kernel - Enhance getcacheblk() (improve saturated write performance (3)).
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
3b998fa9 2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
60f60350 3 *
8c10bfcf
MD
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
60f60350 6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
60f60350 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
60f60350 20 *
8c10bfcf
MD
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
75cdbe6c
MD
33 */
34
35/*
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
8ad65e08
MD
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.h>
46#include <sys/rtprio.h>
b37f18d6 47#include <sys/kinfo.h>
8ad65e08 48#include <sys/queue.h>
7d0bac62 49#include <sys/sysctl.h>
99df837e 50#include <sys/kthread.h>
f1d1c3fa 51#include <machine/cpu.h>
99df837e 52#include <sys/lock.h>
f6bf3af1 53#include <sys/caps.h>
9d265729 54#include <sys/spinlock.h>
57aa743c 55#include <sys/ktr.h>
9d265729
MD
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
684a93c4 59#include <sys/mplock2.h>
f1d1c3fa 60
8c72e3d5
AH
61#include <sys/dsched.h>
62
7d0bac62
MD
63#include <vm/vm.h>
64#include <vm/vm_param.h>
65#include <vm/vm_kern.h>
66#include <vm/vm_object.h>
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
69#include <vm/vm_pager.h>
70#include <vm/vm_extern.h>
7d0bac62 71
99df837e 72#include <machine/stdarg.h>
96728c05 73#include <machine/smp.h>
99df837e 74
d850923c
AE
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
a1f0fb66
AE
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
1541028a 86
40aaf5fc
NT
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
0f7a3396
MD
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
05220613
MD
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
f64b567c 96static __int64_t token_contention_count __debugvar = 0;
fb0f29c4 97static int lwkt_use_spin_port;
40aaf5fc 98static struct objcache *thread_cache;
05220613 99
88ebb169 100#ifdef SMP
e381e77c 101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
88ebb169 102#endif
f9235b6d 103static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
e381e77c 104
0855a2af
JG
105extern void cpu_heavy_restore(void);
106extern void cpu_lwkt_restore(void);
107extern void cpu_kthread_restore(void);
108extern void cpu_idle_restore(void);
109
fb0f29c4
MD
110/*
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
113 */
114TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
115
0f7a3396 116#ifdef INVARIANTS
0c52fa62
SG
117SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
0f7a3396 119#endif
0c52fa62
SG
120SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
9733f757 122SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
0c52fa62 123 "Successful preemption events");
9733f757 124SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
0c52fa62
SG
125 "Failed preemption events");
126SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
38717797
HP
128#ifdef INVARIANTS
129SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
38717797 131#endif
f9235b6d 132static int fairq_enable = 1;
2a418930
MD
133SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
134 &fairq_enable, 0, "Turn on fairq priority accumulators");
135static int lwkt_spin_loops = 10;
136SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
137 &lwkt_spin_loops, 0, "");
138static int lwkt_spin_delay = 1;
139SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
140 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
141static int lwkt_spin_method = 1;
142SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
143 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
d5b2d319
MD
144static int lwkt_spin_fatal = 0; /* disabled */
145SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
146 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
fbc024e4 147static int preempt_enable = 1;
2a418930
MD
148SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
149 &preempt_enable, 0, "Enable preemption");
fbc024e4 150
2a418930
MD
151static __cachealign int lwkt_cseq_rindex;
152static __cachealign int lwkt_cseq_windex;
05220613 153
4b5f931b
MD
154/*
155 * These helper procedures handle the runq, they can only be called from
156 * within a critical section.
75cdbe6c
MD
157 *
158 * WARNING! Prior to SMP being brought up it is possible to enqueue and
159 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
160 * instead of 'mycpu' when referencing the globaldata structure. Once
161 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 162 */
f1d1c3fa
MD
163static __inline
164void
165_lwkt_dequeue(thread_t td)
166{
167 if (td->td_flags & TDF_RUNQ) {
75cdbe6c 168 struct globaldata *gd = td->td_gd;
4b5f931b 169
f1d1c3fa 170 td->td_flags &= ~TDF_RUNQ;
f9235b6d
MD
171 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
172 gd->gd_fairq_total_pri -= td->td_pri;
173 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
2a418930 174 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
f1d1c3fa
MD
175 }
176}
177
f9235b6d
MD
178/*
179 * Priority enqueue.
180 *
181 * NOTE: There are a limited number of lwkt threads runnable since user
182 * processes only schedule one at a time per cpu.
183 */
f1d1c3fa
MD
184static __inline
185void
186_lwkt_enqueue(thread_t td)
187{
f9235b6d
MD
188 thread_t xtd;
189
7f5d7ed7 190 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
75cdbe6c 191 struct globaldata *gd = td->td_gd;
4b5f931b 192
f1d1c3fa 193 td->td_flags |= TDF_RUNQ;
f9235b6d
MD
194 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
195 if (xtd == NULL) {
196 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
2a418930 197 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
f9235b6d
MD
198 } else {
199 while (xtd && xtd->td_pri > td->td_pri)
200 xtd = TAILQ_NEXT(xtd, td_threadq);
201 if (xtd)
202 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
203 else
204 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
205 }
206 gd->gd_fairq_total_pri += td->td_pri;
f1d1c3fa
MD
207 }
208}
8ad65e08 209
40aaf5fc
NT
210static __boolean_t
211_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
212{
213 struct thread *td = (struct thread *)obj;
214
215 td->td_kstack = NULL;
216 td->td_kstack_size = 0;
217 td->td_flags = TDF_ALLOCATED_THREAD;
218 return (1);
219}
220
221static void
222_lwkt_thread_dtor(void *obj, void *privdata)
223{
224 struct thread *td = (struct thread *)obj;
225
226 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
227 ("_lwkt_thread_dtor: not allocated from objcache"));
228 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
229 td->td_kstack_size > 0,
230 ("_lwkt_thread_dtor: corrupted stack"));
231 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
232}
233
234/*
235 * Initialize the lwkt s/system.
236 */
237void
238lwkt_init(void)
239{
240 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
241 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
242 NULL, CACHE_NTHREADS/2,
243 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
244}
245
37af14fe
MD
246/*
247 * Schedule a thread to run. As the current thread we can always safely
248 * schedule ourselves, and a shortcut procedure is provided for that
249 * function.
250 *
251 * (non-blocking, self contained on a per cpu basis)
252 */
253void
254lwkt_schedule_self(thread_t td)
255{
256 crit_enter_quick(td);
f9235b6d
MD
257 KASSERT(td != &td->td_gd->gd_idlethread,
258 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 259 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 260 _lwkt_enqueue(td);
37af14fe
MD
261 crit_exit_quick(td);
262}
263
264/*
265 * Deschedule a thread.
266 *
267 * (non-blocking, self contained on a per cpu basis)
268 */
269void
270lwkt_deschedule_self(thread_t td)
271{
272 crit_enter_quick(td);
37af14fe
MD
273 _lwkt_dequeue(td);
274 crit_exit_quick(td);
275}
276
8ad65e08
MD
277/*
278 * LWKTs operate on a per-cpu basis
279 *
73e4f7b9 280 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
281 */
282void
283lwkt_gdinit(struct globaldata *gd)
284{
f9235b6d 285 TAILQ_INIT(&gd->gd_tdrunq);
73e4f7b9 286 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
287}
288
7d0bac62
MD
289/*
290 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
291 * or LWKT start address before it can be scheduled. If the target cpu is
292 * -1 the thread will be created on the current cpu.
