According to SUSv3 including just regex.h must be enough. Fixes build of
[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 *
40aaf5fc 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.114 2008/05/26 17:11:09 nth 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{
99df837e 252 void *stack;
7d0bac62 253
ef0fdad1 254 if (td == NULL) {
40aaf5fc
NT
255 td = objcache_get(thread_cache, M_WAITOK);
256 KASSERT((td->td_flags &
257 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
258 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
259 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 260 }
f470d0c8
MD
261 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
262 if (flags & TDF_ALLOCATED_STACK) {
e4846942 263 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
264 stack = NULL;
265 }
266 }
267 if (stack == NULL) {
e4846942 268 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 269 flags |= TDF_ALLOCATED_STACK;
99df837e 270 }
75cdbe6c 271 if (cpu < 0)
f470d0c8 272 lwkt_init_thread(td, stack, stksize, flags, mycpu);
75cdbe6c 273 else
f470d0c8 274 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 275 return(td);
7d0bac62
MD
276}
277
278/*
279 * Initialize a preexisting thread structure. This function is used by
280 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
281 *
f8c3996b
MD
282 * All threads start out in a critical section at a priority of
283 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
284 * appropriate. This function may send an IPI message when the
285 * requested cpu is not the current cpu and consequently gd_tdallq may
286 * not be initialized synchronously from the point of view of the originating
287 * cpu.
288 *
289 * NOTE! we have to be careful in regards to creating threads for other cpus
290 * if SMP has not yet been activated.
7d0bac62 291 */
41a01a4d
MD
292#ifdef SMP
293
75cdbe6c
MD
294static void
295lwkt_init_thread_remote(void *arg)
296{
297 thread_t td = arg;
298
52eedfb5
MD
299 /*
300 * Protected by critical section held by IPI dispatch
301 */
75cdbe6c
MD
302 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
303}
304
41a01a4d
MD
305#endif
306
7d0bac62 307void
f470d0c8
MD
308lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
309 struct globaldata *gd)
7d0bac62 310{
37af14fe
MD
311 globaldata_t mygd = mycpu;
312
99df837e
MD
313 bzero(td, sizeof(struct thread));
314 td->td_kstack = stack;
f470d0c8 315 td->td_kstack_size = stksize;
d3d32139 316 td->td_flags = flags;
26a0694b 317 td->td_gd = gd;
f8c3996b 318 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
319#ifdef SMP
320 if ((flags & TDF_MPSAFE) == 0)
321 td->td_mpcount = 1;
322#endif
fb0f29c4
MD
323 if (lwkt_use_spin_port)
324 lwkt_initport_spin(&td->td_msgport);
325 else
326 lwkt_initport_thread(&td->td_msgport, td);
99df837e 327 pmap_init_thread(td);
0f7a3396 328#ifdef SMP
5d21b981
MD
329 /*
330 * Normally initializing a thread for a remote cpu requires sending an
331 * IPI. However, the idlethread is setup before the other cpus are
332 * activated so we have to treat it as a special case. XXX manipulation
333 * of gd_tdallq requires the BGL.
334 */
335 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 336 crit_enter_gd(mygd);
75cdbe6c 337 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 338 crit_exit_gd(mygd);
75cdbe6c 339 } else {
2db3b277 340 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 341 }
0f7a3396 342#else
37af14fe 343 crit_enter_gd(mygd);
0f7a3396 344 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 345 crit_exit_gd(mygd);
0f7a3396 346#endif
73e4f7b9
MD
347}
348
349void
350lwkt_set_comm(thread_t td, const char *ctl, ...)
351{
e2565a42 352 __va_list va;
73e4f7b9 353
e2565a42 354 __va_start(va, ctl);
379210cb 355 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 356 __va_end(va);
7d0bac62
MD
357}
358
99df837e 359void
73e4f7b9 360lwkt_hold(thread_t td)
99df837e 361{
73e4f7b9
MD
362 ++td->td_refs;
363}
364
365void
366lwkt_rele(thread_t td)
367{
368 KKASSERT(td->td_refs > 0);
369 --td->td_refs;
370}
371
372void
373lwkt_wait_free(thread_t td)
374{
375 while (td->td_refs)
377d4740 376 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
377}
378
379void
380lwkt_free_thread(thread_t td)
381{
d9eea1a5 382 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
383 ("lwkt_free_thread: did not exit! %p", td));
384
40aaf5fc
NT
385 if (td->td_flags & TDF_ALLOCATED_THREAD) {
386 objcache_put(thread_cache, td);
387 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
388 /* client-allocated struct with internally allocated stack */
389 KASSERT(td->td_kstack && td->td_kstack_size > 0,
390 ("lwkt_free_thread: corrupted stack"));
391 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
392 td->td_kstack = NULL;
393 td->td_kstack_size = 0;
99df837e
MD
394 }
395}
396
397
8ad65e08
MD
398/*
399 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
400 * switch to the idlethread. Switching must occur within a critical
401 * section to avoid races with the scheduling queue.
