduler until we get it.
689 */
690 if (ntd->td_mpcount) {
691 mpheld = MP_LOCK_HELD();
b402c633 692 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 693 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
694 } else if (mpheld == 0) {
695 cpu_mplock_contested();
df6b8ba0 696 goto again;
b402c633 697 }
df6b8ba0 698 }
a453459d 699#endif
4b5f931b 700 }
f1d1c3fa 701 }
26a0694b
MD
702 KASSERT(ntd->td_pri >= TDPRI_CRIT,
703 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
704
705 /*
706 * Do the actual switch. If the new target does not need the MP lock
707 * and we are holding it, release the MP lock. If the new target requires
708 * the MP lock we have already acquired it for the target.
709 */
710#ifdef SMP
711 if (ntd->td_mpcount == 0 ) {
712 if (MP_LOCK_HELD())
713 cpu_rel_mplock();
714 } else {
a453459d 715 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
716 }
717#endif
94f6d86e
MD
718 if (td != ntd) {
719 ++switch_count;
f1d1c3fa 720 td->td_switch(ntd);
94f6d86e 721 }
37af14fe
MD
722 /* NOTE: current cpu may have changed after switch */
723 crit_exit_quick(td);
8ad65e08
MD
724}
725
b68b7282 726/*
96728c05
MD
727 * Request that the target thread preempt the current thread. Preemption
728 * only works under a specific set of conditions:
b68b7282 729 *
96728c05
MD
730 * - We are not preempting ourselves
731 * - The target thread is owned by the current cpu
732 * - We are not currently being preempted
733 * - The target is not currently being preempted
d3d1cbc8
MD
734 * - We are not holding any spin locks
735 * - The target thread is not holding any tokens
96728c05
MD
736 * - We are able to satisfy the target's MP lock requirements (if any).
737 *
738 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
739 * this is called via lwkt_schedule() through the td_preemptable callback.
740 * critpri is the managed critical priority that we should ignore in order
741 * to determine whether preemption is possible (aka usually just the crit
742 * priority of lwkt_schedule() itself).
b68b7282 743 *
26a0694b
MD
744 * XXX at the moment we run the target thread in a critical section during
745 * the preemption in order to prevent the target from taking interrupts
746 * that *WE* can't. Preemption is strictly limited to interrupt threads
747 * and interrupt-like threads, outside of a critical section, and the
748 * preempted source thread will be resumed the instant the target blocks
749 * whether or not the source is scheduled (i.e. preemption is supposed to
750 * be as transparent as possible).
4b5f931b 751 *
8a8d5d85
MD
752 * The target thread inherits our MP count (added to its own) for the
753 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
754 * MP lock during the preemption. Therefore, any preempting targets must be
755 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
756 * out of sync with the physical mp_lock, but we do not have to preserve
757 * the original ownership of the lock if it was out of synch (that is, we
758 * can leave it synchronized on return).
b68b7282
MD
759 */
760void
96728c05 761lwkt_preempt(thread_t ntd, int critpri)
b68b7282 762{
46a3f46d 763 struct globaldata *gd = mycpu;
0a3f9b47 764 thread_t td;
8a8d5d85
MD
765#ifdef SMP
766 int mpheld;
57c254db 767 int savecnt;
8a8d5d85 768#endif
b68b7282 769
26a0694b 770 /*
96728c05
MD
771 * The caller has put us in a critical section. We can only preempt
772 * if the caller of the caller was not in a critical section (basically
d666840a 773 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 774 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 775 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
776 *
777 * YYY The target thread must be in a critical section (else it must
778 * inherit our critical section? I dunno yet).
