it's -> its
[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 *
1541028a 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.119 2008/09/11 01:11:42 y0netan1 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;
40aaf5fc 91static struct objcache *thread_cache;
05220613 92
fb0f29c4
MD
93/*
94 * We can make all thread ports use the spin backend instead of the thread
95 * backend. This should only be set to debug the spin backend.
96 */
97TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
98
05220613 99SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
100#ifdef INVARIANTS
101SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
102#endif
4b5f931b 103SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 104SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 105SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 106SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
107#ifdef INVARIANTS
108SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
109 &token_contention_count, 0, "spinning due to token contention");
110SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
111 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
112#endif
05220613 113
57aa743c
MD
114/*
115 * Kernel Trace
116 */
57aa743c
MD
117#if !defined(KTR_GIANT_CONTENTION)
118#define KTR_GIANT_CONTENTION KTR_ALL
119#endif
120
121KTR_INFO_MASTER(giant);
122KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
123KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
124
125#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
126
4b5f931b
MD
127/*
128 * These helper procedures handle the runq, they can only be called from
129 * within a critical section.
75cdbe6c
MD
130 *
131 * WARNING! Prior to SMP being brought up it is possible to enqueue and
132 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
133 * instead of 'mycpu' when referencing the globaldata structure. Once
134 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 135 */
f1d1c3fa
MD
136static __inline
137void
138_lwkt_dequeue(thread_t td)
139{
140 if (td->td_flags & TDF_RUNQ) {
4b5f931b 141 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 142 struct globaldata *gd = td->td_gd;
4b5f931b 143
f1d1c3fa 144 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
145 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
146 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
147 }
148}
149
150static __inline
151void
152_lwkt_enqueue(thread_t td)
153{
344ad853 154 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 155 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 156 struct globaldata *gd = td->td_gd;
4b5f931b 157
f1d1c3fa 158 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
159 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
160 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
161 }
162}
8ad65e08 163
40aaf5fc
NT
164static __boolean_t
165_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
166{
167 struct thread *td = (struct thread *)obj;
168
169 td->td_kstack = NULL;
170 td->td_kstack_size = 0;
171 td->td_flags = TDF_ALLOCATED_THREAD;
172 return (1);
173}
174
175static void
176_lwkt_thread_dtor(void *obj, void *privdata)
177{
178 struct thread *td = (struct thread *)obj;
179
180 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
181 ("_lwkt_thread_dtor: not allocated from objcache"));
182 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
183 td->td_kstack_size > 0,
184 ("_lwkt_thread_dtor: corrupted stack"));
185 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
186}
187
188/*
189 * Initialize the lwkt s/system.
190 */
191void
192lwkt_init(void)
193{
194 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
195 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 0,
196 CACHE_NTHREADS/2, _lwkt_thread_ctor, _lwkt_thread_dtor,
197 NULL);
198}
199
37af14fe
MD
200/*
201 * Schedule a thread to run. As the current thread we can always safely
202 * schedule ourselves, and a shortcut procedure is provided for that
203 * function.
204 *
205 * (non-blocking, self contained on a per cpu basis)
206 */
207void
208lwkt_schedule_self(thread_t td)
209{
210 crit_enter_quick(td);
37af14fe 211 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 212 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 213 _lwkt_enqueue(td);
37af14fe
MD
214 crit_exit_quick(td);
215}
216
217/*
218 * Deschedule a thread.
219 *
220 * (non-blocking, self contained on a per cpu basis)
221 */
222void
223lwkt_deschedule_self(thread_t td)
224{
225 crit_enter_quick(td);
37af14fe
MD
226 _lwkt_dequeue(td);
227 crit_exit_quick(td);
228}
229
8ad65e08
MD
230/*
231 * LWKTs operate on a per-cpu basis
232 *
73e4f7b9 233 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
234 */
235void
236lwkt_gdinit(struct globaldata *gd)
237{
4b5f931b
MD
238 int i;
239
240 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
241 TAILQ_INIT(&gd->gd_tdrunq[i]);
242 gd->gd_runqmask = 0;
73e4f7b9 243 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
244}
245
7d0bac62
MD
246/*
247 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
248 * or LWKT start address before it can be scheduled. If the target cpu is
249 * -1 the thread will be created on the current cpu.
0cfcada1
MD
250 *
251 * If you intend to create a thread without a process context this function
252 * does everything except load the startup and switcher function.
7d0bac62
MD
253 */
254thread_t
d3d32139 255lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 256{
c070746a 257 globaldata_t gd = mycpu;
99df837e 258 void *stack;
7d0bac62 259
c070746a
MD
260 /*
261 * If static thread storage is not supplied allocate a thread. Reuse
262 * a cached free thread if possible. gd_freetd is used to keep an exiting
263 * thread intact through the exit.
264 */
ef0fdad1 265 if (td == NULL) {
c070746a
MD
266 if ((td = gd->gd_freetd) != NULL)
267 gd->gd_freetd = NULL;
268 else
269 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
270 KASSERT((td->td_flags &
271 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
272 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
273 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 274 }
c070746a
MD
275
276 /*
277 * Try to reuse cached stack.