0cfcada1
MD
293 *
294 * If you intend to create a thread without a process context this function
295 * does everything except load the startup and switcher function.
7d0bac62
MD
296 */
297thread_t
d3d32139 298lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 299{
c070746a 300 globaldata_t gd = mycpu;
99df837e 301 void *stack;
7d0bac62 302
c070746a
MD
303 /*
304 * If static thread storage is not supplied allocate a thread. Reuse
305 * a cached free thread if possible. gd_freetd is used to keep an exiting
306 * thread intact through the exit.
307 */
ef0fdad1 308 if (td == NULL) {
cf709dd2
MD
309 crit_enter_gd(gd);
310 if ((td = gd->gd_freetd) != NULL) {
311 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
312 TDF_RUNQ)) == 0);
c070746a 313 gd->gd_freetd = NULL;
cf709dd2 314 } else {
c070746a 315 td = objcache_get(thread_cache, M_WAITOK);
cf709dd2
MD
316 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
317 TDF_RUNQ)) == 0);
318 }
319 crit_exit_gd(gd);
40aaf5fc
NT
320 KASSERT((td->td_flags &
321 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
322 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
323 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 324 }
c070746a
MD
325
326 /*
327 * Try to reuse cached stack.
328 */
f470d0c8
MD
329 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
330 if (flags & TDF_ALLOCATED_STACK) {
e4846942 331 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
332 stack = NULL;
333 }
334 }
335 if (stack == NULL) {
e40cfbd7 336 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
ef0fdad1 337 flags |= TDF_ALLOCATED_STACK;
99df837e 338 }
75cdbe6c 339 if (cpu < 0)
c070746a 340 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 341 else
f470d0c8 342 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 343 return(td);
7d0bac62
MD
344}
345
346/*
347 * Initialize a preexisting thread structure. This function is used by
348 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
349 *
f8c3996b
MD
350 * All threads start out in a critical section at a priority of
351 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
352 * appropriate. This function may send an IPI message when the
353 * requested cpu is not the current cpu and consequently gd_tdallq may
354 * not be initialized synchronously from the point of view of the originating
355 * cpu.
356 *
357 * NOTE! we have to be careful in regards to creating threads for other cpus
358 * if SMP has not yet been activated.
7d0bac62 359 */
41a01a4d
MD
360#ifdef SMP
361
75cdbe6c
MD
362static void
363lwkt_init_thread_remote(void *arg)
364{
365 thread_t td = arg;
366
52eedfb5
MD
367 /*
368 * Protected by critical section held by IPI dispatch
369 */
75cdbe6c
MD
370 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
371}
372
41a01a4d
MD
373#endif
374
fdce8919
MD
375/*
376 * lwkt core thread structural initialization.
377 *
378 * NOTE: All threads are initialized as mpsafe threads.
379 */
7d0bac62 380void
f470d0c8
MD
381lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
382 struct globaldata *gd)
7d0bac62 383{
37af14fe
MD
384 globaldata_t mygd = mycpu;
385
99df837e
MD
386 bzero(td, sizeof(struct thread));
387 td->td_kstack = stack;
f470d0c8 388 td->td_kstack_size = stksize;
d3d32139 389 td->td_flags = flags;
26a0694b 390 td->td_gd = gd;
f9235b6d
MD
391 td->td_pri = TDPRI_KERN_DAEMON;
392 td->td_critcount = 1;
3b998fa9 393 td->td_toks_stop = &td->td_toks_base;
fb0f29c4
MD
394 if (lwkt_use_spin_port)
395 lwkt_initport_spin(&td->td_msgport);
396 else
397 lwkt_initport_thread(&td->td_msgport, td);
99df837e 398 pmap_init_thread(td);
0f7a3396 399#ifdef SMP
5d21b981
MD
400 /*
401 * Normally initializing a thread for a remote cpu requires sending an
402 * IPI. However, the idlethread is setup before the other cpus are
403 * activated so we have to treat it as a special case. XXX manipulation
404 * of gd_tdallq requires the BGL.
405 */
406 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 407 crit_enter_gd(mygd);
75cdbe6c 408 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 409 crit_exit_gd(mygd);
75cdbe6c 410 } else {
2db3b277 411 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 412 }
0f7a3396 413#else
37af14fe 414 crit_enter_gd(mygd);
0f7a3396 415 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 416 crit_exit_gd(mygd);
0f7a3396 417#endif
8c72e3d5
AH
418
419 dsched_new_thread(td);
73e4f7b9
MD
420}
421
422void
423lwkt_set_comm(thread_t td, const char *ctl, ...)
424{
e2565a42 425 __va_list va;
73e4f7b9 426
e2565a42 427 __va_start(va, ctl);
379210cb 428 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 429 __va_end(va);
e7c0dbba 430 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
7d0bac62
MD
431}
432
99df837e 433void
73e4f7b9 434lwkt_hold(thread_t td)
99df837e 435{
74c9628e 436 atomic_add_int(&td->td_refs, 1);
73e4f7b9
MD
437}
438
439void
440lwkt_rele(thread_t td)
441{
442 KKASSERT(td->td_refs > 0);
74c9628e 443 atomic_add_int(&td->td_refs, -1);
73e4f7b9
MD
444}
445
446void
447lwkt_wait_free(thread_t td)
448{
449 while (td->td_refs)
377d4740 450 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
451}
452
453void
454lwkt_free_thread(thread_t td)
455{
74c9628e 456 KKASSERT(td->td_refs == 0);
cf709dd2 457 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
40aaf5fc
NT
458 if (td->td_flags & TDF_ALLOCATED_THREAD) {
459 objcache_put(thread_cache, td);
460 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
461 /* client-allocated struct with internally allocated stack */
462 KASSERT(td->td_kstack && td->td_kstack_size > 0,
463 ("lwkt_free_thread: corrupted stack"));
464 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
465 td->td_kstack = NULL;
466 td->td_kstack_size = 0;
99df837e 467 }
e7c0dbba 468 KTR_LOG(ctxsw_deadtd, td);
99df837e
MD
469}
470
471
8ad65e08
MD
472/*
473 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
474 * switch to the idlethread. Switching must occur within a critical
475 * section to avoid races with the scheduling queue.
476 *
477 * We always have full control over our cpu's run queue. Other cpus
478 * that wish to manipulate our queue must use the cpu_*msg() calls to
479 * talk to our cpu, so a critical section is all that is needed and
480 * the result is very, very fast thread switching.
481 *
96728c05
MD
482 * The LWKT scheduler uses a fixed priority model and round-robins at
483 * each priority level. User process scheduling is a totally
484 * different beast and LWKT priorities should not be confused with
485 * user process priorities.
f1d1c3fa 486 *
3933a3ab
MD
487 * Note that the td_switch() function cannot do anything that requires
488 * the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
489 * the target thread (not the current thread). It's nice to have a scheduler
490 * that does not need the MP lock to work because it allows us to do some
491 * really cool high-performance MP lock optimizations.
69d78e99
MD
492 *
493 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
494 * is not called by the current thread in the preemption case, only when
495 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
496 */
497void
498lwkt_switch(void)
499{
37af14fe
MD
500 globaldata_t gd = mycpu;
501 thread_t td = gd->gd_curthread;
8ad65e08 502 thread_t ntd;
f9235b6d 503 thread_t xtd;
2a418930
MD
504 int spinning = lwkt_spin_loops; /* loops before HLTing */
505 int reqflags;
506 int cseq;
0f0466c0 507 int oseq;
d5b2d319 508 int fatal_count;
8ad65e08 509
46a3f46d 510 /*
27e88a6e
MD
511 * Switching from within a 'fast' (non thread switched) interrupt or IPI
512 * is illegal. However, we may have to do it anyway if we hit a fatal
513 * kernel trap or we have paniced.