41a01a4d 779 *
0a3f9b47 780 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
781 */
782 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 783
0a3f9b47 784 td = gd->gd_curthread;
0a3f9b47 785 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
786 ++preempt_miss;
787 return;
788 }
96728c05
MD
789 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
790 ++preempt_miss;
8ec60c3f 791 need_lwkt_resched();
96728c05
MD
792 return;
793 }
794#ifdef SMP
46a3f46d 795 if (ntd->td_gd != gd) {
96728c05 796 ++preempt_miss;
8ec60c3f 797 need_lwkt_resched();
96728c05
MD
798 return;
799 }
800#endif
41a01a4d 801 /*
d3d1cbc8 802 * Take the easy way out and do not preempt if we are holding
d666840a 803 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
804 * preempted interlock against the target thread's tokens and whether
805 * we can get all the target thread's tokens, but this situation
806 * should not occur very often so its easier to simply not preempt.
d666840a
MD
807 * Also, plain spinlocks are impossible to figure out at this point so
808 * just don't preempt.
d3d1cbc8
MD
809 *
810 * Do not try to preempt if the target thread is holding any tokens.
811 * We could try to acquire the tokens but this case is so rare there
812 * is no need to support it.
41a01a4d 813 */
bbb31c5d 814 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 815 ++preempt_miss;
8ec60c3f 816 need_lwkt_resched();
41a01a4d
MD
817 return;
818 }
d3d1cbc8
MD
819 if (ntd->td_toks) {
820 ++preempt_miss;
821 need_lwkt_resched();
822 return;
823 }
26a0694b
MD
824 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
825 ++preempt_weird;
8ec60c3f 826 need_lwkt_resched();
26a0694b
MD
827 return;
828 }
829 if (ntd->td_preempted) {
4b5f931b 830 ++preempt_hit;
8ec60c3f 831 need_lwkt_resched();
26a0694b 832 return;
b68b7282 833 }
8a8d5d85 834#ifdef SMP
a2a5ad0d
MD
835 /*
836 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
837 * to or from the MP lock. In this case td_mpcount will be pre-disposed
838 * (non-zero) but not actually synchronized with the actual state of the
839 * lock. We can use it to imply an MP lock requirement for the
840 * preemption but we cannot use it to test whether we hold the MP lock
841 * or not.
a2a5ad0d 842 */
96728c05 843 savecnt = td->td_mpcount;
71ef2f5c 844 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
845 ntd->td_mpcount += td->td_mpcount;
846 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
847 ntd->td_mpcount -= td->td_mpcount;
848 ++preempt_miss;
8ec60c3f 849 need_lwkt_resched();
8a8d5d85
MD
850 return;
851 }
852#endif
26a0694b 853
8ec60c3f
MD
854 /*
855 * Since we are able to preempt the current thread, there is no need to
856 * call need_lwkt_resched().
857 */
26a0694b
MD
858 ++preempt_hit;
859 ntd->td_preempted = td;
860 td->td_flags |= TDF_PREEMPT_LOCK;
861 td->td_switch(ntd);
862 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
863#ifdef SMP
864 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
865 mpheld = MP_LOCK_HELD();
866 if (mpheld && td->td_mpcount == 0)
96728c05 867 cpu_rel_mplock();
71ef2f5c 868 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
869 panic("lwkt_preempt(): MP lock was not held through");
870#endif
26a0694b
MD
871 ntd->td_preempted = NULL;
872 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
873}
874
f1d1c3fa
MD
875/*
876 * Yield our thread while higher priority threads are pending. This is
877 * typically called when we leave a critical section but it can be safely
878 * called while we are in a critical section.
879 *
880 * This function will not generally yield to equal priority threads but it
881 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 882 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
883 * lwkt_yield_quick().
884 *
235957ed 885 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
886 * came along but was blocked and made pending.
887 *
f1d1c3fa
MD
888 * (self contained on a per cpu basis)
889 */
890void
891lwkt_yield_quick(void)
892{
7966cb69
MD
893 globaldata_t gd = mycpu;
894 thread_t td = gd->gd_curthread;
ef0fdad1 895
a2a5ad0d 896 /*
235957ed 897 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
898 * it with a non-zero cpl then we might not wind up calling splz after
899 * a task switch when the critical section is exited even though the
46a3f46d 900 * new task could accept the interrupt.
a2a5ad0d
MD
901 *
902 * XXX from crit_exit() only called after last crit section is released.