278 */
f470d0c8
MD
279 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
280 if (flags & TDF_ALLOCATED_STACK) {
e4846942 281 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
282 stack = NULL;
283 }
284 }
285 if (stack == NULL) {
e4846942 286 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 287 flags |= TDF_ALLOCATED_STACK;
99df837e 288 }
75cdbe6c 289 if (cpu < 0)
c070746a 290 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 291 else
f470d0c8 292 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 293 return(td);
7d0bac62
MD
294}
295
296/*
297 * Initialize a preexisting thread structure. This function is used by
298 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
299 *
f8c3996b
MD
300 * All threads start out in a critical section at a priority of
301 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
302 * appropriate. This function may send an IPI message when the
303 * requested cpu is not the current cpu and consequently gd_tdallq may
304 * not be initialized synchronously from the point of view of the originating
305 * cpu.
306 *
307 * NOTE! we have to be careful in regards to creating threads for other cpus
308 * if SMP has not yet been activated.
7d0bac62 309 */
41a01a4d
MD
310#ifdef SMP
311
75cdbe6c
MD
312static void
313lwkt_init_thread_remote(void *arg)
314{
315 thread_t td = arg;
316
52eedfb5
MD
317 /*
318 * Protected by critical section held by IPI dispatch
319 */
75cdbe6c
MD
320 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
321}
322
41a01a4d
MD
323#endif
324
7d0bac62 325void
f470d0c8
MD
326lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
327 struct globaldata *gd)
7d0bac62 328{
37af14fe
MD
329 globaldata_t mygd = mycpu;
330
99df837e
MD
331 bzero(td, sizeof(struct thread));
332 td->td_kstack = stack;
f470d0c8 333 td->td_kstack_size = stksize;
d3d32139 334 td->td_flags = flags;
26a0694b 335 td->td_gd = gd;
f8c3996b 336 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
337#ifdef SMP
338 if ((flags & TDF_MPSAFE) == 0)
339 td->td_mpcount = 1;
340#endif
fb0f29c4
MD
341 if (lwkt_use_spin_port)
342 lwkt_initport_spin(&td->td_msgport);
343 else
344 lwkt_initport_thread(&td->td_msgport, td);
99df837e 345 pmap_init_thread(td);
0f7a3396 346#ifdef SMP
5d21b981
MD
347 /*
348 * Normally initializing a thread for a remote cpu requires sending an
349 * IPI. However, the idlethread is setup before the other cpus are
350 * activated so we have to treat it as a special case. XXX manipulation
351 * of gd_tdallq requires the BGL.
352 */
353 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 354 crit_enter_gd(mygd);
75cdbe6c 355 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 356 crit_exit_gd(mygd);
75cdbe6c 357 } else {
2db3b277 358 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 359 }
0f7a3396 360#else
37af14fe 361 crit_enter_gd(mygd);
0f7a3396 362 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 363 crit_exit_gd(mygd);
0f7a3396 364#endif
73e4f7b9
MD
365}
366
367void
368lwkt_set_comm(thread_t td, const char *ctl, ...)
369{
e2565a42 370 __va_list va;
73e4f7b9 371
e2565a42 372 __va_start(va, ctl);
379210cb 373 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 374 __va_end(va);
7d0bac62
MD
375}
376
99df837e 377void
73e4f7b9 378lwkt_hold(thread_t td)
99df837e 379{
73e4f7b9
MD
380 ++td->td_refs;
381}
382
383void
384lwkt_rele(thread_t td)
385{
386 KKASSERT(td->td_refs > 0);
387 --td->td_refs;
388}
389
390void
391lwkt_wait_free(thread_t td)
392{
393 while (td->td_refs)
377d4740 394 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
395}
396
397void
398lwkt_free_thread(thread_t td)
399{
d9eea1a5 400 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
401 ("lwkt_free_thread: did not exit! %p", td));
402
40aaf5fc
NT
403 if (td->td_flags & TDF_ALLOCATED_THREAD) {
404 objcache_put(thread_cache, td);
405 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
406 /* client-allocated struct with internally allocated stack */
407 KASSERT(td->td_kstack && td->td_kstack_size > 0,
408 ("lwkt_free_thread: corrupted stack"));
409 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
410 td->td_kstack = NULL;
411 td->td_kstack_size = 0;
99df837e
MD
412 }
413}
414
415
8ad65e08
MD
416/*
417 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
418 * switch to the idlethread. Switching must occur within a critical
419 * section to avoid races with the scheduling queue.
420 *
421 * We always have full control over our cpu's run queue. Other cpus
422 * that wish to manipulate our queue must use the cpu_*msg() calls to
423 * talk to our cpu, so a critical section is all that is needed and
424 * the result is very, very fast thread switching.
425 *
96728c05
MD
426 * The LWKT scheduler uses a fixed priority model and round-robins at
427 * each priority level. User process scheduling is a totally
428 * different beast and LWKT priorities should not be confused with
429 * user process priorities.
f1d1c3fa 430 *
96728c05
MD
431 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
432 * cleans it up. Note that the td_switch() function cannot do anything that
433 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
434 * the target thread (not the current thread). It's nice to have a scheduler
435 * that does not need the MP lock to work because it allows us to do some
436 * really cool high-performance MP lock optimizations.
69d78e99
MD
437 *
438 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
439 * is not called by the current thread in the preemption case, only when
440 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
441 */
442void
443lwkt_switch(void)
444{
37af14fe
MD
445 globaldata_t gd = mycpu;
446 thread_t td = gd->gd_curthread;
8ad65e08 447 thread_t ntd;
8a8d5d85
MD
448#ifdef SMP
449 int mpheld;
450#endif
8ad65e08 451
46a3f46d 452 /*
27e88a6e
MD
453 * Switching from within a 'fast' (non thread switched) interrupt or IPI
454 * is illegal. However, we may have to do it anyway if we hit a fatal
455 * kernel trap or we have paniced.