514 *
515 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 516 */
27e88a6e
MD
517 if (gd->gd_intr_nesting_level) {
518 int savegdnest;
519 int savegdtrap;
520
5fddbda2 521 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
4a28fe22
MD
522 panic("lwkt_switch: Attempt to switch from a "
523 "a fast interrupt, ipi, or hard code section, "
524 "td %p\n",
525 td);
27e88a6e
MD
526 } else {
527 savegdnest = gd->gd_intr_nesting_level;
528 savegdtrap = gd->gd_trap_nesting_level;
529 gd->gd_intr_nesting_level = 0;
530 gd->gd_trap_nesting_level = 0;
a7422615
MD
531 if ((td->td_flags & TDF_PANICWARN) == 0) {
532 td->td_flags |= TDF_PANICWARN;
4a28fe22
MD
533 kprintf("Warning: thread switch from interrupt, IPI, "
534 "or hard code section.\n"
a7422615 535 "thread %p (%s)\n", td, td->td_comm);
7ce2998e 536 print_backtrace(-1);
a7422615 537 }
27e88a6e
MD
538 lwkt_switch();
539 gd->gd_intr_nesting_level = savegdnest;
540 gd->gd_trap_nesting_level = savegdtrap;
541 return;
542 }
96728c05 543 }
ef0fdad1 544
cb973d15
MD
545 /*
546 * Passive release (used to transition from user to kernel mode
547 * when we block or switch rather then when we enter the kernel).
548 * This function is NOT called if we are switching into a preemption
549 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
550 * our current process designation (if we have one) and become a true
551 * LWKT thread, and may also hand the current process designation to
552 * another process and schedule thread.
cb973d15
MD
553 */
554 if (td->td_release)
555 td->td_release(td);
556
37af14fe 557 crit_enter_gd(gd);
3b998fa9 558 if (TD_TOKS_HELD(td))
9d265729
MD
559 lwkt_relalltokens(td);
560
561 /*
b02926de
MD
562 * We had better not be holding any spin locks, but don't get into an
563 * endless panic loop.
9d265729 564 */
d666840a
MD
565 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
566 ("lwkt_switch: still holding %d exclusive spinlocks!",
567 gd->gd_spinlocks_wr));
9d265729 568
8a8d5d85
MD
569
570#ifdef SMP
0f7a3396
MD
571#ifdef INVARIANTS
572 if (td->td_cscount) {
6ea70f76 573 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
574 td);
575 if (panic_on_cscount)
576 panic("switching while mastering cpusync");
577 }
578#endif
8a8d5d85 579#endif
f9235b6d
MD
580
581 /*
582 * If we had preempted another thread on this cpu, resume the preempted
583 * thread. This occurs transparently, whether the preempted thread
584 * was scheduled or not (it may have been preempted after descheduling
585 * itself).
586 *
587 * We have to setup the MP lock for the original thread after backing
588 * out the adjustment that was made to curthread when the original
589 * was preempted.
590 */
99df837e 591 if ((ntd = td->td_preempted) != NULL) {
26a0694b
MD
592 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
593 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
594
595 /*
b9eb1c19
MD
596 * The interrupt may have woken a thread up, we need to properly
597 * set the reschedule flag if the originally interrupted thread is
598 * at a lower priority.
8ec60c3f 599 */
f9235b6d
MD
600 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
601 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
8ec60c3f 602 need_lwkt_resched();
f9235b6d 603 }
8a8d5d85 604 /* YYY release mp lock on switchback if original doesn't need it */
f9235b6d
MD
605 goto havethread_preempted;
606 }
607
608 /*
609 * Implement round-robin fairq with priority insertion. The priority
610 * insertion is handled by _lwkt_enqueue()
611 *
612 * We have to adjust the MP lock for the target thread. If we
613 * need the MP lock and cannot obtain it we try to locate a
614 * thread that does not need the MP lock. If we cannot, we spin
615 * instead of HLT.
616 *
617 * A similar issue exists for the tokens held by the target thread.
618 * If we cannot obtain ownership of the tokens we cannot immediately
619 * schedule the thread.
620 */
621 for (;;) {
2a418930
MD
622 /*
623 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
624 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
625 * received).
626 */
627 for (;;) {
628 reqflags = gd->gd_reqflags;
629 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
630 (reqflags & ~RQF_AST_LWKT_RESCHED) |
631 RQF_WAKEUP)) {
632 break;
633 }
634 }
f9235b6d 635
4b5f931b 636 /*
2a418930
MD
637 * Hotpath - pull the head of the run queue and attempt to schedule
638 * it. Fairq exhaustion moves the task to the end of the list. If
639 * no threads are runnable we switch to the idle thread.
41a01a4d 640 */
2a418930
MD
641 for (;;) {
642 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
643
644 if (ntd == NULL) {
645 /*
646 * Runq is empty, switch to idle and clear RQF_WAKEUP
647 * to allow it to halt.
648 */
649 ntd = &gd->gd_idlethread;
6f207a2c 650#ifdef SMP
2a418930 651 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
b5d16701 652 ASSERT_NO_TOKENS_HELD(ntd);
6f207a2c 653#endif
2a418930
MD
654 cpu_time.cp_msg[0] = 0;
655 cpu_time.cp_stallpc = 0;
656 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
657 goto haveidle;
658 }
659
660 if (ntd->td_fairq_accum >= 0)
661 break;
662
663 splz_check();
664 lwkt_fairq_accumulate(gd, ntd);
665 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
666 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
f9235b6d 667 }
41a01a4d 668
8ec60c3f 669 /*
2a418930
MD
670 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
671 * unwanted decontention IPIs.
6f207a2c
MD
672 *
673 * NOTE: For UP there is no mplock and lwkt_getalltokens()
674 * always succeeds.
8ec60c3f 675 */
2a418930 676 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 677 goto havethread;
f9235b6d 678
f9235b6d 679 /*
2a418930
MD
680 * Coldpath (SMP only since tokens always succeed on UP)
681 *
682 * We had some contention on the thread we wanted to schedule.
683 * What we do now is try to find a thread that we can schedule
684 * in its stead until decontention reschedules on our cpu.
685 *
686 * The coldpath scan does NOT rearrange threads in the run list
687 * and it also ignores the accumulator.
688 *
689 * We do not immediately schedule a user priority thread, instead
690 * we record it in xtd and continue looking for kernel threads.
691 * A cpu can only have one user priority thread (normally) so just
692 * record the first one.
693 *
694 * NOTE: This scan will also include threads whos fairq's were
695 * accumulated in the first loop.
f9235b6d 696 */
2a418930
MD
697 ++token_contention_count;
698 xtd = NULL;
699 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
41a01a4d 700 /*
2a418930
MD
701 * Try to switch to this thread. If the thread is running at
702 * user priority we clear WAKEUP to allow decontention IPIs
703 * (since this thread is simply running until the one we wanted
704 * decontends), and we make sure that LWKT_RESCHED is not set.
df6b8ba0 705 *
2a418930
MD
706 * Otherwise for kernel threads we leave WAKEUP set to avoid
707 * unnecessary decontention IPIs.
