HAMMER UTIL - Add -C cachesize option to improve the 'show' command.
[dragonfly.git] / sys / kern / lwkt_thread.c
CommitLineData
8ad65e08 1/*
8c10bfcf
MD
2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
8ad65e08
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
8c10bfcf 10 *
8ad65e08
MD
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
8c10bfcf
MD
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
8ad65e08 32 * SUCH DAMAGE.
8c10bfcf 33 *
0aa16b5d 34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe Exp $
75cdbe6c
MD
35 */
36
37/*
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
8ad65e08
MD
42 */
43
44#include <sys/param.h>
45#include <sys/systm.h>
46#include <sys/kernel.h>
47#include <sys/proc.h>
48#include <sys/rtprio.h>
49#include <sys/queue.h>
7d0bac62 50#include <sys/sysctl.h>
99df837e 51#include <sys/kthread.h>
f1d1c3fa 52#include <machine/cpu.h>
99df837e 53#include <sys/lock.h>
f6bf3af1 54#include <sys/caps.h>
9d265729 55#include <sys/spinlock.h>
57aa743c 56#include <sys/ktr.h>
9d265729
MD
57
58#include <sys/thread2.h>
59#include <sys/spinlock2.h>
f1d1c3fa 60
7d0bac62
MD
61#include <vm/vm.h>
62#include <vm/vm_param.h>
63#include <vm/vm_kern.h>
64#include <vm/vm_object.h>
65#include <vm/vm_page.h>
66#include <vm/vm_map.h>
67#include <vm/vm_pager.h>
68#include <vm/vm_extern.h>
7d0bac62 69
99df837e 70#include <machine/stdarg.h>
96728c05 71#include <machine/smp.h>
99df837e 72
1541028a 73
40aaf5fc
NT
74static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
75
7d0bac62 76static int untimely_switch = 0;
0f7a3396
MD
77#ifdef INVARIANTS
78static int panic_on_cscount = 0;
79#endif
05220613
MD
80static __int64_t switch_count = 0;
81static __int64_t preempt_hit = 0;
82static __int64_t preempt_miss = 0;
83static __int64_t preempt_weird = 0;
38717797
HP
84static __int64_t token_contention_count = 0;
85static __int64_t mplock_contention_count = 0;
fb0f29c4 86static int lwkt_use_spin_port;
d2f86ad2 87#ifdef SMP
b9eb1c19 88static int chain_mplock = 0;
d2f86ad2 89#endif
40aaf5fc 90static struct objcache *thread_cache;
05220613 91
b9eb1c19
MD
92volatile cpumask_t mp_lock_contention_mask;
93
0855a2af
JG
94extern void cpu_heavy_restore(void);
95extern void cpu_lwkt_restore(void);
96extern void cpu_kthread_restore(void);
97extern void cpu_idle_restore(void);
98
85514115
MD
99#ifdef __amd64__
100
101static int
0855a2af
JG
102jg_tos_ok(struct thread *td)
103{
85514115
MD
104 void *tos;
105 int tos_ok;
106
0855a2af
JG
107 if (td == NULL) {
108 return 1;
109 }
110 KKASSERT(td->td_sp != NULL);
85514115
MD
111 tos = ((void **)td->td_sp)[0];
112 tos_ok = 0;
113 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
114 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
0855a2af
JG
115 tos_ok = 1;
116 }
117 return tos_ok;
118}
119
85514115
MD
120#endif
121
fb0f29c4
MD
122/*
123 * We can make all thread ports use the spin backend instead of the thread
124 * backend. This should only be set to debug the spin backend.
125 */
126TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
127
05220613 128SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
0f7a3396
MD
129#ifdef INVARIANTS
130SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
131#endif
b9eb1c19
MD
132#ifdef SMP
133SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
134#endif
4b5f931b 135SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
4b5f931b 136SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
4b5f931b 137SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
26a0694b 138SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
38717797
HP
139#ifdef INVARIANTS
140SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
141 &token_contention_count, 0, "spinning due to token contention");
142SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
143 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
144#endif
05220613 145
57aa743c
MD
146/*
147 * Kernel Trace
148 */
57aa743c
MD
149#if !defined(KTR_GIANT_CONTENTION)
150#define KTR_GIANT_CONTENTION KTR_ALL
151#endif
152
153KTR_INFO_MASTER(giant);
154KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
155KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
156
157#define loggiant(name) KTR_LOG(giant_ ## name, curthread)
158
4b5f931b
MD
159/*
160 * These helper procedures handle the runq, they can only be called from
161 * within a critical section.