402 *
403 * We always have full control over our cpu's run queue. Other cpus
404 * that wish to manipulate our queue must use the cpu_*msg() calls to
405 * talk to our cpu, so a critical section is all that is needed and
406 * the result is very, very fast thread switching.
407 *
96728c05
MD
408 * The LWKT scheduler uses a fixed priority model and round-robins at
409 * each priority level. User process scheduling is a totally
410 * different beast and LWKT priorities should not be confused with
411 * user process priorities.
f1d1c3fa 412 *
96728c05
MD
413 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
414 * cleans it up. Note that the td_switch() function cannot do anything that
415 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
416 * the target thread (not the current thread). It's nice to have a scheduler
417 * that does not need the MP lock to work because it allows us to do some
418 * really cool high-performance MP lock optimizations.
69d78e99
MD
419 *
420 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
421 * is not called by the current thread in the preemption case, only when
422 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
423 */
424void
425lwkt_switch(void)
426{
37af14fe
MD
427 globaldata_t gd = mycpu;
428 thread_t td = gd->gd_curthread;
8ad65e08 429 thread_t ntd;
8a8d5d85
MD
430#ifdef SMP
431 int mpheld;
432#endif
8ad65e08 433
46a3f46d 434 /*
27e88a6e
MD
435 * Switching from within a 'fast' (non thread switched) interrupt or IPI
436 * is illegal. However, we may have to do it anyway if we hit a fatal
437 * kernel trap or we have paniced.
438 *
439 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 440 */
27e88a6e
MD
441 if (gd->gd_intr_nesting_level) {
442 int savegdnest;
443 int savegdtrap;
444
445 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
446 panic("lwkt_switch: cannot switch from within "
447 "a fast interrupt, yet, td %p\n", td);
448 } else {
449 savegdnest = gd->gd_intr_nesting_level;
450 savegdtrap = gd->gd_trap_nesting_level;
451 gd->gd_intr_nesting_level = 0;
452 gd->gd_trap_nesting_level = 0;
a7422615
MD
453 if ((td->td_flags & TDF_PANICWARN) == 0) {
454 td->td_flags |= TDF_PANICWARN;
6ea70f76 455 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615
MD
456 "thread %p (%s)\n", td, td->td_comm);
457#ifdef DDB
458 db_print_backtrace();
459#endif
460 }
27e88a6e
MD
461 lwkt_switch();
462 gd->gd_intr_nesting_level = savegdnest;
463 gd->gd_trap_nesting_level = savegdtrap;
464 return;
465 }
96728c05 466 }
ef0fdad1 467
cb973d15
MD
468 /*
469 * Passive release (used to transition from user to kernel mode
470 * when we block or switch rather then when we enter the kernel).
471 * This function is NOT called if we are switching into a preemption
472 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
473 * our current process designation (if we have one) and become a true
474 * LWKT thread, and may also hand the current process designation to
475 * another process and schedule thread.
cb973d15
MD
476 */
477 if (td->td_release)
478 td->td_release(td);
479
37af14fe 480 crit_enter_gd(gd);
9d265729
MD
481 if (td->td_toks)
482 lwkt_relalltokens(td);
483
484 /*
b02926de
MD
485 * We had better not be holding any spin locks, but don't get into an
486 * endless panic loop.
9d265729 487 */
bbb31c5d
MD
488 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
489 ("lwkt_switch: still holding a shared spinlock %p!",
490 gd->gd_spinlock_rd));
d666840a
MD
491 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
492 ("lwkt_switch: still holding %d exclusive spinlocks!",
493 gd->gd_spinlocks_wr));
9d265729 494
8a8d5d85
MD
495
496#ifdef SMP
497 /*
498 * td_mpcount cannot be used to determine if we currently hold the
499 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
500 * to get the lock, and switch out if it can't. Our ownership of
501 * the actual lock will remain stable while we are in a critical section
502 * (but, of course, another cpu may own or release the lock so the
503 * actual value of mp_lock is not stable).
8a8d5d85
MD
504 */
505 mpheld = MP_LOCK_HELD();
0f7a3396
MD
506#ifdef INVARIANTS
507 if (td->td_cscount) {
6ea70f76 508 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
509 td);
510 if (panic_on_cscount)
511 panic("switching while mastering cpusync");
512 }
513#endif
8a8d5d85 514#endif
99df837e
MD
515 if ((ntd = td->td_preempted) != NULL) {
516 /*
517 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
518 * thread. This occurs transparently, whether the preempted thread
519 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
520 * itself).