41a01a4d 708 */
2a418930
MD
709 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
710 if (xtd == NULL)
711 xtd = ntd;
712 continue;
f9235b6d 713 }
a453459d 714
f9235b6d 715 /*
2a418930
MD
716 * Do not let the fairq get too negative. Even though we are
717 * ignoring it atm once the scheduler decontends a very negative
718 * thread will get moved to the end of the queue.
f9235b6d 719 */
2a418930
MD
720 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
721 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
722 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
723 goto havethread;
8a8d5d85 724 }
f9235b6d 725
df6b8ba0 726 /*
2a418930 727 * Well fubar, this thread is contended as well, loop
df6b8ba0 728 */
2a418930
MD
729 /* */
730 }
731
732 /*
733 * We exhausted the run list but we may have recorded a user
734 * thread to try. We have three choices based on
735 * lwkt.decontention_method.
736 *
737 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
738 * IPIs (to interrupt the user process) and test
739 * RQF_AST_LWKT_RESCHED at the same time.
740 *
741 * This results in significant decontention IPI traffic but may
742 * be more responsive.
743 *
744 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
745 * An automatic LWKT reschedule will occur on the next hardclock
746 * (typically 100hz).
747 *
748 * This results in no decontention IPI traffic but may be less
749 * responsive. This is the default.
750 *
751 * (2) Refuse to schedule the user process at this time.
752 *
753 * This is highly experimental and should not be used under
754 * normal circumstances. This can cause a user process to
755 * get starved out in situations where kernel threads are
756 * fighting each other for tokens.
757 */
758 if (xtd) {
759 ntd = xtd;
760
761 switch(lwkt_spin_method) {
762 case 0:
763 for (;;) {
764 reqflags = gd->gd_reqflags;
765 if (atomic_cmpset_int(&gd->gd_reqflags,
766 reqflags,
767 reqflags & ~RQF_WAKEUP)) {
768 break;
769 }
770 }
771 break;
772 case 1:
773 reqflags = gd->gd_reqflags;
774 break;
775 default:
776 goto skip;
777 break;
778 }
779 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
b5d16701 780 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
f9235b6d 781 ) {
2a418930
MD
782 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
783 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
784 goto havethread;
df6b8ba0 785 }
9ac1ee6e 786
2a418930 787skip:
9ac1ee6e 788 /*
2a418930
MD
789 * Make sure RQF_WAKEUP is set if we failed to schedule the
790 * user thread to prevent the idle thread from halting.
9ac1ee6e 791 */
2a418930
MD
792 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
793 }
794
795 /*
796 * We exhausted the run list, meaning that all runnable threads
797 * are contended.
798 */
799 cpu_pause();
800 ntd = &gd->gd_idlethread;
801#ifdef SMP
802 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
803 ASSERT_NO_TOKENS_HELD(ntd);
804 /* contention case, do not clear contention mask */
805#endif
806
807 /*
808 * Ok, we might want to spin a few times as some tokens are held for
809 * very short periods of time and IPI overhead is 1uS or worse
810 * (meaning it is usually better to spin). Regardless we have to
811 * call splz_check() to be sure to service any interrupts blocked
812 * by our critical section, otherwise we could livelock e.g. IPIs.
813 *
814 * The IPI mechanic is really a last resort. In nearly all other
815 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
816 *
817 * When we decide not to spin we clear RQF_WAKEUP and switch to
818 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
819 * to halt and decontended tokens will issue an IPI to us. The
820 * idle thread will check for pending reschedules already set
821 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
822 * to here.
823 */
824 if (spinning <= 0) {
825 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
826 goto haveidle;
4b5f931b 827 }
2a418930 828 --spinning;
c5724852
MD
829
830 /*
2a418930
MD
831 * When spinning a delay is required both to avoid livelocks from
832 * token order reversals (a thread may be trying to acquire multiple
833 * tokens), and also to reduce cpu cache management traffic.
834 *
835 * In order to scale to a large number of CPUs we use a time slot
836 * resequencer to force contending cpus into non-contending
837 * time-slots. The scheduler may still contend with the lock holder
838 * but will not (generally) contend with all the other cpus trying
839 * trying to get the same token.
840 *
841 * The resequencer uses a FIFO counter mechanic. The owner of the
842 * rindex at the head of the FIFO is allowed to pull itself off
843 * the FIFO and fetchadd is used to enter into the FIFO. This bit
844 * of code is VERY cache friendly and forces all spinning schedulers
845 * into their own time slots.
c5724852 846 *
2a418930
MD
847 * This code has been tested to 48-cpus and caps the cache
848 * contention load at ~1uS intervals regardless of the number of
849 * cpus. Scaling beyond 64 cpus might require additional smarts
850 * (such as separate FIFOs for specific token cases).
851 *
852 * WARNING! We can't call splz_check() or anything else here as
853 * it could cause a deadlock.
c5724852 854 */
d5b2d319
MD
855#ifdef __amd64__
856 if ((read_rflags() & PSL_I) == 0) {
857 cpu_enable_intr();
858 panic("lwkt_switch() called with interrupts disabled");
859 }
860#endif
2a418930 861 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
d5b2d319 862 fatal_count = lwkt_spin_fatal;
0f0466c0
MD
863 while ((oseq = lwkt_cseq_rindex) != cseq) {
864 cpu_ccfence();
8b402283 865#if !defined(_KERNEL_VIRTUAL)
0f0466c0
MD
866 if (cpu_mi_feature & CPU_MI_MONITOR) {
867 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
8b402283
SW
868 } else
869#endif
870 {
0f0466c0
MD
871 DELAY(1);
872 cpu_lfence();
873 }
d5b2d319
MD
874 if (fatal_count && --fatal_count == 0)
875 panic("lwkt_switch: fatal spin wait");
2a418930
MD
876 }
877 cseq = lwkt_spin_delay; /* don't trust the system operator */
878 cpu_ccfence();
879 if (cseq < 1)
880 cseq = 1;
881 if (cseq > 1000)
882 cseq = 1000;
883 DELAY(cseq);
884 atomic_add_int(&lwkt_cseq_rindex, 1);
c5724852 885 splz_check();
2a418930 886 /* highest level for(;;) loop */
f1d1c3fa 887 }
8a8d5d85 888
2a418930 889havethread:
8a8d5d85 890 /*
f9235b6d
MD
891 * We must always decrement td_fairq_accum on non-idle threads just
892 * in case a thread never gets a tick due to being in a continuous
2a418930 893 * critical section. The page-zeroing code does this, for example.
f9235b6d
MD
894 *
895 * If the thread we came up with is a higher or equal priority verses
896 * the thread at the head of the queue we move our thread to the
897 * front. This way we can always check the front of the queue.
be71787b
MD
898 *
899 * Clear gd_idle_repeat when doing a normal switch to a non-idle
900 * thread.
f9235b6d 901 */
f9235b6d
MD
902 ++gd->gd_cnt.v_swtch;
903 --ntd->td_fairq_accum;
9ac1ee6e 904 ntd->td_wmesg = NULL;
f9235b6d
MD
905 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
906 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
907 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
908 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
909 }
be71787b 910 gd->gd_idle_repeat = 0;
2a418930 911
f9235b6d 912havethread_preempted:
f9235b6d
MD
913 /*
914 * If the new target does not need the MP lock and we are holding it,
915 * release the MP lock. If the new target requires the MP lock we have
916 * already acquired it for the target.