903 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
904 *
905 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
906 * except for this special case, we MUST call splz() here to handle any
907 * pending ints, particularly after we switch, or we might accidently
908 * halt the cpu with interrupts pending.
a2a5ad0d 909 */
46a3f46d 910 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 911 splz();
f1d1c3fa
MD
912
913 /*
914 * YYY enabling will cause wakeup() to task-switch, which really
915 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
916 * preemption and MP without actually doing preemption or MP, because a
917 * lot of code assumes that wakeup() does not block.
f1d1c3fa 918 */
46a3f46d
MD
919 if (untimely_switch && td->td_nest_count == 0 &&
920 gd->gd_intr_nesting_level == 0
921 ) {
37af14fe 922 crit_enter_quick(td);
f1d1c3fa
MD
923 /*
924 * YYY temporary hacks until we disassociate the userland scheduler
925 * from the LWKT scheduler.
926 */
927 if (td->td_flags & TDF_RUNQ) {
928 lwkt_switch(); /* will not reenter yield function */
929 } else {
37af14fe 930 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 931 lwkt_switch(); /* will not reenter yield function */
37af14fe 932 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 933 }
7966cb69 934 crit_exit_noyield(td);
f1d1c3fa 935 }
f1d1c3fa
MD
936}
937
8ad65e08 938/*
f1d1c3fa 939 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 940 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
941 * the crit_exit() call in lwkt_switch().
942 *
943 * (self contained on a per cpu basis)
8ad65e08
MD
944 */
945void
f1d1c3fa 946lwkt_yield(void)
8ad65e08 947{
37af14fe 948 lwkt_schedule_self(curthread);
f1d1c3fa
MD
949 lwkt_switch();
950}
951
8ad65e08 952/*
f1d1c3fa
MD
953 * Generic schedule. Possibly schedule threads belonging to other cpus and
954 * deal with threads that might be blocked on a wait queue.
955 *
0a3f9b47
MD
956 * We have a little helper inline function which does additional work after
957 * the thread has been enqueued, including dealing with preemption and
958 * setting need_lwkt_resched() (which prevents the kernel from returning
959 * to userland until it has processed higher priority threads).
6330a558
MD
960 *
961 * It is possible for this routine to be called after a failed _enqueue
962 * (due to the target thread migrating, sleeping, or otherwise blocked).
963 * We have to check that the thread is actually on the run queue!
8ad65e08 964 */
0a3f9b47
MD
965static __inline
966void
8ec60c3f 967_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 968{
c730be20
MD
969 int mypri;
970
6330a558
MD
971 if (ntd->td_flags & TDF_RUNQ) {
972 if (ntd->td_preemptable) {
973 ntd->td_preemptable(ntd, cpri); /* YYY +token */
c730be20
MD
974 } else {
975 /*
976 * This is a little sticky. Due to the passive release function
977 * the LWKT priority can wiggle around for threads acting in
978 * the kernel on behalf of a user process. We do not want this
979 * to effect the comparison per-say.
980 *
981 * What will happen is that the current user process will be
982 * allowed to run until the next hardclock at which time a
983 * forced need_lwkt_resched() will allow the other kernel mode
984 * threads to get in their two cents. This prevents cavitation.
985 */
986 mypri = gd->gd_curthread->td_pri & TDPRI_MASK;
987 if (mypri >= TDPRI_USER_IDLE && mypri <= TDPRI_USER_REAL)
988 mypri = TDPRI_KERN_USER;
989
990 if ((ntd->td_pri & TDPRI_MASK) > mypri)
991 need_lwkt_resched();
6330a558 992 }
0a3f9b47
MD
993 }
994}
995
8ad65e08
MD
996void
997lwkt_schedule(thread_t td)
998{
37af14fe
MD
999 globaldata_t mygd = mycpu;
1000
41a01a4d 1001 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1002 crit_enter_gd(mygd);
9388413d 1003 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1004 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1005 _lwkt_enqueue(td);
1006 } else {
f1d1c3fa 1007 /*
7cd8d145
MD
1008 * If we own the thread, there is no race (since we are in a
1009 * critical section). If we do not own the thread there might
1010 * be a race but the target cpu will deal with it.
