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