8a8d5d85 917 */
2a418930 918 ;
f9235b6d
MD
919haveidle:
920 KASSERT(ntd->td_critcount,
b5d16701
MD
921 ("priority problem in lwkt_switch %d %d",
922 td->td_critcount, ntd->td_critcount));
923
94f6d86e
MD
924 if (td != ntd) {
925 ++switch_count;
a1f0fb66 926 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
f1d1c3fa 927 td->td_switch(ntd);
94f6d86e 928 }
37af14fe
MD
929 /* NOTE: current cpu may have changed after switch */
930 crit_exit_quick(td);
8ad65e08
MD
931}
932
b68b7282 933/*
96728c05
MD
934 * Request that the target thread preempt the current thread. Preemption
935 * only works under a specific set of conditions:
b68b7282 936 *
96728c05
MD
937 * - We are not preempting ourselves
938 * - The target thread is owned by the current cpu
939 * - We are not currently being preempted
940 * - The target is not currently being preempted
d3d1cbc8
MD
941 * - We are not holding any spin locks
942 * - The target thread is not holding any tokens
96728c05
MD
943 * - We are able to satisfy the target's MP lock requirements (if any).
944 *
945 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
946 * this is called via lwkt_schedule() through the td_preemptable callback.
f9235b6d 947 * critcount is the managed critical priority that we should ignore in order
96728c05
MD
948 * to determine whether preemption is possible (aka usually just the crit
949 * priority of lwkt_schedule() itself).
b68b7282 950 *
26a0694b
MD
951 * XXX at the moment we run the target thread in a critical section during
952 * the preemption in order to prevent the target from taking interrupts
953 * that *WE* can't. Preemption is strictly limited to interrupt threads
954 * and interrupt-like threads, outside of a critical section, and the
955 * preempted source thread will be resumed the instant the target blocks
956 * whether or not the source is scheduled (i.e. preemption is supposed to
957 * be as transparent as possible).
b68b7282
MD
958 */
959void
f9235b6d 960lwkt_preempt(thread_t ntd, int critcount)
b68b7282 961{
46a3f46d 962 struct globaldata *gd = mycpu;
0a3f9b47 963 thread_t td;
2d910aaf 964 int save_gd_intr_nesting_level;
b68b7282 965
26a0694b 966 /*
96728c05
MD
967 * The caller has put us in a critical section. We can only preempt
968 * if the caller of the caller was not in a critical section (basically
f9235b6d 969 * a local interrupt), as determined by the 'critcount' parameter. We
47737962 970 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 971 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
972 *
973 * YYY The target thread must be in a critical section (else it must
974 * inherit our critical section? I dunno yet).
41a01a4d 975 *
0a3f9b47 976 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b 977 */
f9235b6d 978 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 979
fbc024e4
MD
980 if (preempt_enable == 0) {
981 ++preempt_miss;
982 return;
983 }
984
0a3f9b47 985 td = gd->gd_curthread;
f9235b6d 986 if (ntd->td_pri <= td->td_pri) {
57c254db
MD
987 ++preempt_miss;
988 return;
989 }
f9235b6d 990 if (td->td_critcount > critcount) {
96728c05 991 ++preempt_miss;
8ec60c3f 992 need_lwkt_resched();
96728c05
MD
993 return;
994 }
995#ifdef SMP
46a3f46d 996 if (ntd->td_gd != gd) {
96728c05 997 ++preempt_miss;
8ec60c3f 998 need_lwkt_resched();
96728c05
MD
999 return;
1000 }
1001#endif
41a01a4d 1002 /*
77912481
MD
1003 * We don't have to check spinlocks here as they will also bump
1004 * td_critcount.
d3d1cbc8
MD
1005 *
1006 * Do not try to preempt if the target thread is holding any tokens.
1007 * We could try to acquire the tokens but this case is so rare there
1008 * is no need to support it.
41a01a4d 1009 */
77912481
MD
1010 KKASSERT(gd->gd_spinlocks_wr == 0);
1011
3b998fa9 1012 if (TD_TOKS_HELD(ntd)) {
d3d1cbc8
MD
1013 ++preempt_miss;
1014 need_lwkt_resched();
1015 return;
1016 }
26a0694b
MD
1017 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1018 ++preempt_weird;
8ec60c3f 1019 need_lwkt_resched();
26a0694b
MD
1020 return;
1021 }
1022 if (ntd->td_preempted) {
4b5f931b 1023 ++preempt_hit;
8ec60c3f 1024 need_lwkt_resched();
26a0694b 1025 return;
b68b7282 1026 }
26a0694b 1027
8ec60c3f
MD
1028 /*
1029 * Since we are able to preempt the current thread, there is no need to
1030 * call need_lwkt_resched().
2d910aaf
MD
1031 *
1032 * We must temporarily clear gd_intr_nesting_level around the switch
1033 * since switchouts from the target thread are allowed (they will just
1034 * return to our thread), and since the target thread has its own stack.
8ec60c3f 1035 */
26a0694b
MD
1036 ++preempt_hit;
1037 ntd->td_preempted = td;
1038 td->td_flags |= TDF_PREEMPT_LOCK;
a1f0fb66 1039 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
2d910aaf
MD
1040 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1041 gd->gd_intr_nesting_level = 0;
26a0694b 1042 td->td_switch(ntd);
2d910aaf 1043 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
b9eb1c19 1044
26a0694b
MD
1045 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1046 ntd->td_preempted = NULL;
1047 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
1048}
1049
f1d1c3fa 1050/*
faaeffac 1051 * Conditionally call splz() if gd_reqflags indicates work is pending.
4a28fe22
MD
1052 * This will work inside a critical section but not inside a hard code
1053 * section.
ef0fdad1 1054 *
f1d1c3fa
MD
1055 * (self contained on a per cpu basis)
1056 */
1057void
faaeffac 1058splz_check(void)
f1d1c3fa 1059{
7966cb69
MD
1060 globaldata_t gd = mycpu;
1061 thread_t td = gd->gd_curthread;
ef0fdad1 1062
4a28fe22
MD
1063 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1064 gd->gd_intr_nesting_level == 0 &&
1065 td->td_nest_count < 2)
1066 {
f1d1c3fa 1067 splz();
4a28fe22
MD
1068 }
1069}
1070
1071/*
1072 * This version is integrated into crit_exit, reqflags has already
1073 * been tested but td_critcount has not.
1074 *
1075 * We only want to execute the splz() on the 1->0 transition of
1076 * critcount and not in a hard code section or if too deeply nested.
1077 */
1078void
1079lwkt_maybe_splz(thread_t td)
1080{
1081 globaldata_t gd = td->td_gd;
1082
1083 if (td->td_critcount == 0 &&
1084 gd->gd_intr_nesting_level == 0 &&
1085 td->td_nest_count < 2)
1086 {
1087 splz();
1088 }
f1d1c3fa
MD
1089}
1090
8ad65e08 1091/*
f9235b6d
MD
1092 * This function is used to negotiate a passive release of the current
1093 * process/lwp designation with the user scheduler, allowing the user
1094 * scheduler to schedule another user thread. The related kernel thread
1095 * (curthread) continues running in the released state.
8ad65e08
MD
1096 */
1097void
f9235b6d 1098lwkt_passive_release(struct thread *td)
8ad65e08 1099{
f9235b6d
MD
1100 struct lwp *lp = td->td_lwp;
1101
1102 td->td_release = NULL;
1103 lwkt_setpri_self(TDPRI_KERN_USER);
1104 lp->lwp_proc->p_usched->release_curproc(lp);
f1d1c3fa
MD
1105}
1106
f9235b6d 1107
3824f392 1108/*
f9235b6d
MD
1109 * This implements a normal yield. This routine is virtually a nop if
1110 * there is nothing to yield to but it will always run any pending interrupts
1111 * if called from a critical section.