456 *
457 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 458 */
27e88a6e
MD
459 if (gd->gd_intr_nesting_level) {
460 int savegdnest;
461 int savegdtrap;
462
463 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
464 panic("lwkt_switch: cannot switch from within "
465 "a fast interrupt, yet, td %p\n", td);
466 } else {
467 savegdnest = gd->gd_intr_nesting_level;
468 savegdtrap = gd->gd_trap_nesting_level;
469 gd->gd_intr_nesting_level = 0;
470 gd->gd_trap_nesting_level = 0;
a7422615
MD
471 if ((td->td_flags & TDF_PANICWARN) == 0) {
472 td->td_flags |= TDF_PANICWARN;
6ea70f76 473 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615
MD
474 "thread %p (%s)\n", td, td->td_comm);
475#ifdef DDB
476 db_print_backtrace();
477#endif
478 }
27e88a6e
MD
479 lwkt_switch();
480 gd->gd_intr_nesting_level = savegdnest;
481 gd->gd_trap_nesting_level = savegdtrap;
482 return;
483 }
96728c05 484 }
ef0fdad1 485
cb973d15
MD
486 /*
487 * Passive release (used to transition from user to kernel mode
488 * when we block or switch rather then when we enter the kernel).
489 * This function is NOT called if we are switching into a preemption
490 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
491 * our current process designation (if we have one) and become a true
492 * LWKT thread, and may also hand the current process designation to
493 * another process and schedule thread.
cb973d15
MD
494 */
495 if (td->td_release)
496 td->td_release(td);
497
37af14fe 498 crit_enter_gd(gd);
9d265729
MD
499 if (td->td_toks)
500 lwkt_relalltokens(td);
501
502 /*
b02926de
MD
503 * We had better not be holding any spin locks, but don't get into an
504 * endless panic loop.
9d265729 505 */
bbb31c5d
MD
506 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
507 ("lwkt_switch: still holding a shared spinlock %p!",
508 gd->gd_spinlock_rd));
d666840a
MD
509 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
510 ("lwkt_switch: still holding %d exclusive spinlocks!",
511 gd->gd_spinlocks_wr));
9d265729 512
8a8d5d85
MD
513
514#ifdef SMP
515 /*
516 * td_mpcount cannot be used to determine if we currently hold the
517 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
518 * to get the lock, and switch out if it can't. Our ownership of
519 * the actual lock will remain stable while we are in a critical section
520 * (but, of course, another cpu may own or release the lock so the
521 * actual value of mp_lock is not stable).
8a8d5d85
MD
522 */
523 mpheld = MP_LOCK_HELD();
0f7a3396
MD
524#ifdef INVARIANTS
525 if (td->td_cscount) {
6ea70f76 526 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
527 td);
528 if (panic_on_cscount)
529 panic("switching while mastering cpusync");
530 }
531#endif
8a8d5d85 532#endif
99df837e
MD
533 if ((ntd = td->td_preempted) != NULL) {
534 /*
535 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
536 * thread. This occurs transparently, whether the preempted thread
537 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
538 * itself).
539 *
540 * We have to setup the MP lock for the original thread after backing
541 * out the adjustment that was made to curthread when the original
542 * was preempted.
99df837e 543 */
26a0694b 544 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 545#ifdef SMP
96728c05 546 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 547 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
548 td, ntd, td->td_mpcount, ntd->td_mpcount);
549 }
8a8d5d85
MD
550 if (ntd->td_mpcount) {
551 td->td_mpcount -= ntd->td_mpcount;
552 KKASSERT(td->td_mpcount >= 0);
553 }
554#endif
26a0694b 555 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
556
557 /*
558 * XXX. The interrupt may have woken a thread up, we need to properly
559 * set the reschedule flag if the originally interrupted thread is at
560 * a lower priority.
561 */
562 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
563 need_lwkt_resched();
8a8d5d85 564 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 565 } else {
4b5f931b
MD
566 /*
567 * Priority queue / round-robin at each priority. Note that user
568 * processes run at a fixed, low priority and the user process
569 * scheduler deals with interactions between user processes
570 * by scheduling and descheduling them from the LWKT queue as
571 * necessary.
8a8d5d85
MD
572 *
573 * We have to adjust the MP lock for the target thread. If we
574 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
575 * thread that does not need the MP lock. If we cannot, we spin
576 * instead of HLT.
577 *
578 * A similar issue exists for the tokens held by the target thread.
579 * If we cannot obtain ownership of the tokens we cannot immediately
580 * schedule the thread.
581 */
582
8ec60c3f
MD
583 /*
584 * If an LWKT reschedule was requested, well that is what we are
585 * doing now so clear it.
586 */
587 clear_lwkt_resched();
4b5f931b
MD
588again:
589 if (gd->gd_runqmask) {
590 int nq = bsrl(gd->gd_runqmask);
591 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
592 gd->gd_runqmask &= ~(1 << nq);
593 goto again;
594 }
8a8d5d85 595#ifdef SMP
41a01a4d 596 /*
df6b8ba0
MD
597 * THREAD SELECTION FOR AN SMP MACHINE BUILD
598 *
41a01a4d
MD
599 * If the target needs the MP lock and we couldn't get it,
600 * or if the target is holding tokens and we could not
601 * gain ownership of the tokens, continue looking for a
602 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
603 *
604 * NOTE: the mpheld variable invalid after this conditional, it
605 * can change due to both cpu_try_mplock() returning success
9d265729 606 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
607 * we are trying to check the mpcount of a thread other then
608 * the current thread. Because of this, if the current thread
609 * is not holding td_mpcount, an IPI indirectly run via
9d265729 610 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 611 * cause the core MP lock to be released.