f1d1c3fa 1011 */
0f7a3396 1012#ifdef SMP
7cd8d145 1013 if (td->td_gd == mygd) {
9d265729
MD
1014 _lwkt_enqueue(td);
1015 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 1016 } else {
7cd8d145
MD
1017 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1018 }
0f7a3396 1019#else
7cd8d145
MD
1020 _lwkt_enqueue(td);
1021 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 1022#endif
8ad65e08 1023 }
37af14fe 1024 crit_exit_gd(mygd);
8ad65e08
MD
1025}
1026
52eedfb5
MD
1027#ifdef SMP
1028
d9eea1a5 1029/*
52eedfb5
MD
1030 * Thread migration using a 'Pull' method. The thread may or may not be
1031 * the current thread. It MUST be descheduled and in a stable state.
1032 * lwkt_giveaway() must be called on the cpu owning the thread.
1033 *
1034 * At any point after lwkt_giveaway() is called, the target cpu may
1035 * 'pull' the thread by calling lwkt_acquire().
1036 *
1037 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1038 */
52eedfb5
MD
1039void
1040lwkt_giveaway(thread_t td)
1041{
1042 globaldata_t gd = mycpu;
1043
1044 crit_enter_gd(gd);
1045 KKASSERT(td->td_gd == gd);
1046 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1047 td->td_flags |= TDF_MIGRATING;
1048 crit_exit_gd(gd);
1049}
1050
a2a5ad0d
MD
1051void
1052lwkt_acquire(thread_t td)
1053{
37af14fe
MD
1054 globaldata_t gd;
1055 globaldata_t mygd;
a2a5ad0d 1056
52eedfb5 1057 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1058 gd = td->td_gd;
37af14fe 1059 mygd = mycpu;
52eedfb5 1060 if (gd != mycpu) {
35238fa5 1061 cpu_lfence();
52eedfb5 1062 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1063 crit_enter_gd(mygd);
df910c23
MD
1064 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1065#ifdef SMP
1066 lwkt_process_ipiq();
1067#endif
52eedfb5 1068 cpu_lfence();
df910c23 1069 }
37af14fe 1070 td->td_gd = mygd;
52eedfb5
MD
1071 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1072 td->td_flags &= ~TDF_MIGRATING;
1073 crit_exit_gd(mygd);
1074 } else {
1075 crit_enter_gd(mygd);
1076 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1077 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1078 crit_exit_gd(mygd);
a2a5ad0d
MD
1079 }
1080}
1081
52eedfb5
MD
1082#endif
1083
f1d1c3fa
MD
1084/*
1085 * Generic deschedule. Descheduling threads other then your own should be
1086 * done only in carefully controlled circumstances. Descheduling is
1087 * asynchronous.
1088 *
1089 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1090 */
1091void
1092lwkt_deschedule(thread_t td)
1093{
f1d1c3fa 1094 crit_enter();
b8a98473 1095#ifdef SMP
f1d1c3fa
MD
1096 if (td == curthread) {
1097 _lwkt_dequeue(td);
1098 } else {
a72187e9 1099 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1100 _lwkt_dequeue(td);
1101 } else {
b8a98473 1102 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1103 }
1104 }
b8a98473
MD
1105#else
1106 _lwkt_dequeue(td);
1107#endif
f1d1c3fa
MD
1108 crit_exit();
1109}
1110
4b5f931b
MD
1111/*
1112 * Set the target thread's priority. This routine does not automatically
1113 * switch to a higher priority thread, LWKT threads are not designed for
1114 * continuous priority changes. Yield if you want to switch.
1115 *
1116 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1117 * of the td_pri field. The specified priority may also indicate zero or
1118 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1119 *
1120 * Note that we requeue the thread whether it winds up on a different runq
1121 * or not. uio_yield() depends on this and the routine is not normally
1122 * called with the same priority otherwise.