75cdbe6c
MD
162 *
163 * WARNING! Prior to SMP being brought up it is possible to enqueue and
164 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
165 * instead of 'mycpu' when referencing the globaldata structure. Once
166 * SMP live enqueuing and dequeueing only occurs on the current cpu.
4b5f931b 167 */
f1d1c3fa
MD
168static __inline
169void
170_lwkt_dequeue(thread_t td)
171{
172 if (td->td_flags & TDF_RUNQ) {
4b5f931b 173 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 174 struct globaldata *gd = td->td_gd;
4b5f931b 175
f1d1c3fa 176 td->td_flags &= ~TDF_RUNQ;
4b5f931b
MD
177 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
178 /* runqmask is passively cleaned up by the switcher */
f1d1c3fa
MD
179 }
180}
181
182static __inline
183void
184_lwkt_enqueue(thread_t td)
185{
344ad853 186 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
4b5f931b 187 int nq = td->td_pri & TDPRI_MASK;
75cdbe6c 188 struct globaldata *gd = td->td_gd;
4b5f931b 189
f1d1c3fa 190 td->td_flags |= TDF_RUNQ;
4b5f931b
MD
191 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
192 gd->gd_runqmask |= 1 << nq;
f1d1c3fa
MD
193 }
194}
8ad65e08 195
40aaf5fc
NT
196static __boolean_t
197_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
198{
199 struct thread *td = (struct thread *)obj;
200
201 td->td_kstack = NULL;
202 td->td_kstack_size = 0;
203 td->td_flags = TDF_ALLOCATED_THREAD;
204 return (1);
205}
206
207static void
208_lwkt_thread_dtor(void *obj, void *privdata)
209{
210 struct thread *td = (struct thread *)obj;
211
212 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
213 ("_lwkt_thread_dtor: not allocated from objcache"));
214 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
215 td->td_kstack_size > 0,
216 ("_lwkt_thread_dtor: corrupted stack"));
217 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
218}
219
220/*
221 * Initialize the lwkt s/system.
222 */
223void
224lwkt_init(void)
225{
226 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
0aa16b5d
SZ
227 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
228 NULL, CACHE_NTHREADS/2,
229 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
40aaf5fc
NT
230}
231
37af14fe
MD
232/*
233 * Schedule a thread to run. As the current thread we can always safely
234 * schedule ourselves, and a shortcut procedure is provided for that
235 * function.
236 *
237 * (non-blocking, self contained on a per cpu basis)
238 */
239void
240lwkt_schedule_self(thread_t td)
241{
242 crit_enter_quick(td);
37af14fe 243 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
9388413d 244 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 245 _lwkt_enqueue(td);
37af14fe
MD
246 crit_exit_quick(td);
247}
248
249/*
250 * Deschedule a thread.
251 *
252 * (non-blocking, self contained on a per cpu basis)
253 */
254void
255lwkt_deschedule_self(thread_t td)
256{
257 crit_enter_quick(td);
37af14fe
MD
258 _lwkt_dequeue(td);
259 crit_exit_quick(td);
260}
261
8ad65e08
MD
262/*
263 * LWKTs operate on a per-cpu basis
264 *
73e4f7b9 265 * WARNING! Called from early boot, 'mycpu' may not work yet.
8ad65e08
MD
266 */
267void
268lwkt_gdinit(struct globaldata *gd)
269{
4b5f931b
MD
270 int i;
271
272 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
273 TAILQ_INIT(&gd->gd_tdrunq[i]);
274 gd->gd_runqmask = 0;
73e4f7b9 275 TAILQ_INIT(&gd->gd_tdallq);
8ad65e08
MD
276}
277
7d0bac62
MD
278/*
279 * Create a new thread. The thread must be associated with a process context
75cdbe6c
MD
280 * or LWKT start address before it can be scheduled. If the target cpu is
281 * -1 the thread will be created on the current cpu.
0cfcada1
MD
282 *
283 * If you intend to create a thread without a process context this function
284 * does everything except load the startup and switcher function.