41a01a4d
MD
612 */
613 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 614 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 615 ) {
8a8d5d85 616 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
617
618 mpheld = MP_LOCK_HELD();
619 ntd = NULL;
8a8d5d85
MD
620 while (rqmask) {
621 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 622 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 623 /* spinning due to MP lock being held */
38717797 624#ifdef INVARIANTS
a453459d 625 ++mplock_contention_count;
38717797 626#endif
a453459d 627 /* mplock still not held, 'mpheld' still valid */
41a01a4d 628 continue;
38717797 629 }
a453459d
MD
630
631 /*
9d265729 632 * mpheld state invalid after getalltokens call returns
a453459d
MD
633 * failure, but the variable is only needed for
634 * the loop.
635 */
9d265729 636 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 637 /* spinning due to token contention */
38717797 638#ifdef INVARIANTS
a453459d 639 ++token_contention_count;
38717797 640#endif
a453459d 641 mpheld = MP_LOCK_HELD();
41a01a4d 642 continue;
38717797 643 }
41a01a4d 644 break;
8a8d5d85
MD
645 }
646 if (ntd)
647 break;
648 rqmask &= ~(1 << nq);
649 nq = bsrl(rqmask);
650 }
651 if (ntd == NULL) {
b402c633 652 cpu_mplock_contested();
a2a5ad0d
MD
653 ntd = &gd->gd_idlethread;
654 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 655 goto using_idle_thread;
8a8d5d85 656 } else {
344ad853 657 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
658 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
659 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
660 }
661 } else {
344ad853 662 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
663 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
664 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
665 }
666#else
df6b8ba0
MD
667 /*
668 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
669 * worry about tokens or the BGL. However, we still have
670 * to call lwkt_getalltokens() in order to properly detect
671 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 672 */
7eb611ef 673 lwkt_getalltokens(ntd);
344ad853 674 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
675 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
676 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 677#endif
4b5f931b 678 } else {
3c23a41a 679 /*
60f945af
MD
680 * We have nothing to run but only let the idle loop halt
681 * the cpu if there are no pending interrupts.
3c23a41a 682 */
a2a5ad0d 683 ntd = &gd->gd_idlethread;
60f945af 684 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 685 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 686#ifdef SMP
df6b8ba0
MD
687using_idle_thread:
688 /*
689 * The idle thread should not be holding the MP lock unless we
690 * are trapping in the kernel or in a panic. Since we select the
691 * idle thread unconditionally when no other thread is available,
692 * if the MP lock is desired during a panic or kernel trap, we
693 * have to loop in the scheduler until we get it.
694 */
695 if (ntd->td_mpcount) {
696 mpheld = MP_LOCK_HELD();
b402c633 697 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 698 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
699 } else if (mpheld == 0) {
700 cpu_mplock_contested();
df6b8ba0 701 goto again;
b402c633 702 }
df6b8ba0 703 }
a453459d 704#endif
4b5f931b 705 }
f1d1c3fa 706 }
26a0694b
MD
707 KASSERT(ntd->td_pri >= TDPRI_CRIT,
708 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
709
710 /*
711 * Do the actual switch. If the new target does not need the MP lock
712 * and we are holding it, release the MP lock. If the new target requires
713 * the MP lock we have already acquired it for the target.
714 */
715#ifdef SMP
716 if (ntd->td_mpcount == 0 ) {
717 if (MP_LOCK_HELD())
718 cpu_rel_mplock();
719 } else {
a453459d 720 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
721 }
722#endif
94f6d86e
MD
723 if (td != ntd) {
724 ++switch_count;
f1d1c3fa 725 td->td_switch(ntd);
94f6d86e 726 }
37af14fe
MD
727 /* NOTE: current cpu may have changed after switch */
728 crit_exit_quick(td);
8ad65e08
MD
729}
730
b68b7282 731/*
96728c05
MD
732 * Request that the target thread preempt the current thread. Preemption
733 * only works under a specific set of conditions:
b68b7282 734 *
96728c05
MD
735 * - We are not preempting ourselves
736 * - The target thread is owned by the current cpu
737 * - We are not currently being preempted
738 * - The target is not currently being preempted
d3d1cbc8
MD
739 * - We are not holding any spin locks
740 * - The target thread is not holding any tokens
96728c05
MD
741 * - We are able to satisfy the target's MP lock requirements (if any).
742 *
743 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
744 * this is called via lwkt_schedule() through the td_preemptable callback.
745 * critpri is the managed critical priority that we should ignore in order
746 * to determine whether preemption is possible (aka usually just the crit
747 * priority of lwkt_schedule() itself).
b68b7282 748 *
26a0694b
MD
749 * XXX at the moment we run the target thread in a critical section during
750 * the preemption in order to prevent the target from taking interrupts
751 * that *WE* can't. Preemption is strictly limited to interrupt threads
752 * and interrupt-like threads, outside of a critical section, and the
753 * preempted source thread will be resumed the instant the target blocks
754 * whether or not the source is scheduled (i.e. preemption is supposed to
755 * be as transparent as possible).