521 *
522 * We have to setup the MP lock for the original thread after backing
523 * out the adjustment that was made to curthread when the original
524 * was preempted.
99df837e 525 */
26a0694b 526 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 527#ifdef SMP
96728c05 528 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 529 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
530 td, ntd, td->td_mpcount, ntd->td_mpcount);
531 }
8a8d5d85
MD
532 if (ntd->td_mpcount) {
533 td->td_mpcount -= ntd->td_mpcount;
534 KKASSERT(td->td_mpcount >= 0);
535 }
536#endif
26a0694b 537 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
538
539 /*
540 * XXX. The interrupt may have woken a thread up, we need to properly
541 * set the reschedule flag if the originally interrupted thread is at
542 * a lower priority.
543 */
544 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
545 need_lwkt_resched();
8a8d5d85 546 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 547 } else {
4b5f931b
MD
548 /*
549 * Priority queue / round-robin at each priority. Note that user
550 * processes run at a fixed, low priority and the user process
551 * scheduler deals with interactions between user processes
552 * by scheduling and descheduling them from the LWKT queue as
553 * necessary.
8a8d5d85
MD
554 *
555 * We have to adjust the MP lock for the target thread. If we
556 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
557 * thread that does not need the MP lock. If we cannot, we spin
558 * instead of HLT.
559 *
560 * A similar issue exists for the tokens held by the target thread.
561 * If we cannot obtain ownership of the tokens we cannot immediately
562 * schedule the thread.
563 */
564
8ec60c3f
MD
565 /*
566 * If an LWKT reschedule was requested, well that is what we are
567 * doing now so clear it.
568 */
569 clear_lwkt_resched();
4b5f931b
MD
570again:
571 if (gd->gd_runqmask) {
572 int nq = bsrl(gd->gd_runqmask);
573 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
574 gd->gd_runqmask &= ~(1 << nq);
575 goto again;
576 }
8a8d5d85 577#ifdef SMP
41a01a4d 578 /*
df6b8ba0
MD
579 * THREAD SELECTION FOR AN SMP MACHINE BUILD
580 *
41a01a4d
MD
581 * If the target needs the MP lock and we couldn't get it,
582 * or if the target is holding tokens and we could not
583 * gain ownership of the tokens, continue looking for a
584 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
585 *
586 * NOTE: the mpheld variable invalid after this conditional, it
587 * can change due to both cpu_try_mplock() returning success
9d265729 588 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
589 * we are trying to check the mpcount of a thread other then
590 * the current thread. Because of this, if the current thread
591 * is not holding td_mpcount, an IPI indirectly run via
9d265729 592 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 593 * cause the core MP lock to be released.
41a01a4d
MD
594 */
595 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 596 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 597 ) {
8a8d5d85 598 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
599
600 mpheld = MP_LOCK_HELD();
601 ntd = NULL;
8a8d5d85
MD
602 while (rqmask) {
603 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 604 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 605 /* spinning due to MP lock being held */
38717797 606#ifdef INVARIANTS
a453459d 607 ++mplock_contention_count;
38717797 608#endif
a453459d 609 /* mplock still not held, 'mpheld' still valid */
41a01a4d 610 continue;
38717797 611 }
a453459d
MD
612
613 /*
9d265729 614 * mpheld state invalid after getalltokens call returns
a453459d
MD
615 * failure, but the variable is only needed for
616 * the loop.
617 */
9d265729 618 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 619 /* spinning due to token contention */
38717797 620#ifdef INVARIANTS
a453459d 621 ++token_contention_count;
38717797 622#endif
a453459d 623 mpheld = MP_LOCK_HELD();
41a01a4d 624 continue;
38717797 625 }
41a01a4d 626 break;
8a8d5d85
MD
627 }
628 if (ntd)
629 break;
630 rqmask &= ~(1 << nq);
631 nq = bsrl(rqmask);
632 }
633 if (ntd == NULL) {
b402c633 634 cpu_mplock_contested();
a2a5ad0d
MD
635 ntd = &gd->gd_idlethread;
636 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 637 goto using_idle_thread;
8a8d5d85 638 } else {
344ad853 639 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
640 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
641 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
642 }
643 } else {
344ad853 644 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
645 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
646 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
647 }
648#else
df6b8ba0
MD
649 /*
650 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
651 * worry about tokens or the BGL. However, we still have
652 * to call lwkt_getalltokens() in order to properly detect
653 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 654 */
7eb611ef 655 lwkt_getalltokens(ntd);
344ad853 656 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
657 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
658 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 659#endif
4b5f931b 660 } else {
3c23a41a 661 /*
60f945af
MD
662 * We have nothing to run but only let the idle loop halt
663 * the cpu if there are no pending interrupts.