7d0bac62
MD
285 */
286thread_t
d3d32139 287lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
7d0bac62 288{
c070746a 289 globaldata_t gd = mycpu;
99df837e 290 void *stack;
7d0bac62 291
c070746a
MD
292 /*
293 * If static thread storage is not supplied allocate a thread. Reuse
294 * a cached free thread if possible. gd_freetd is used to keep an exiting
295 * thread intact through the exit.
296 */
ef0fdad1 297 if (td == NULL) {
c070746a
MD
298 if ((td = gd->gd_freetd) != NULL)
299 gd->gd_freetd = NULL;
300 else
301 td = objcache_get(thread_cache, M_WAITOK);
40aaf5fc
NT
302 KASSERT((td->td_flags &
303 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
304 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
305 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
ef0fdad1 306 }
c070746a
MD
307
308 /*
309 * Try to reuse cached stack.
310 */
f470d0c8
MD
311 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
312 if (flags & TDF_ALLOCATED_STACK) {
e4846942 313 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
f470d0c8
MD
314 stack = NULL;
315 }
316 }
317 if (stack == NULL) {
e4846942 318 stack = (void *)kmem_alloc(&kernel_map, stksize);
ef0fdad1 319 flags |= TDF_ALLOCATED_STACK;
99df837e 320 }
75cdbe6c 321 if (cpu < 0)
c070746a 322 lwkt_init_thread(td, stack, stksize, flags, gd);
75cdbe6c 323 else
f470d0c8 324 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
99df837e 325 return(td);
7d0bac62
MD
326}
327
328/*
329 * Initialize a preexisting thread structure. This function is used by
330 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
331 *
f8c3996b
MD
332 * All threads start out in a critical section at a priority of
333 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
75cdbe6c
MD
334 * appropriate. This function may send an IPI message when the
335 * requested cpu is not the current cpu and consequently gd_tdallq may
336 * not be initialized synchronously from the point of view of the originating
337 * cpu.
338 *
339 * NOTE! we have to be careful in regards to creating threads for other cpus
340 * if SMP has not yet been activated.
7d0bac62 341 */
41a01a4d
MD
342#ifdef SMP
343
75cdbe6c
MD
344static void
345lwkt_init_thread_remote(void *arg)
346{
347 thread_t td = arg;
348
52eedfb5
MD
349 /*
350 * Protected by critical section held by IPI dispatch
351 */
75cdbe6c
MD
352 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
353}
354
41a01a4d
MD
355#endif
356
7d0bac62 357void
f470d0c8
MD
358lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
359 struct globaldata *gd)
7d0bac62 360{
37af14fe
MD
361 globaldata_t mygd = mycpu;
362
99df837e
MD
363 bzero(td, sizeof(struct thread));
364 td->td_kstack = stack;
f470d0c8 365 td->td_kstack_size = stksize;
d3d32139 366 td->td_flags = flags;
26a0694b 367 td->td_gd = gd;
f8c3996b 368 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
d3d32139
MD
369#ifdef SMP
370 if ((flags & TDF_MPSAFE) == 0)
371 td->td_mpcount = 1;
372#endif
fb0f29c4
MD
373 if (lwkt_use_spin_port)
374 lwkt_initport_spin(&td->td_msgport);
375 else
376 lwkt_initport_thread(&td->td_msgport, td);
99df837e 377 pmap_init_thread(td);
0f7a3396 378#ifdef SMP
5d21b981
MD
379 /*
380 * Normally initializing a thread for a remote cpu requires sending an
381 * IPI. However, the idlethread is setup before the other cpus are
382 * activated so we have to treat it as a special case. XXX manipulation
383 * of gd_tdallq requires the BGL.
384 */
385 if (gd == mygd || td == &gd->gd_idlethread) {
37af14fe 386 crit_enter_gd(mygd);
75cdbe6c 387 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 388 crit_exit_gd(mygd);
75cdbe6c 389 } else {
2db3b277 390 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
75cdbe6c 391 }
0f7a3396 392#else
37af14fe 393 crit_enter_gd(mygd);
0f7a3396 394 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
37af14fe 395 crit_exit_gd(mygd);
0f7a3396 396#endif
73e4f7b9
MD
397}
398
399void
400lwkt_set_comm(thread_t td, const char *ctl, ...)