4b5f931b 756 *
8a8d5d85
MD
757 * The target thread inherits our MP count (added to its own) for the
758 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
759 * MP lock during the preemption. Therefore, any preempting targets must be
760 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
761 * out of sync with the physical mp_lock, but we do not have to preserve
762 * the original ownership of the lock if it was out of synch (that is, we
763 * can leave it synchronized on return).
b68b7282
MD
764 */
765void
96728c05 766lwkt_preempt(thread_t ntd, int critpri)
b68b7282 767{
46a3f46d 768 struct globaldata *gd = mycpu;
0a3f9b47 769 thread_t td;
8a8d5d85
MD
770#ifdef SMP
771 int mpheld;
57c254db 772 int savecnt;
8a8d5d85 773#endif
b68b7282 774
26a0694b 775 /*
96728c05
MD
776 * The caller has put us in a critical section. We can only preempt
777 * if the caller of the caller was not in a critical section (basically
d666840a 778 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 779 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 780 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
781 *
782 * YYY The target thread must be in a critical section (else it must
783 * inherit our critical section? I dunno yet).
41a01a4d 784 *
0a3f9b47 785 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
786 */
787 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 788
0a3f9b47 789 td = gd->gd_curthread;
0a3f9b47 790 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
791 ++preempt_miss;
792 return;
793 }
96728c05
MD
794 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
795 ++preempt_miss;
8ec60c3f 796 need_lwkt_resched();
96728c05
MD
797 return;
798 }
799#ifdef SMP
46a3f46d 800 if (ntd->td_gd != gd) {
96728c05 801 ++preempt_miss;
8ec60c3f 802 need_lwkt_resched();
96728c05
MD
803 return;
804 }
805#endif
41a01a4d 806 /*
d3d1cbc8 807 * Take the easy way out and do not preempt if we are holding
d666840a 808 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
809 * preempted interlock against the target thread's tokens and whether
810 * we can get all the target thread's tokens, but this situation
811 * should not occur very often so its easier to simply not preempt.
d666840a
MD
812 * Also, plain spinlocks are impossible to figure out at this point so
813 * just don't preempt.
d3d1cbc8
MD
814 *
815 * Do not try to preempt if the target thread is holding any tokens.
816 * We could try to acquire the tokens but this case is so rare there
817 * is no need to support it.
41a01a4d 818 */
bbb31c5d 819 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 820 ++preempt_miss;
8ec60c3f 821 need_lwkt_resched();
41a01a4d
MD
822 return;
823 }
d3d1cbc8
MD
824 if (ntd->td_toks) {
825 ++preempt_miss;
826 need_lwkt_resched();
827 return;
828 }
26a0694b
MD
829 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
830 ++preempt_weird;
8ec60c3f 831 need_lwkt_resched();
26a0694b
MD
832 return;
833 }
834 if (ntd->td_preempted) {
4b5f931b 835 ++preempt_hit;
8ec60c3f 836 need_lwkt_resched();
26a0694b 837 return;
b68b7282 838 }
8a8d5d85 839#ifdef SMP
a2a5ad0d
MD
840 /*
841 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
842 * to or from the MP lock. In this case td_mpcount will be pre-disposed
843 * (non-zero) but not actually synchronized with the actual state of the
844 * lock. We can use it to imply an MP lock requirement for the
845 * preemption but we cannot use it to test whether we hold the MP lock
846 * or not.
a2a5ad0d 847 */
96728c05 848 savecnt = td->td_mpcount;
71ef2f5c 849 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
850 ntd->td_mpcount += td->td_mpcount;
851 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
852 ntd->td_mpcount -= td->td_mpcount;
853 ++preempt_miss;
8ec60c3f 854 need_lwkt_resched();
8a8d5d85
MD
855 return;
856 }
857#endif
26a0694b 858
8ec60c3f
MD
859 /*
860 * Since we are able to preempt the current thread, there is no need to
861 * call need_lwkt_resched().
862 */
26a0694b
MD
863 ++preempt_hit;
864 ntd->td_preempted = td;
865 td->td_flags |= TDF_PREEMPT_LOCK;
866 td->td_switch(ntd);
867 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
868#ifdef SMP
869 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
870 mpheld = MP_LOCK_HELD();
871 if (mpheld && td->td_mpcount == 0)
96728c05 872 cpu_rel_mplock();
71ef2f5c 873 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
874 panic("lwkt_preempt(): MP lock was not held through");
875#endif
26a0694b
MD
876 ntd->td_preempted = NULL;
877 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
878}
879
f1d1c3fa
MD
880/*
881 * Yield our thread while higher priority threads are pending. This is
882 * typically called when we leave a critical section but it can be safely
883 * called while we are in a critical section.
884 *
885 * This function will not generally yield to equal priority threads but it
886 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 887 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
888 * lwkt_yield_quick().
889 *
235957ed 890 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
891 * came along but was blocked and made pending.
892 *
f1d1c3fa
MD
893 * (self contained on a per cpu basis)
894 */
895void
896lwkt_yield_quick(void)
897{
7966cb69
MD
898 globaldata_t gd = mycpu;
899 thread_t td = gd->gd_curthread;
ef0fdad1 900
a2a5ad0d 901 /*
235957ed 902 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
903 * it with a non-zero cpl then we might not wind up calling splz after
904 * a task switch when the critical section is exited even though the
46a3f46d 905 * new task could accept the interrupt.
a2a5ad0d
MD
906 *
907 * XXX from crit_exit() only called after last crit section is released.