3c23a41a 664 */
a2a5ad0d 665 ntd = &gd->gd_idlethread;
60f945af 666 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 667 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 668#ifdef SMP
df6b8ba0
MD
669using_idle_thread:
670 /*
671 * The idle thread should not be holding the MP lock unless we
672 * are trapping in the kernel or in a panic. Since we select the
673 * idle thread unconditionally when no other thread is available,
674 * if the MP lock is desired during a panic or kernel trap, we
675 * have to loop in the scheduler until we get it.
676 */
677 if (ntd->td_mpcount) {
678 mpheld = MP_LOCK_HELD();
b402c633 679 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 680 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
681 } else if (mpheld == 0) {
682 cpu_mplock_contested();
df6b8ba0 683 goto again;
b402c633 684 }
df6b8ba0 685 }
a453459d 686#endif
4b5f931b 687 }
f1d1c3fa 688 }
26a0694b
MD
689 KASSERT(ntd->td_pri >= TDPRI_CRIT,
690 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
691
692 /*
693 * Do the actual switch. If the new target does not need the MP lock
694 * and we are holding it, release the MP lock. If the new target requires
695 * the MP lock we have already acquired it for the target.
696 */
697#ifdef SMP
698 if (ntd->td_mpcount == 0 ) {
699 if (MP_LOCK_HELD())
700 cpu_rel_mplock();
701 } else {
a453459d 702 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
703 }
704#endif
94f6d86e
MD
705 if (td != ntd) {
706 ++switch_count;
f1d1c3fa 707 td->td_switch(ntd);
94f6d86e 708 }
37af14fe
MD
709 /* NOTE: current cpu may have changed after switch */
710 crit_exit_quick(td);
8ad65e08
MD
711}
712
b68b7282 713/*
96728c05
MD
714 * Request that the target thread preempt the current thread. Preemption
715 * only works under a specific set of conditions:
b68b7282 716 *
96728c05
MD
717 * - We are not preempting ourselves
718 * - The target thread is owned by the current cpu
719 * - We are not currently being preempted
720 * - The target is not currently being preempted
d3d1cbc8
MD
721 * - We are not holding any spin locks
722 * - The target thread is not holding any tokens
96728c05
MD
723 * - We are able to satisfy the target's MP lock requirements (if any).
724 *
725 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
726 * this is called via lwkt_schedule() through the td_preemptable callback.
727 * critpri is the managed critical priority that we should ignore in order
728 * to determine whether preemption is possible (aka usually just the crit
729 * priority of lwkt_schedule() itself).
b68b7282 730 *
26a0694b
MD
731 * XXX at the moment we run the target thread in a critical section during
732 * the preemption in order to prevent the target from taking interrupts
733 * that *WE* can't. Preemption is strictly limited to interrupt threads
734 * and interrupt-like threads, outside of a critical section, and the
735 * preempted source thread will be resumed the instant the target blocks
736 * whether or not the source is scheduled (i.e. preemption is supposed to
737 * be as transparent as possible).
4b5f931b 738 *
8a8d5d85
MD
739 * The target thread inherits our MP count (added to its own) for the
740 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
741 * MP lock during the preemption. Therefore, any preempting targets must be
742 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
743 * out of sync with the physical mp_lock, but we do not have to preserve
744 * the original ownership of the lock if it was out of synch (that is, we
745 * can leave it synchronized on return).
b68b7282
MD
746 */
747void
96728c05 748lwkt_preempt(thread_t ntd, int critpri)
b68b7282 749{
46a3f46d 750 struct globaldata *gd = mycpu;
0a3f9b47 751 thread_t td;
8a8d5d85
MD
752#ifdef SMP
753 int mpheld;
57c254db 754 int savecnt;
8a8d5d85 755#endif
b68b7282 756
26a0694b 757 /*
96728c05
MD
758 * The caller has put us in a critical section. We can only preempt
759 * if the caller of the caller was not in a critical section (basically
d666840a 760 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 761 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 762 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
763 *
764 * YYY The target thread must be in a critical section (else it must
765 * inherit our critical section? I dunno yet).
41a01a4d 766 *
0a3f9b47 767 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
768 */
769 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 770
0a3f9b47 771 td = gd->gd_curthread;
0a3f9b47 772 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
773 ++preempt_miss;
774 return;
775 }
96728c05
MD
776 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
777 ++preempt_miss;
8ec60c3f 778 need_lwkt_resched();
96728c05
MD
779 return;
780 }
781#ifdef SMP
46a3f46d 782 if (ntd->td_gd != gd) {
96728c05 783 ++preempt_miss;
8ec60c3f 784 need_lwkt_resched();
96728c05
MD
785 return;
786 }
787#endif
41a01a4d 788 /*
d3d1cbc8 789 * Take the easy way out and do not preempt if we are holding
d666840a 790 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
791 * preempted interlock against the target thread's tokens and whether
792 * we can get all the target thread's tokens, but this situation
793 * should not occur very often so its easier to simply not preempt.
d666840a
MD
794 * Also, plain spinlocks are impossible to figure out at this point so
795 * just don't preempt.
d3d1cbc8
MD
796 *
797 * Do not try to preempt if the target thread is holding any tokens.