1112 *
1113 * This yield is designed for kernel threads without a user context.
1114 *
1115 * (self contained on a per cpu basis)
3824f392
MD
1116 */
1117void
f9235b6d 1118lwkt_yield(void)
3824f392 1119{
f9235b6d
MD
1120 globaldata_t gd = mycpu;
1121 thread_t td = gd->gd_curthread;
1122 thread_t xtd;
3824f392 1123
f9235b6d
MD
1124 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1125 splz();
1126 if (td->td_fairq_accum < 0) {
1127 lwkt_schedule_self(curthread);
1128 lwkt_switch();
1129 } else {
1130 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1131 if (xtd && xtd->td_pri > td->td_pri) {
1132 lwkt_schedule_self(curthread);
1133 lwkt_switch();
1134 }
1135 }
3824f392
MD
1136}
1137
1138/*
f9235b6d
MD
1139 * This yield is designed for kernel threads with a user context.
1140 *
1141 * The kernel acting on behalf of the user is potentially cpu-bound,
1142 * this function will efficiently allow other threads to run and also
1143 * switch to other processes by releasing.
3824f392
MD
1144 *
1145 * The lwkt_user_yield() function is designed to have very low overhead
1146 * if no yield is determined to be needed.
1147 */
1148void
1149lwkt_user_yield(void)
1150{
f9235b6d
MD
1151 globaldata_t gd = mycpu;
1152 thread_t td = gd->gd_curthread;
1153
1154 /*
1155 * Always run any pending interrupts in case we are in a critical
1156 * section.
1157 */
1158 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1159 splz();
3824f392 1160
3824f392 1161 /*
f9235b6d
MD
1162 * Switch (which forces a release) if another kernel thread needs
1163 * the cpu, if userland wants us to resched, or if our kernel
1164 * quantum has run out.
3824f392 1165 */
f9235b6d
MD
1166 if (lwkt_resched_wanted() ||
1167 user_resched_wanted() ||
1168 td->td_fairq_accum < 0)
1169 {
3824f392 1170 lwkt_switch();
3824f392
MD
1171 }
1172
f9235b6d 1173#if 0
3824f392 1174 /*
f9235b6d
MD
1175 * Reacquire the current process if we are released.
1176 *
1177 * XXX not implemented atm. The kernel may be holding locks and such,
1178 * so we want the thread to continue to receive cpu.
3824f392 1179 */
f9235b6d
MD
1180 if (td->td_release == NULL && lp) {
1181 lp->lwp_proc->p_usched->acquire_curproc(lp);
1182 td->td_release = lwkt_passive_release;
1183 lwkt_setpri_self(TDPRI_USER_NORM);
3824f392 1184 }
f9235b6d 1185#endif
b9eb1c19
MD
1186}
1187
8ad65e08 1188/*
f1d1c3fa
MD
1189 * Generic schedule. Possibly schedule threads belonging to other cpus and
1190 * deal with threads that might be blocked on a wait queue.
1191 *
0a3f9b47
MD
1192 * We have a little helper inline function which does additional work after
1193 * the thread has been enqueued, including dealing with preemption and
1194 * setting need_lwkt_resched() (which prevents the kernel from returning
1195 * to userland until it has processed higher priority threads).
6330a558
MD
1196 *
1197 * It is possible for this routine to be called after a failed _enqueue
1198 * (due to the target thread migrating, sleeping, or otherwise blocked).
1199 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1200 *
1201 * reschedok is an optimized constant propagated from lwkt_schedule() or
1202 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1203 * reschedule to be requested if the target thread has a higher priority.
1204 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1205 * be 0, prevented undesired reschedules.
8ad65e08 1206 */
0a3f9b47
MD
1207static __inline
1208void
f9235b6d 1209_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
0a3f9b47 1210{
b9eb1c19 1211 thread_t otd;
c730be20 1212
6330a558 1213 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1214 if (ntd->td_preemptable && reschedok) {
f9235b6d 1215 ntd->td_preemptable(ntd, ccount); /* YYY +token */
361d01dd 1216 } else if (reschedok) {
b9eb1c19 1217 otd = curthread;
f9235b6d 1218 if (ntd->td_pri > otd->td_pri)
c730be20 1219 need_lwkt_resched();
6330a558 1220 }
f9235b6d
MD
1221
1222 /*
1223 * Give the thread a little fair share scheduler bump if it
1224 * has been asleep for a while. This is primarily to avoid
1225 * a degenerate case for interrupt threads where accumulator
1226 * crosses into negative territory unnecessarily.
1227 */
1228 if (ntd->td_fairq_lticks != ticks) {
1229 ntd->td_fairq_lticks = ticks;
1230 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1231 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1232 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1233 }
0a3f9b47
MD
1234 }
1235}
1236
361d01dd 1237static __inline
8ad65e08 1238void
361d01dd 1239_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1240{
37af14fe
MD
1241 globaldata_t mygd = mycpu;
1242
cf709dd2
MD
1243 KASSERT(td != &td->td_gd->gd_idlethread,
1244 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1245 crit_enter_gd(mygd);
9388413d 1246 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1247 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1248 _lwkt_enqueue(td);
1249 } else {
f1d1c3fa 1250 /*
7cd8d145
MD
1251 * If we own the thread, there is no race (since we are in a
1252 * critical section). If we do not own the thread there might
1253 * be a race but the target cpu will deal with it.
f1d1c3fa 1254 */
0f7a3396 1255#ifdef SMP
7cd8d145 1256 if (td->td_gd == mygd) {
9d265729 1257 _lwkt_enqueue(td);
f9235b6d 1258 _lwkt_schedule_post(mygd, td, 1, reschedok);
f1d1c3fa 1259 } else {
e381e77c 1260 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
7cd8d145 1261 }
0f7a3396 1262#else
7cd8d145 1263 _lwkt_enqueue(td);
f9235b6d 1264 _lwkt_schedule_post(mygd, td, 1, reschedok);
0f7a3396 1265#endif
8ad65e08 1266 }
37af14fe 1267 crit_exit_gd(mygd);
8ad65e08
MD
1268}
1269
361d01dd
MD
1270void
1271lwkt_schedule(thread_t td)
1272{
1273 _lwkt_schedule(td, 1);
1274}
1275
1276void
1277lwkt_schedule_noresched(thread_t td)
1278{
1279 _lwkt_schedule(td, 0);
1280}
1281
88ebb169
SW
1282#ifdef SMP
1283
e381e77c
MD
1284/*
1285 * When scheduled remotely if frame != NULL the IPIQ is being
1286 * run via doreti or an interrupt then preemption can be allowed.
1287 *
1288 * To allow preemption we have to drop the critical section so only
1289 * one is present in _lwkt_schedule_post.
1290 */
1291static void
1292lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1293{
1294 thread_t td = curthread;
1295 thread_t ntd = arg;
1296
1297 if (frame && ntd->td_preemptable) {
1298 crit_exit_noyield(td);
1299 _lwkt_schedule(ntd, 1);
1300 crit_enter_quick(td);
1301 } else {
1302 _lwkt_schedule(ntd, 1);
1303 }
1304}
1305
d9eea1a5 1306/*
52eedfb5
MD
1307 * Thread migration using a 'Pull' method. The thread may or may not be
1308 * the current thread. It MUST be descheduled and in a stable state.