4b5f931b
MD
1123 */
1124void
1125lwkt_setpri(thread_t td, int pri)
1126{
26a0694b 1127 KKASSERT(pri >= 0);
a72187e9 1128 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1129 crit_enter();
1130 if (td->td_flags & TDF_RUNQ) {
1131 _lwkt_dequeue(td);
1132 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1133 _lwkt_enqueue(td);
1134 } else {
1135 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1136 }
1137 crit_exit();
1138}
1139
1140void
1141lwkt_setpri_self(int pri)
1142{
1143 thread_t td = curthread;
1144
4b5f931b
MD
1145 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1146 crit_enter();
1147 if (td->td_flags & TDF_RUNQ) {
1148 _lwkt_dequeue(td);
1149 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1150 _lwkt_enqueue(td);
1151 } else {
1152 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1153 }
1154 crit_exit();
1155}
1156
5d21b981 1157/*
52eedfb5
MD
1158 * Migrate the current thread to the specified cpu.
1159 *
1160 * This is accomplished by descheduling ourselves from the current cpu,
1161 * moving our thread to the tdallq of the target cpu, IPI messaging the
1162 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1163 * races while the thread is being migrated.
5d21b981 1164 */
3d28ff59 1165#ifdef SMP
5d21b981 1166static void lwkt_setcpu_remote(void *arg);
3d28ff59 1167#endif
5d21b981
MD
1168
1169void
1170lwkt_setcpu_self(globaldata_t rgd)
1171{
1172#ifdef SMP
1173 thread_t td = curthread;
1174
1175 if (td->td_gd != rgd) {
1176 crit_enter_quick(td);
1177 td->td_flags |= TDF_MIGRATING;
1178 lwkt_deschedule_self(td);
52eedfb5 1179 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1180 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1181 lwkt_switch();
1182 /* we are now on the target cpu */
52eedfb5 1183 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1184 crit_exit_quick(td);
1185 }
1186#endif
1187}
1188
ecdefdda
MD
1189void
1190lwkt_migratecpu(int cpuid)
1191{
1192#ifdef SMP
1193 globaldata_t rgd;
1194
1195 rgd = globaldata_find(cpuid);
1196 lwkt_setcpu_self(rgd);
1197#endif
1198}
1199
5d21b981
MD
1200/*
1201 * Remote IPI for cpu migration (called while in a critical section so we
1202 * do not have to enter another one). The thread has already been moved to
1203 * our cpu's allq, but we must wait for the thread to be completely switched
1204 * out on the originating cpu before we schedule it on ours or the stack
1205 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1206 * change to main memory.
1207 *
1208 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1209 * against wakeups. It is best if this interface is used only when there
1210 * are no pending events that might try to schedule the thread.
1211 */
3d28ff59 1212#ifdef SMP
5d21b981
MD
1213static void
1214lwkt_setcpu_remote(void *arg)
1215{
1216 thread_t td = arg;
1217 globaldata_t gd = mycpu;
1218
df910c23
MD
1219 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1220#ifdef SMP
1221 lwkt_process_ipiq();
1222#endif
35238fa5 1223 cpu_lfence();
df910c23 1224 }
5d21b981 1225 td->td_gd = gd;
35238fa5 1226 cpu_sfence();
5d21b981 1227 td->td_flags &= ~TDF_MIGRATING;
9388413d 1228 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1229 _lwkt_enqueue(td);
1230}
3d28ff59 1231#endif
5d21b981 1232
553ea3c8 1233struct lwp *
4b5f931b
MD
1234lwkt_preempted_proc(void)
1235{
73e4f7b9 1236 thread_t td = curthread;
4b5f931b
MD
1237 while (td->td_preempted)
1238 td = td->td_preempted;
553ea3c8 1239 return(td->td_lwp);
4b5f931b
MD
1240}
1241
99df837e
MD
1242/*
1243 * Create a kernel process/thread/whatever. It shares it's address space
1244 * with proc0 - ie: kernel only.
1245 *
365fa13f
MD
1246 * NOTE! By default new threads are created with the MP lock held. A
1247 * thread which does not require the MP lock should release it by calling
1248 * rel_mplock() at the start of the new thread.