798 * We could try to acquire the tokens but this case is so rare there
799 * is no need to support it.
41a01a4d 800 */
bbb31c5d 801 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 802 ++preempt_miss;
8ec60c3f 803 need_lwkt_resched();
41a01a4d
MD
804 return;
805 }
d3d1cbc8
MD
806 if (ntd->td_toks) {
807 ++preempt_miss;
808 need_lwkt_resched();
809 return;
810 }
26a0694b
MD
811 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
812 ++preempt_weird;
8ec60c3f 813 need_lwkt_resched();
26a0694b
MD
814 return;
815 }
816 if (ntd->td_preempted) {
4b5f931b 817 ++preempt_hit;
8ec60c3f 818 need_lwkt_resched();
26a0694b 819 return;
b68b7282 820 }
8a8d5d85 821#ifdef SMP
a2a5ad0d
MD
822 /*
823 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
824 * to or from the MP lock. In this case td_mpcount will be pre-disposed
825 * (non-zero) but not actually synchronized with the actual state of the
826 * lock. We can use it to imply an MP lock requirement for the
827 * preemption but we cannot use it to test whether we hold the MP lock
828 * or not.
a2a5ad0d 829 */
96728c05 830 savecnt = td->td_mpcount;
71ef2f5c 831 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
832 ntd->td_mpcount += td->td_mpcount;
833 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
834 ntd->td_mpcount -= td->td_mpcount;
835 ++preempt_miss;
8ec60c3f 836 need_lwkt_resched();
8a8d5d85
MD
837 return;
838 }
839#endif
26a0694b 840
8ec60c3f
MD
841 /*
842 * Since we are able to preempt the current thread, there is no need to
843 * call need_lwkt_resched().
844 */
26a0694b
MD
845 ++preempt_hit;
846 ntd->td_preempted = td;
847 td->td_flags |= TDF_PREEMPT_LOCK;
848 td->td_switch(ntd);
849 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
850#ifdef SMP
851 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
852 mpheld = MP_LOCK_HELD();
853 if (mpheld && td->td_mpcount == 0)
96728c05 854 cpu_rel_mplock();
71ef2f5c 855 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
856 panic("lwkt_preempt(): MP lock was not held through");
857#endif
26a0694b
MD
858 ntd->td_preempted = NULL;
859 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
860}
861
f1d1c3fa
MD
862/*
863 * Yield our thread while higher priority threads are pending. This is
864 * typically called when we leave a critical section but it can be safely
865 * called while we are in a critical section.
866 *
867 * This function will not generally yield to equal priority threads but it
868 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 869 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
870 * lwkt_yield_quick().
871 *
235957ed 872 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
873 * came along but was blocked and made pending.
874 *
f1d1c3fa
MD
875 * (self contained on a per cpu basis)
876 */
877void
878lwkt_yield_quick(void)
879{
7966cb69
MD
880 globaldata_t gd = mycpu;
881 thread_t td = gd->gd_curthread;
ef0fdad1 882
a2a5ad0d 883 /*
235957ed 884 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
885 * it with a non-zero cpl then we might not wind up calling splz after
886 * a task switch when the critical section is exited even though the
46a3f46d 887 * new task could accept the interrupt.
a2a5ad0d
MD
888 *
889 * XXX from crit_exit() only called after last crit section is released.
890 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
891 *
892 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
893 * except for this special case, we MUST call splz() here to handle any
894 * pending ints, particularly after we switch, or we might accidently
895 * halt the cpu with interrupts pending.
a2a5ad0d 896 */
46a3f46d 897 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 898 splz();
f1d1c3fa
MD
899
900 /*
901 * YYY enabling will cause wakeup() to task-switch, which really
902 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
903 * preemption and MP without actually doing preemption or MP, because a
904 * lot of code assumes that wakeup() does not block.
f1d1c3fa 905 */
46a3f46d
MD
906 if (untimely_switch && td->td_nest_count == 0 &&
907 gd->gd_intr_nesting_level == 0
908 ) {
37af14fe 909 crit_enter_quick(td);
f1d1c3fa
MD
910 /*
911 * YYY temporary hacks until we disassociate the userland scheduler
912 * from the LWKT scheduler.