1309 * lwkt_giveaway() must be called on the cpu owning the thread.
1310 *
1311 * At any point after lwkt_giveaway() is called, the target cpu may
1312 * 'pull' the thread by calling lwkt_acquire().
1313 *
ae8e83e6
MD
1314 * We have to make sure the thread is not sitting on a per-cpu tsleep
1315 * queue or it will blow up when it moves to another cpu.
1316 *
52eedfb5 1317 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1318 */
52eedfb5
MD
1319void
1320lwkt_giveaway(thread_t td)
1321{
3b4192fb 1322 globaldata_t gd = mycpu;
52eedfb5 1323
3b4192fb
MD
1324 crit_enter_gd(gd);
1325 if (td->td_flags & TDF_TSLEEPQ)
1326 tsleep_remove(td);
1327 KKASSERT(td->td_gd == gd);
1328 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1329 td->td_flags |= TDF_MIGRATING;
1330 crit_exit_gd(gd);
52eedfb5
MD
1331}
1332
a2a5ad0d
MD
1333void
1334lwkt_acquire(thread_t td)
1335{
37af14fe
MD
1336 globaldata_t gd;
1337 globaldata_t mygd;
a2a5ad0d 1338
52eedfb5 1339 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1340 gd = td->td_gd;
37af14fe 1341 mygd = mycpu;
52eedfb5 1342 if (gd != mycpu) {
35238fa5 1343 cpu_lfence();
52eedfb5 1344 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1345 crit_enter_gd(mygd);
df910c23
MD
1346 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1347#ifdef SMP
1348 lwkt_process_ipiq();
1349#endif
52eedfb5 1350 cpu_lfence();
df910c23 1351 }
562273ea 1352 cpu_mfence();
37af14fe 1353 td->td_gd = mygd;
52eedfb5
MD
1354 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1355 td->td_flags &= ~TDF_MIGRATING;
1356 crit_exit_gd(mygd);
1357 } else {
1358 crit_enter_gd(mygd);
1359 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1360 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1361 crit_exit_gd(mygd);
a2a5ad0d
MD
1362 }
1363}
1364
52eedfb5
MD
1365#endif
1366
f1d1c3fa
MD
1367/*
1368 * Generic deschedule. Descheduling threads other then your own should be
1369 * done only in carefully controlled circumstances. Descheduling is
1370 * asynchronous.
1371 *
1372 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1373 */
1374void
1375lwkt_deschedule(thread_t td)
1376{
f1d1c3fa 1377 crit_enter();
b8a98473 1378#ifdef SMP
f1d1c3fa
MD
1379 if (td == curthread) {
1380 _lwkt_dequeue(td);
1381 } else {
a72187e9 1382 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1383 _lwkt_dequeue(td);
1384 } else {
b8a98473 1385 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1386 }
1387 }
b8a98473
MD
1388#else
1389 _lwkt_dequeue(td);
1390#endif
f1d1c3fa
MD
1391 crit_exit();
1392}
1393
4b5f931b
MD
1394/*
1395 * Set the target thread's priority. This routine does not automatically
1396 * switch to a higher priority thread, LWKT threads are not designed for
1397 * continuous priority changes. Yield if you want to switch.
4b5f931b
MD
1398 */
1399void
1400lwkt_setpri(thread_t td, int pri)
1401{
a72187e9 1402 KKASSERT(td->td_gd == mycpu);
f9235b6d
MD
1403 if (td->td_pri != pri) {
1404 KKASSERT(pri >= 0);
1405 crit_enter();
1406 if (td->td_flags & TDF_RUNQ) {
1407 _lwkt_dequeue(td);
1408 td->td_pri = pri;
1409 _lwkt_enqueue(td);
1410 } else {
1411 td->td_pri = pri;
1412 }
1413 crit_exit();
26a0694b 1414 }
26a0694b
MD
1415}
1416
03bd0a5e
MD
1417/*
1418 * Set the initial priority for a thread prior to it being scheduled for
1419 * the first time. The thread MUST NOT be scheduled before or during
1420 * this call. The thread may be assigned to a cpu other then the current
1421 * cpu.
1422 *
1423 * Typically used after a thread has been created with TDF_STOPPREQ,
1424 * and before the thread is initially scheduled.
1425 */
1426void
1427lwkt_setpri_initial(thread_t td, int pri)
1428{
1429 KKASSERT(pri >= 0);
1430 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
f9235b6d 1431 td->td_pri = pri;
03bd0a5e
MD
1432}
1433
26a0694b
MD
1434void
1435lwkt_setpri_self(int pri)
1436{
1437 thread_t td = curthread;
1438
4b5f931b
MD
1439 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1440 crit_enter();
1441 if (td->td_flags & TDF_RUNQ) {
1442 _lwkt_dequeue(td);
f9235b6d 1443 td->td_pri = pri;
4b5f931b
MD
1444 _lwkt_enqueue(td);
1445 } else {
f9235b6d 1446 td->td_pri = pri;
4b5f931b
MD
1447 }
1448 crit_exit();
1449}
1450
f9235b6d
MD
1451/*
1452 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1453 *
1454 * Example: two competing threads, same priority N. decrement by (2*N)
1455 * increment by N*8, each thread will get 4 ticks.
1456 */
1457void
1458lwkt_fairq_schedulerclock(thread_t td)
1459{
2a418930
MD
1460 globaldata_t gd;
1461
f9235b6d
MD
1462 if (fairq_enable) {
1463 while (td) {
2a418930
MD
1464 gd = td->td_gd;
1465 if (td != &gd->gd_idlethread) {
1466 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1467 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1468 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
f9235b6d
MD
1469 if (td->td_fairq_accum < 0)
1470 need_lwkt_resched();
1471 td->td_fairq_lticks = ticks;
1472 }
1473 td = td->td_preempted;
1474 }
1475 }
1476}
1477
1478static void
1479lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1480{
1481 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1482 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1483 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1484}
1485
5d21b981 1486/*
52eedfb5
MD
1487 * Migrate the current thread to the specified cpu.
1488 *
1489 * This is accomplished by descheduling ourselves from the current cpu,
1490 * moving our thread to the tdallq of the target cpu, IPI messaging the
1491 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1492 * races while the thread is being migrated.
ae8e83e6
MD
1493 *
1494 * We must be sure to remove ourselves from the current cpu's tsleepq
1495 * before potentially moving to another queue. The thread can be on
1496 * a tsleepq due to a left-over tsleep_interlock().
5d21b981 1497 */
3d28ff59 1498#ifdef SMP
5d21b981 1499static void lwkt_setcpu_remote(void *arg);
3d28ff59 1500#endif
5d21b981
MD
1501
1502void
1503lwkt_setcpu_self(globaldata_t rgd)
1504{
1505#ifdef SMP
1506 thread_t td = curthread;
1507
1508 if (td->td_gd != rgd) {
1509 crit_enter_quick(td);
ae8e83e6 1510 if (td->td_flags & TDF_TSLEEPQ)
3b4192fb 1511 tsleep_remove(td);
5d21b981
MD
1512 td->td_flags |= TDF_MIGRATING;
1513 lwkt_deschedule_self(td);
52eedfb5 1514 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1515 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1516 lwkt_switch();
1517 /* we are now on the target cpu */
52eedfb5 1518 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1519 crit_exit_quick(td);
1520 }
1521#endif
1522}
1523
ecdefdda
MD
1524void
1525lwkt_migratecpu(int cpuid)
1526{
1527#ifdef SMP
1528 globaldata_t rgd;
1529
1530 rgd = globaldata_find(cpuid);
1531 lwkt_setcpu_self(rgd);
1532#endif
1533}
1534
5d21b981
MD
1535/*
1536 * Remote IPI for cpu migration (called while in a critical section so we
1537 * do not have to enter another one). The thread has already been moved to
1538 * our cpu's allq, but we must wait for the thread to be completely switched
1539 * out on the originating cpu before we schedule it on ours or the stack
1540 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1541 * change to main memory.