908 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
909 *
910 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
911 * except for this special case, we MUST call splz() here to handle any
912 * pending ints, particularly after we switch, or we might accidently
913 * halt the cpu with interrupts pending.
a2a5ad0d 914 */
46a3f46d 915 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 916 splz();
f1d1c3fa
MD
917
918 /*
919 * YYY enabling will cause wakeup() to task-switch, which really
920 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
921 * preemption and MP without actually doing preemption or MP, because a
922 * lot of code assumes that wakeup() does not block.
f1d1c3fa 923 */
46a3f46d
MD
924 if (untimely_switch && td->td_nest_count == 0 &&
925 gd->gd_intr_nesting_level == 0
926 ) {
37af14fe 927 crit_enter_quick(td);
f1d1c3fa
MD
928 /*
929 * YYY temporary hacks until we disassociate the userland scheduler
930 * from the LWKT scheduler.
931 */
932 if (td->td_flags & TDF_RUNQ) {
933 lwkt_switch(); /* will not reenter yield function */
934 } else {
37af14fe 935 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 936 lwkt_switch(); /* will not reenter yield function */
37af14fe 937 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 938 }
7966cb69 939 crit_exit_noyield(td);
f1d1c3fa 940 }
f1d1c3fa
MD
941}
942
8ad65e08 943/*
f1d1c3fa 944 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 945 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
946 * the crit_exit() call in lwkt_switch().
947 *
948 * (self contained on a per cpu basis)
8ad65e08
MD
949 */
950void
f1d1c3fa 951lwkt_yield(void)
8ad65e08 952{
37af14fe 953 lwkt_schedule_self(curthread);
f1d1c3fa
MD
954 lwkt_switch();
955}
956
8ad65e08 957/*
f1d1c3fa
MD
958 * Generic schedule. Possibly schedule threads belonging to other cpus and
959 * deal with threads that might be blocked on a wait queue.
960 *
0a3f9b47
MD
961 * We have a little helper inline function which does additional work after
962 * the thread has been enqueued, including dealing with preemption and
963 * setting need_lwkt_resched() (which prevents the kernel from returning
964 * to userland until it has processed higher priority threads).
6330a558
MD
965 *
966 * It is possible for this routine to be called after a failed _enqueue
967 * (due to the target thread migrating, sleeping, or otherwise blocked).
968 * We have to check that the thread is actually on the run queue!
361d01dd
MD
969 *
970 * reschedok is an optimized constant propagated from lwkt_schedule() or
971 * lwkt_schedule_noresched(). By default it is non-zero, causing a
972 * reschedule to be requested if the target thread has a higher priority.
973 * The port messaging code will set MSG_NORESCHED and cause reschedok to
974 * be 0, prevented undesired reschedules.
8ad65e08 975 */
0a3f9b47
MD
976static __inline
977void
361d01dd 978_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 979{
c730be20
MD
980 int mypri;
981
6330a558 982 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 983 if (ntd->td_preemptable && reschedok) {
6330a558 984 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 985 } else if (reschedok) {
c730be20
MD
986 /*
987 * This is a little sticky. Due to the passive release function
988 * the LWKT priority can wiggle around for threads acting in
989 * the kernel on behalf of a user process. We do not want this
990 * to effect the comparison per-say.
991 *
992 * What will happen is that the current user process will be
993 * allowed to run until the next hardclock at which time a
994 * forced need_lwkt_resched() will allow the other kernel mode
995 * threads to get in their two cents. This prevents cavitation.
996 */
997 mypri = gd->gd_curthread->td_pri & TDPRI_MASK;
998 if (mypri >= TDPRI_USER_IDLE && mypri <= TDPRI_USER_REAL)
999 mypri = TDPRI_KERN_USER;
1000
1001 if ((ntd->td_pri & TDPRI_MASK) > mypri)
1002 need_lwkt_resched();
6330a558 1003 }
0a3f9b47
MD
1004 }
1005}
1006
361d01dd 1007static __inline
8ad65e08 1008void
361d01dd 1009_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1010{
37af14fe
MD
1011 globaldata_t mygd = mycpu;
1012
41a01a4d 1013 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1014 crit_enter_gd(mygd);
9388413d 1015 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1016 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1017 _lwkt_enqueue(td);
1018 } else {
f1d1c3fa 1019 /*
7cd8d145
MD
1020 * If we own the thread, there is no race (since we are in a
1021 * critical section). If we do not own the thread there might
1022 * be a race but the target cpu will deal with it.
f1d1c3fa 1023 */
0f7a3396 1024#ifdef SMP
7cd8d145 1025 if (td->td_gd == mygd) {
9d265729 1026 _lwkt_enqueue(td);
361d01dd 1027 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1028 } else {
7cd8d145
MD
1029 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1030 }
0f7a3396 1031#else
7cd8d145 1032 _lwkt_enqueue(td);
361d01dd 1033 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1034#endif
8ad65e08 1035 }
37af14fe 1036 crit_exit_gd(mygd);
8ad65e08
MD
1037}
1038
361d01dd
MD
1039void
1040lwkt_schedule(thread_t td)
1041{
1042 _lwkt_schedule(td, 1);
1043}
1044
1045void
1046lwkt_schedule_noresched(thread_t td)
1047{
1048 _lwkt_schedule(td, 0);
1049}
1050
52eedfb5
MD
1051#ifdef SMP
1052
d9eea1a5 1053/*
52eedfb5
MD
1054 * Thread migration using a 'Pull' method. The thread may or may not be
1055 * the current thread. It MUST be descheduled and in a stable state.
1056 * lwkt_giveaway() must be called on the cpu owning the thread.