913 */
914 if (td->td_flags & TDF_RUNQ) {
915 lwkt_switch(); /* will not reenter yield function */
916 } else {
37af14fe 917 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 918 lwkt_switch(); /* will not reenter yield function */
37af14fe 919 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 920 }
7966cb69 921 crit_exit_noyield(td);
f1d1c3fa 922 }
f1d1c3fa
MD
923}
924
8ad65e08 925/*
f1d1c3fa 926 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 927 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
928 * the crit_exit() call in lwkt_switch().
929 *
930 * (self contained on a per cpu basis)
8ad65e08
MD
931 */
932void
f1d1c3fa 933lwkt_yield(void)
8ad65e08 934{
37af14fe 935 lwkt_schedule_self(curthread);
f1d1c3fa
MD
936 lwkt_switch();
937}
938
8ad65e08 939/*
f1d1c3fa
MD
940 * Generic schedule. Possibly schedule threads belonging to other cpus and
941 * deal with threads that might be blocked on a wait queue.
942 *
0a3f9b47
MD
943 * We have a little helper inline function which does additional work after
944 * the thread has been enqueued, including dealing with preemption and
945 * setting need_lwkt_resched() (which prevents the kernel from returning
946 * to userland until it has processed higher priority threads).
6330a558
MD
947 *
948 * It is possible for this routine to be called after a failed _enqueue
949 * (due to the target thread migrating, sleeping, or otherwise blocked).
950 * We have to check that the thread is actually on the run queue!
8ad65e08 951 */
0a3f9b47
MD
952static __inline
953void
8ec60c3f 954_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
0a3f9b47 955{
6330a558
MD
956 if (ntd->td_flags & TDF_RUNQ) {
957 if (ntd->td_preemptable) {
958 ntd->td_preemptable(ntd, cpri); /* YYY +token */
959 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
960 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
961 ) {
962 need_lwkt_resched();
963 }
0a3f9b47
MD
964 }
965}
966
8ad65e08
MD
967void
968lwkt_schedule(thread_t td)
969{
37af14fe
MD
970 globaldata_t mygd = mycpu;
971
41a01a4d 972 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 973 crit_enter_gd(mygd);
9388413d 974 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 975 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
976 _lwkt_enqueue(td);
977 } else {
f1d1c3fa 978 /*
7cd8d145
MD
979 * If we own the thread, there is no race (since we are in a
980 * critical section). If we do not own the thread there might
981 * be a race but the target cpu will deal with it.
f1d1c3fa 982 */
0f7a3396 983#ifdef SMP
7cd8d145 984 if (td->td_gd == mygd) {
9d265729
MD
985 _lwkt_enqueue(td);
986 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
f1d1c3fa 987 } else {
7cd8d145
MD
988 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
989 }
0f7a3396 990#else
7cd8d145
MD
991 _lwkt_enqueue(td);
992 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
0f7a3396 993#endif
8ad65e08 994 }
37af14fe 995 crit_exit_gd(mygd);
8ad65e08
MD
996}
997
52eedfb5
MD
998#ifdef SMP
999
d9eea1a5 1000/*
52eedfb5
MD
1001 * Thread migration using a 'Pull' method. The thread may or may not be
1002 * the current thread. It MUST be descheduled and in a stable state.
1003 * lwkt_giveaway() must be called on the cpu owning the thread.
1004 *
1005 * At any point after lwkt_giveaway() is called, the target cpu may
1006 * 'pull' the thread by calling lwkt_acquire().
1007 *
1008 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1009 */
52eedfb5
MD
1010void
1011lwkt_giveaway(thread_t td)
1012{
1013 globaldata_t gd = mycpu;
1014
1015 crit_enter_gd(gd);
1016 KKASSERT(td->td_gd == gd);
1017 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1018 td->td_flags |= TDF_MIGRATING;
1019 crit_exit_gd(gd);
1020}
1021
a2a5ad0d
MD
1022void
1023lwkt_acquire(thread_t td)
1024{
37af14fe
MD
1025 globaldata_t gd;
1026 globaldata_t mygd;
a2a5ad0d 1027
52eedfb5 1028 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1029 gd = td->td_gd;
37af14fe 1030 mygd = mycpu;
52eedfb5 1031 if (gd != mycpu) {
35238fa5 1032 cpu_lfence();
52eedfb5 1033 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1034 crit_enter_gd(mygd);
52eedfb5
MD
1035 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1036 cpu_lfence();
37af14fe 1037 td->td_gd = mygd;
52eedfb5
MD
1038 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1039 td->td_flags &= ~TDF_MIGRATING;
1040 crit_exit_gd(mygd);
1041 } else {
1042 crit_enter_gd(mygd);
1043 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1044 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1045 crit_exit_gd(mygd);
a2a5ad0d
MD
1046 }
1047}
1048
52eedfb5
MD
1049#endif
1050
f1d1c3fa
MD
1051/*
1052 * Generic deschedule. Descheduling threads other then your own should be
1053 * done only in carefully controlled circumstances. Descheduling is
1054 * asynchronous.