99df837e
MD
1249 */
1250int
1251lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1252 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1253 const char *fmt, ...)
99df837e 1254{
73e4f7b9 1255 thread_t td;
e2565a42 1256 __va_list ap;
99df837e 1257
d3d32139 1258 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1259 tdflags);
a2a5ad0d
MD
1260 if (tdp)
1261 *tdp = td;
709799ea 1262 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1263
1264 /*
1265 * Set up arg0 for 'ps' etc
1266 */
e2565a42 1267 __va_start(ap, fmt);
379210cb 1268 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1269 __va_end(ap);
99df837e
MD
1270
1271 /*
1272 * Schedule the thread to run
1273 */
ef0fdad1
MD
1274 if ((td->td_flags & TDF_STOPREQ) == 0)
1275 lwkt_schedule(td);
1276 else
1277 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1278 return 0;
1279}
1280
1281/*
1282 * Destroy an LWKT thread. Warning! This function is not called when
1283 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1284 * uses a different reaping mechanism.
1285 */
1286void
1287lwkt_exit(void)
1288{
1289 thread_t td = curthread;
c070746a 1290 thread_t std;
8826f33a 1291 globaldata_t gd;
99df837e
MD
1292
1293 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1294 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1295 caps_exit(td);
c070746a
MD
1296
1297 /*
1298 * Get us into a critical section to interlock gd_freetd and loop
1299 * until we can get it freed.
1300 *
1301 * We have to cache the current td in gd_freetd because objcache_put()ing
1302 * it would rip it out from under us while our thread is still active.
1303 */
1304 gd = mycpu;
37af14fe 1305 crit_enter_quick(td);
c070746a
MD
1306 while ((std = gd->gd_freetd) != NULL) {
1307 gd->gd_freetd = NULL;
1308 objcache_put(thread_cache, std);
1309 }
37af14fe 1310 lwkt_deschedule_self(td);
e56e4dea 1311 lwkt_remove_tdallq(td);
c070746a
MD
1312 if (td->td_flags & TDF_ALLOCATED_THREAD)
1313 gd->gd_freetd = td;
99df837e
MD
1314 cpu_thread_exit();
1315}
1316
e56e4dea
MD
1317void
1318lwkt_remove_tdallq(thread_t td)
1319{
1320 KKASSERT(td->td_gd == mycpu);
1321 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1322}
1323
2d93b37a
MD
1324void
1325crit_panic(void)
1326{
1327 thread_t td = curthread;
1328 int lpri = td->td_pri;
1329
1330 td->td_pri = 0;
1331 panic("td_pri is/would-go negative! %p %d", td, lpri);
1332}
1333
d165e668
MD
1334#ifdef SMP
1335
bd8015ca
MD
1336/*
1337 * Called from debugger/panic on cpus which have been stopped. We must still
1338 * process the IPIQ while stopped, even if we were stopped while in a critical
1339 * section (XXX).
1340 *
1341 * If we are dumping also try to process any pending interrupts. This may
1342 * or may not work depending on the state of the cpu at the point it was
1343 * stopped.
1344 */
1345void
1346lwkt_smp_stopped(void)
1347{
1348 globaldata_t gd = mycpu;
1349
1350 crit_enter_gd(gd);
1351 if (dumping) {
1352 lwkt_process_ipiq();
1353 splz();
1354 } else {
1355 lwkt_process_ipiq();
1356 }
1357 crit_exit_gd(gd);
1358}
1359
57aa743c
MD
1360/*
1361 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1362 * get_mplock() has already incremented td_mpcount. We must block and
1363 * not return until giant is held.
1364 *
1365 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1366 * reschedule the thread until it can obtain the giant lock for it.
1367 */
1368void
1369lwkt_mp_lock_contested(void)
1370{
57aa743c 1371 loggiant(beg);
57aa743c 1372 lwkt_switch();
57aa743c 1373 loggiant(end);
57aa743c
MD
1374}
1375
d165e668 1376#endif