1057 *
1058 * At any point after lwkt_giveaway() is called, the target cpu may
1059 * 'pull' the thread by calling lwkt_acquire().
1060 *
1061 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1062 */
52eedfb5
MD
1063void
1064lwkt_giveaway(thread_t td)
1065{
1066 globaldata_t gd = mycpu;
1067
1068 crit_enter_gd(gd);
1069 KKASSERT(td->td_gd == gd);
1070 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1071 td->td_flags |= TDF_MIGRATING;
1072 crit_exit_gd(gd);
1073}
1074
a2a5ad0d
MD
1075void
1076lwkt_acquire(thread_t td)
1077{
37af14fe
MD
1078 globaldata_t gd;
1079 globaldata_t mygd;
a2a5ad0d 1080
52eedfb5 1081 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1082 gd = td->td_gd;
37af14fe 1083 mygd = mycpu;
52eedfb5 1084 if (gd != mycpu) {
35238fa5 1085 cpu_lfence();
52eedfb5 1086 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1087 crit_enter_gd(mygd);
df910c23
MD
1088 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1089#ifdef SMP
1090 lwkt_process_ipiq();
1091#endif
52eedfb5 1092 cpu_lfence();
df910c23 1093 }
37af14fe 1094 td->td_gd = mygd;
52eedfb5
MD
1095 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1096 td->td_flags &= ~TDF_MIGRATING;
1097 crit_exit_gd(mygd);
1098 } else {
1099 crit_enter_gd(mygd);
1100 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1101 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1102 crit_exit_gd(mygd);
a2a5ad0d
MD
1103 }
1104}
1105
52eedfb5
MD
1106#endif
1107
f1d1c3fa
MD
1108/*
1109 * Generic deschedule. Descheduling threads other then your own should be
1110 * done only in carefully controlled circumstances. Descheduling is
1111 * asynchronous.
1112 *
1113 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1114 */
1115void
1116lwkt_deschedule(thread_t td)
1117{
f1d1c3fa 1118 crit_enter();
b8a98473 1119#ifdef SMP
f1d1c3fa
MD
1120 if (td == curthread) {
1121 _lwkt_dequeue(td);
1122 } else {
a72187e9 1123 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1124 _lwkt_dequeue(td);
1125 } else {
b8a98473 1126 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1127 }
1128 }
b8a98473
MD
1129#else
1130 _lwkt_dequeue(td);
1131#endif
f1d1c3fa
MD
1132 crit_exit();
1133}
1134
4b5f931b
MD
1135/*
1136 * Set the target thread's priority. This routine does not automatically
1137 * switch to a higher priority thread, LWKT threads are not designed for
1138 * continuous priority changes. Yield if you want to switch.
1139 *
1140 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1141 * of the td_pri field. The specified priority may also indicate zero or
1142 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1143 *
1144 * Note that we requeue the thread whether it winds up on a different runq
1145 * or not. uio_yield() depends on this and the routine is not normally
1146 * called with the same priority otherwise.
4b5f931b
MD
1147 */
1148void
1149lwkt_setpri(thread_t td, int pri)
1150{
26a0694b 1151 KKASSERT(pri >= 0);
a72187e9 1152 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1153 crit_enter();
1154 if (td->td_flags & TDF_RUNQ) {
1155 _lwkt_dequeue(td);
1156 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1157 _lwkt_enqueue(td);
1158 } else {
1159 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1160 }
1161 crit_exit();
1162}
1163
1164void
1165lwkt_setpri_self(int pri)
1166{
1167 thread_t td = curthread;
1168
4b5f931b
MD
1169 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1170 crit_enter();
1171 if (td->td_flags & TDF_RUNQ) {
1172 _lwkt_dequeue(td);
1173 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1174 _lwkt_enqueue(td);
1175 } else {
1176 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1177 }
1178 crit_exit();
1179}
1180
5d21b981 1181/*
52eedfb5
MD
1182 * Migrate the current thread to the specified cpu.
1183 *
1184 * This is accomplished by descheduling ourselves from the current cpu,
1185 * moving our thread to the tdallq of the target cpu, IPI messaging the
1186 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1187 * races while the thread is being migrated.
5d21b981 1188 */
3d28ff59 1189#ifdef SMP
5d21b981 1190static void lwkt_setcpu_remote(void *arg);
3d28ff59 1191#endif
5d21b981
MD
1192
1193void
1194lwkt_setcpu_self(globaldata_t rgd)
1195{
1196#ifdef SMP
1197 thread_t td = curthread;
1198
1199 if (td->td_gd != rgd) {
1200 crit_enter_quick(td);
1201 td->td_flags |= TDF_MIGRATING;
1202 lwkt_deschedule_self(td);
52eedfb5 1203 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1204 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1205 lwkt_switch();
1206 /* we are now on the target cpu */
52eedfb5 1207 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1208 crit_exit_quick(td);
1209 }
1210#endif
1211}
1212
ecdefdda
MD
1213void
1214lwkt_migratecpu(int cpuid)
1215{
1216#ifdef SMP
1217 globaldata_t rgd;
1218
1219 rgd = globaldata_find(cpuid);
1220 lwkt_setcpu_self(rgd);
1221#endif
1222}
1223
5d21b981
MD
1224/*
1225 * Remote IPI for cpu migration (called while in a critical section so we
1226 * do not have to enter another one). The thread has already been moved to
1227 * our cpu's allq, but we must wait for the thread to be completely switched
1228 * out on the originating cpu before we schedule it on ours or the stack
1229 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1230 * change to main memory.