1055 *
1056 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1057 */
1058void
1059lwkt_deschedule(thread_t td)
1060{
f1d1c3fa 1061 crit_enter();
b8a98473 1062#ifdef SMP
f1d1c3fa
MD
1063 if (td == curthread) {
1064 _lwkt_dequeue(td);
1065 } else {
a72187e9 1066 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1067 _lwkt_dequeue(td);
1068 } else {
b8a98473 1069 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1070 }
1071 }
b8a98473
MD
1072#else
1073 _lwkt_dequeue(td);
1074#endif
f1d1c3fa
MD
1075 crit_exit();
1076}
1077
4b5f931b
MD
1078/*
1079 * Set the target thread's priority. This routine does not automatically
1080 * switch to a higher priority thread, LWKT threads are not designed for
1081 * continuous priority changes. Yield if you want to switch.
1082 *
1083 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1084 * of the td_pri field. The specified priority may also indicate zero or
1085 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1086 *
1087 * Note that we requeue the thread whether it winds up on a different runq
1088 * or not. uio_yield() depends on this and the routine is not normally
1089 * called with the same priority otherwise.
4b5f931b
MD
1090 */
1091void
1092lwkt_setpri(thread_t td, int pri)
1093{
26a0694b 1094 KKASSERT(pri >= 0);
a72187e9 1095 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1096 crit_enter();
1097 if (td->td_flags & TDF_RUNQ) {
1098 _lwkt_dequeue(td);
1099 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1100 _lwkt_enqueue(td);
1101 } else {
1102 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1103 }
1104 crit_exit();
1105}
1106
1107void
1108lwkt_setpri_self(int pri)
1109{
1110 thread_t td = curthread;
1111
4b5f931b
MD
1112 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1113 crit_enter();
1114 if (td->td_flags & TDF_RUNQ) {
1115 _lwkt_dequeue(td);
1116 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1117 _lwkt_enqueue(td);
1118 } else {
1119 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1120 }
1121 crit_exit();
1122}
1123
8ec60c3f
MD
1124/*
1125 * Determine if there is a runnable thread at a higher priority then
1126 * the current thread. lwkt_setpri() does not check this automatically.
1127 * Return 1 if there is, 0 if there isn't.
1128 *
1129 * Example: if bit 31 of runqmask is set and the current thread is priority
1130 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1131 *
1132 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1133 * up comparing against 0xffffffff, a comparison that will always be false.
1134 */
1135int
1136lwkt_checkpri_self(void)
1137{
1138 globaldata_t gd = mycpu;
1139 thread_t td = gd->gd_curthread;
1140 int nq = td->td_pri & TDPRI_MASK;
1141
1142 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1143 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1144 return(1);
1145 ++nq;
1146 }
1147 return(0);
1148}
1149
5d21b981 1150/*
52eedfb5
MD
1151 * Migrate the current thread to the specified cpu.
1152 *
1153 * This is accomplished by descheduling ourselves from the current cpu,
1154 * moving our thread to the tdallq of the target cpu, IPI messaging the
1155 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1156 * races while the thread is being migrated.
5d21b981 1157 */
3d28ff59 1158#ifdef SMP
5d21b981 1159static void lwkt_setcpu_remote(void *arg);
3d28ff59 1160#endif
5d21b981
MD
1161
1162void
1163lwkt_setcpu_self(globaldata_t rgd)
1164{
1165#ifdef SMP
1166 thread_t td = curthread;
1167
1168 if (td->td_gd != rgd) {
1169 crit_enter_quick(td);
1170 td->td_flags |= TDF_MIGRATING;
1171 lwkt_deschedule_self(td);
52eedfb5 1172 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1173 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1174 lwkt_switch();
1175 /* we are now on the target cpu */
52eedfb5 1176 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1177 crit_exit_quick(td);
1178 }
1179#endif
1180}
1181
ecdefdda
MD
1182void
1183lwkt_migratecpu(int cpuid)
1184{
1185#ifdef SMP
1186 globaldata_t rgd;
1187
1188 rgd = globaldata_find(cpuid);
1189 lwkt_setcpu_self(rgd);
1190#endif
1191}
1192
5d21b981
MD
1193/*
1194 * Remote IPI for cpu migration (called while in a critical section so we
1195 * do not have to enter another one). The thread has already been moved to
1196 * our cpu's allq, but we must wait for the thread to be completely switched
1197 * out on the originating cpu before we schedule it on ours or the stack
1198 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1199 * change to main memory.
1200 *
1201 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1202 * against wakeups. It is best if this interface is used only when there
1203 * are no pending events that might try to schedule the thread.