1542 *
1543 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1544 * against wakeups. It is best if this interface is used only when there
1545 * are no pending events that might try to schedule the thread.
1546 */
3d28ff59 1547#ifdef SMP
5d21b981
MD
1548static void
1549lwkt_setcpu_remote(void *arg)
1550{
1551 thread_t td = arg;
1552 globaldata_t gd = mycpu;
1553
df910c23
MD
1554 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1555#ifdef SMP
1556 lwkt_process_ipiq();
1557#endif
35238fa5 1558 cpu_lfence();
562273ea 1559 cpu_pause();
df910c23 1560 }
5d21b981 1561 td->td_gd = gd;
562273ea 1562 cpu_mfence();
5d21b981 1563 td->td_flags &= ~TDF_MIGRATING;
9388413d 1564 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1565 _lwkt_enqueue(td);
1566}
3d28ff59 1567#endif
5d21b981 1568
553ea3c8 1569struct lwp *
4b5f931b
MD
1570lwkt_preempted_proc(void)
1571{
73e4f7b9 1572 thread_t td = curthread;
4b5f931b
MD
1573 while (td->td_preempted)
1574 td = td->td_preempted;
553ea3c8 1575 return(td->td_lwp);
4b5f931b
MD
1576}
1577
99df837e
MD
1578/*
1579 * Create a kernel process/thread/whatever. It shares it's address space
1580 * with proc0 - ie: kernel only.
1581 *
365fa13f
MD
1582 * NOTE! By default new threads are created with the MP lock held. A
1583 * thread which does not require the MP lock should release it by calling
1584 * rel_mplock() at the start of the new thread.
99df837e
MD
1585 */
1586int
c9e9fb21
MD
1587lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1588 thread_t template, int tdflags, int cpu, const char *fmt, ...)
99df837e 1589{
73e4f7b9 1590 thread_t td;
e2565a42 1591 __va_list ap;
99df837e 1592
d3d32139 1593 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1594 tdflags);
a2a5ad0d
MD
1595 if (tdp)
1596 *tdp = td;
709799ea 1597 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1598
1599 /*
1600 * Set up arg0 for 'ps' etc
1601 */
e2565a42 1602 __va_start(ap, fmt);
379210cb 1603 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1604 __va_end(ap);
99df837e
MD
1605
1606 /*
1607 * Schedule the thread to run
1608 */
ef0fdad1
MD
1609 if ((td->td_flags & TDF_STOPREQ) == 0)
1610 lwkt_schedule(td);
1611 else
1612 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1613 return 0;
1614}
1615
1616/*
1617 * Destroy an LWKT thread. Warning! This function is not called when
1618 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1619 * uses a different reaping mechanism.
1620 */
1621void
1622lwkt_exit(void)
1623{
1624 thread_t td = curthread;
c070746a 1625 thread_t std;
8826f33a 1626 globaldata_t gd;
99df837e 1627
2883d2d8
MD
1628 /*
1629 * Do any cleanup that might block here
1630 */
99df837e 1631 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1632 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1633 caps_exit(td);
2883d2d8
MD
1634 biosched_done(td);
1635 dsched_exit_thread(td);
c070746a
MD
1636
1637 /*
1638 * Get us into a critical section to interlock gd_freetd and loop
1639 * until we can get it freed.
1640 *
1641 * We have to cache the current td in gd_freetd because objcache_put()ing
1642 * it would rip it out from under us while our thread is still active.
1643 */
1644 gd = mycpu;
37af14fe 1645 crit_enter_quick(td);
c070746a 1646 while ((std = gd->gd_freetd) != NULL) {
cf709dd2 1647 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
c070746a
MD
1648 gd->gd_freetd = NULL;
1649 objcache_put(thread_cache, std);
1650 }
3b4192fb
MD
1651
1652 /*
1653 * Remove thread resources from kernel lists and deschedule us for
2883d2d8
MD
1654 * the last time. We cannot block after this point or we may end
1655 * up with a stale td on the tsleepq.
3b4192fb
MD
1656 */
1657 if (td->td_flags & TDF_TSLEEPQ)
1658 tsleep_remove(td);
37af14fe 1659 lwkt_deschedule_self(td);
e56e4dea 1660 lwkt_remove_tdallq(td);
74c9628e 1661 KKASSERT(td->td_refs == 0);
2883d2d8
MD
1662
1663 /*
1664 * Final cleanup
1665 */
1666 KKASSERT(gd->gd_freetd == NULL);
c070746a
MD
1667 if (td->td_flags & TDF_ALLOCATED_THREAD)
1668 gd->gd_freetd = td;
99df837e
MD
1669 cpu_thread_exit();
1670}
1671
e56e4dea
MD
1672void
1673lwkt_remove_tdallq(thread_t td)
1674{
1675 KKASSERT(td->td_gd == mycpu);
1676 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1677}
1678
9cf43f91
MD
1679/*
1680 * Code reduction and branch prediction improvements. Call/return
1681 * overhead on modern cpus often degenerates into 0 cycles due to
1682 * the cpu's branch prediction hardware and return pc cache. We
1683 * can take advantage of this by not inlining medium-complexity
1684 * functions and we can also reduce the branch prediction impact
1685 * by collapsing perfectly predictable branches into a single
1686 * procedure instead of duplicating it.
1687 *
1688 * Is any of this noticeable? Probably not, so I'll take the
1689 * smaller code size.
1690 */
1691void
b6468f56 1692crit_exit_wrapper(__DEBUG_CRIT_ARG__)
9cf43f91 1693{
b6468f56 1694 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
9cf43f91
MD
1695}
1696
2d93b37a
MD
1697void
1698crit_panic(void)
1699{
1700 thread_t td = curthread;
850634cc 1701 int lcrit = td->td_critcount;
2d93b37a 1702
850634cc
AH
1703 td->td_critcount = 0;
1704 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
4a28fe22 1705 /* NOT REACHED */
2d93b37a
MD
1706}
1707
d165e668
MD
1708#ifdef SMP
1709
bd8015ca
MD
1710/*
1711 * Called from debugger/panic on cpus which have been stopped. We must still
1712 * process the IPIQ while stopped, even if we were stopped while in a critical
1713 * section (XXX).
1714 *
1715 * If we are dumping also try to process any pending interrupts. This may
1716 * or may not work depending on the state of the cpu at the point it was
1717 * stopped.
1718 */
1719void
1720lwkt_smp_stopped(void)
1721{
1722 globaldata_t gd = mycpu;
1723
1724 crit_enter_gd(gd);
1725 if (dumping) {
1726 lwkt_process_ipiq();
1727 splz();
1728 } else {
1729 lwkt_process_ipiq();
1730 }
1731 crit_exit_gd(gd);
1732}
1733
d165e668 1734#endif