1231 *
1232 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1233 * against wakeups. It is best if this interface is used only when there
1234 * are no pending events that might try to schedule the thread.
1235 */
3d28ff59 1236#ifdef SMP
5d21b981
MD
1237static void
1238lwkt_setcpu_remote(void *arg)
1239{
1240 thread_t td = arg;
1241 globaldata_t gd = mycpu;
1242
df910c23
MD
1243 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1244#ifdef SMP
1245 lwkt_process_ipiq();
1246#endif
35238fa5 1247 cpu_lfence();
df910c23 1248 }
5d21b981 1249 td->td_gd = gd;
35238fa5 1250 cpu_sfence();
5d21b981 1251 td->td_flags &= ~TDF_MIGRATING;
9388413d 1252 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1253 _lwkt_enqueue(td);
1254}
3d28ff59 1255#endif
5d21b981 1256
553ea3c8 1257struct lwp *
4b5f931b
MD
1258lwkt_preempted_proc(void)
1259{
73e4f7b9 1260 thread_t td = curthread;
4b5f931b
MD
1261 while (td->td_preempted)
1262 td = td->td_preempted;
553ea3c8 1263 return(td->td_lwp);
4b5f931b
MD
1264}
1265
99df837e
MD
1266/*
1267 * Create a kernel process/thread/whatever. It shares it's address space
1268 * with proc0 - ie: kernel only.
1269 *
365fa13f
MD
1270 * NOTE! By default new threads are created with the MP lock held. A
1271 * thread which does not require the MP lock should release it by calling
1272 * rel_mplock() at the start of the new thread.
99df837e
MD
1273 */
1274int
1275lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1276 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1277 const char *fmt, ...)
99df837e 1278{
73e4f7b9 1279 thread_t td;
e2565a42 1280 __va_list ap;
99df837e 1281
d3d32139 1282 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1283 tdflags);
a2a5ad0d
MD
1284 if (tdp)
1285 *tdp = td;
709799ea 1286 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1287
1288 /*
1289 * Set up arg0 for 'ps' etc
1290 */
e2565a42 1291 __va_start(ap, fmt);
379210cb 1292 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1293 __va_end(ap);
99df837e
MD
1294
1295 /*
1296 * Schedule the thread to run
1297 */
ef0fdad1
MD
1298 if ((td->td_flags & TDF_STOPREQ) == 0)
1299 lwkt_schedule(td);
1300 else
1301 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1302 return 0;
1303}
1304
1305/*
1306 * Destroy an LWKT thread. Warning! This function is not called when
1307 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1308 * uses a different reaping mechanism.
1309 */
1310void
1311lwkt_exit(void)
1312{
1313 thread_t td = curthread;
c070746a 1314 thread_t std;
8826f33a 1315 globaldata_t gd;
99df837e
MD
1316
1317 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1318 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1319 caps_exit(td);
c070746a
MD
1320
1321 /*
1322 * Get us into a critical section to interlock gd_freetd and loop
1323 * until we can get it freed.
1324 *
1325 * We have to cache the current td in gd_freetd because objcache_put()ing
1326 * it would rip it out from under us while our thread is still active.
1327 */
1328 gd = mycpu;
37af14fe 1329 crit_enter_quick(td);
c070746a
MD
1330 while ((std = gd->gd_freetd) != NULL) {
1331 gd->gd_freetd = NULL;
1332 objcache_put(thread_cache, std);
1333 }
37af14fe 1334 lwkt_deschedule_self(td);
e56e4dea 1335 lwkt_remove_tdallq(td);
c070746a
MD
1336 if (td->td_flags & TDF_ALLOCATED_THREAD)
1337 gd->gd_freetd = td;
99df837e
MD
1338 cpu_thread_exit();
1339}
1340
e56e4dea
MD
1341void
1342lwkt_remove_tdallq(thread_t td)
1343{
1344 KKASSERT(td->td_gd == mycpu);
1345 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1346}
1347
2d93b37a
MD
1348void
1349crit_panic(void)
1350{
1351 thread_t td = curthread;
1352 int lpri = td->td_pri;
1353
1354 td->td_pri = 0;
1355 panic("td_pri is/would-go negative! %p %d", td, lpri);
1356}
1357
d165e668
MD
1358#ifdef SMP
1359
bd8015ca
MD
1360/*
1361 * Called from debugger/panic on cpus which have been stopped. We must still
1362 * process the IPIQ while stopped, even if we were stopped while in a critical
1363 * section (XXX).
1364 *
1365 * If we are dumping also try to process any pending interrupts. This may
1366 * or may not work depending on the state of the cpu at the point it was
1367 * stopped.
1368 */
1369void
1370lwkt_smp_stopped(void)
1371{
1372 globaldata_t gd = mycpu;
1373
1374 crit_enter_gd(gd);
1375 if (dumping) {
1376 lwkt_process_ipiq();
1377 splz();
1378 } else {
1379 lwkt_process_ipiq();
1380 }
1381 crit_exit_gd(gd);
1382}
1383
57aa743c
MD
1384/*
1385 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1386 * get_mplock() has already incremented td_mpcount. We must block and
1387 * not return until giant is held.
1388 *
1389 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1390 * reschedule the thread until it can obtain the giant lock for it.
1391 */
1392void
1393lwkt_mp_lock_contested(void)
1394{
57aa743c 1395 loggiant(beg);
57aa743c 1396 lwkt_switch();
57aa743c 1397 loggiant(end);
57aa743c
MD
1398}
1399
d165e668 1400#endif