1204 */
3d28ff59 1205#ifdef SMP
5d21b981
MD
1206static void
1207lwkt_setcpu_remote(void *arg)
1208{
1209 thread_t td = arg;
1210 globaldata_t gd = mycpu;
1211
c1102e9f 1212 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
35238fa5 1213 cpu_lfence();
5d21b981 1214 td->td_gd = gd;
35238fa5 1215 cpu_sfence();
5d21b981 1216 td->td_flags &= ~TDF_MIGRATING;
9388413d 1217 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1218 _lwkt_enqueue(td);
1219}
3d28ff59 1220#endif
5d21b981 1221
553ea3c8 1222struct lwp *
4b5f931b
MD
1223lwkt_preempted_proc(void)
1224{
73e4f7b9 1225 thread_t td = curthread;
4b5f931b
MD
1226 while (td->td_preempted)
1227 td = td->td_preempted;
553ea3c8 1228 return(td->td_lwp);
4b5f931b
MD
1229}
1230
99df837e
MD
1231/*
1232 * Create a kernel process/thread/whatever. It shares it's address space
1233 * with proc0 - ie: kernel only.
1234 *
365fa13f
MD
1235 * NOTE! By default new threads are created with the MP lock held. A
1236 * thread which does not require the MP lock should release it by calling
1237 * rel_mplock() at the start of the new thread.
99df837e
MD
1238 */
1239int
1240lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1241 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1242 const char *fmt, ...)
99df837e 1243{
73e4f7b9 1244 thread_t td;
e2565a42 1245 __va_list ap;
99df837e 1246
d3d32139 1247 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1248 tdflags);
a2a5ad0d
MD
1249 if (tdp)
1250 *tdp = td;
709799ea 1251 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1252
1253 /*
1254 * Set up arg0 for 'ps' etc
1255 */
e2565a42 1256 __va_start(ap, fmt);
379210cb 1257 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1258 __va_end(ap);
99df837e
MD
1259
1260 /*
1261 * Schedule the thread to run
1262 */
ef0fdad1
MD
1263 if ((td->td_flags & TDF_STOPREQ) == 0)
1264 lwkt_schedule(td);
1265 else
1266 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1267 return 0;
1268}
1269
1270/*
1271 * Destroy an LWKT thread. Warning! This function is not called when
1272 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1273 * uses a different reaping mechanism.
1274 */
1275void
1276lwkt_exit(void)
1277{
1278 thread_t td = curthread;
8826f33a 1279 globaldata_t gd;
99df837e
MD
1280
1281 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1282 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1283 caps_exit(td);
37af14fe
MD
1284 crit_enter_quick(td);
1285 lwkt_deschedule_self(td);
8826f33a 1286 gd = mycpu;
e56e4dea 1287 lwkt_remove_tdallq(td);
8826f33a 1288 if (td->td_flags & TDF_ALLOCATED_THREAD) {
40aaf5fc 1289 objcache_put(thread_cache, td);
8826f33a 1290 }
99df837e
MD
1291 cpu_thread_exit();
1292}
1293
e56e4dea
MD
1294void
1295lwkt_remove_tdallq(thread_t td)
1296{
1297 KKASSERT(td->td_gd == mycpu);
1298 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1299}
1300
2d93b37a
MD
1301void
1302crit_panic(void)
1303{
1304 thread_t td = curthread;
1305 int lpri = td->td_pri;
1306
1307 td->td_pri = 0;
1308 panic("td_pri is/would-go negative! %p %d", td, lpri);
1309}
1310
d165e668
MD
1311#ifdef SMP
1312
bd8015ca
MD
1313/*
1314 * Called from debugger/panic on cpus which have been stopped. We must still
1315 * process the IPIQ while stopped, even if we were stopped while in a critical
1316 * section (XXX).
1317 *
1318 * If we are dumping also try to process any pending interrupts. This may
1319 * or may not work depending on the state of the cpu at the point it was
1320 * stopped.
1321 */
1322void
1323lwkt_smp_stopped(void)
1324{
1325 globaldata_t gd = mycpu;
1326
1327 crit_enter_gd(gd);
1328 if (dumping) {
1329 lwkt_process_ipiq();
1330 splz();
1331 } else {
1332 lwkt_process_ipiq();
1333 }
1334 crit_exit_gd(gd);
1335}
1336
57aa743c
MD
1337/*
1338 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1339 * get_mplock() has already incremented td_mpcount. We must block and
1340 * not return until giant is held.
1341 *
1342 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1343 * reschedule the thread until it can obtain the giant lock for it.
1344 */
1345void
1346lwkt_mp_lock_contested(void)
1347{
57aa743c 1348 loggiant(beg);
57aa743c 1349 lwkt_switch();
57aa743c 1350 loggiant(end);
57aa743c
MD
1351}
1352
d165e668 1353#endif