401{
e2565a42 402 __va_list va;
73e4f7b9 403
e2565a42 404 __va_start(va, ctl);
379210cb 405 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
e2565a42 406 __va_end(va);
7d0bac62
MD
407}
408
99df837e 409void
73e4f7b9 410lwkt_hold(thread_t td)
99df837e 411{
73e4f7b9
MD
412 ++td->td_refs;
413}
414
415void
416lwkt_rele(thread_t td)
417{
418 KKASSERT(td->td_refs > 0);
419 --td->td_refs;
420}
421
422void
423lwkt_wait_free(thread_t td)
424{
425 while (td->td_refs)
377d4740 426 tsleep(td, 0, "tdreap", hz);
73e4f7b9
MD
427}
428
429void
430lwkt_free_thread(thread_t td)
431{
d9eea1a5 432 KASSERT((td->td_flags & TDF_RUNNING) == 0,
99df837e
MD
433 ("lwkt_free_thread: did not exit! %p", td));
434
40aaf5fc
NT
435 if (td->td_flags & TDF_ALLOCATED_THREAD) {
436 objcache_put(thread_cache, td);
437 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
438 /* client-allocated struct with internally allocated stack */
439 KASSERT(td->td_kstack && td->td_kstack_size > 0,
440 ("lwkt_free_thread: corrupted stack"));
441 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
442 td->td_kstack = NULL;
443 td->td_kstack_size = 0;
99df837e
MD
444 }
445}
446
447
8ad65e08
MD
448/*
449 * Switch to the next runnable lwkt. If no LWKTs are runnable then
f1d1c3fa
MD
450 * switch to the idlethread. Switching must occur within a critical
451 * section to avoid races with the scheduling queue.
452 *
453 * We always have full control over our cpu's run queue. Other cpus
454 * that wish to manipulate our queue must use the cpu_*msg() calls to
455 * talk to our cpu, so a critical section is all that is needed and
456 * the result is very, very fast thread switching.
457 *
96728c05
MD
458 * The LWKT scheduler uses a fixed priority model and round-robins at
459 * each priority level. User process scheduling is a totally
460 * different beast and LWKT priorities should not be confused with
461 * user process priorities.
f1d1c3fa 462 *
96728c05
MD
463 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
464 * cleans it up. Note that the td_switch() function cannot do anything that
465 * requires the MP lock since the MP lock will have already been setup for
71ef2f5c
MD
466 * the target thread (not the current thread). It's nice to have a scheduler
467 * that does not need the MP lock to work because it allows us to do some
468 * really cool high-performance MP lock optimizations.
69d78e99
MD
469 *
470 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
471 * is not called by the current thread in the preemption case, only when
472 * the preempting thread blocks (in order to return to the original thread).
8ad65e08
MD
473 */
474void
475lwkt_switch(void)
476{
37af14fe
MD
477 globaldata_t gd = mycpu;
478 thread_t td = gd->gd_curthread;
8ad65e08 479 thread_t ntd;
8a8d5d85
MD
480#ifdef SMP
481 int mpheld;
482#endif
8ad65e08 483
46a3f46d 484 /*
27e88a6e
MD
485 * Switching from within a 'fast' (non thread switched) interrupt or IPI
486 * is illegal. However, we may have to do it anyway if we hit a fatal
487 * kernel trap or we have paniced.
488 *
489 * If this case occurs save and restore the interrupt nesting level.
46a3f46d 490 */
27e88a6e
MD
491 if (gd->gd_intr_nesting_level) {
492 int savegdnest;
493 int savegdtrap;
494
495 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
496 panic("lwkt_switch: cannot switch from within "
497 "a fast interrupt, yet, td %p\n", td);
498 } else {
499 savegdnest = gd->gd_intr_nesting_level;
500 savegdtrap = gd->gd_trap_nesting_level;
501 gd->gd_intr_nesting_level = 0;
502 gd->gd_trap_nesting_level = 0;
a7422615
MD
503 if ((td->td_flags & TDF_PANICWARN) == 0) {
504 td->td_flags |= TDF_PANICWARN;
6ea70f76 505 kprintf("Warning: thread switch from interrupt or IPI, "
a7422615 506 "thread %p (%s)\n", td, td->td_comm);
1e5fb84b 507 print_backtrace();
a7422615 508 }
27e88a6e
MD
509 lwkt_switch();
510 gd->gd_intr_nesting_level = savegdnest;
511 gd->gd_trap_nesting_level = savegdtrap;
512 return;
513 }
96728c05 514 }
ef0fdad1 515
cb973d15
MD
516 /*
517 * Passive release (used to transition from user to kernel mode
518 * when we block or switch rather then when we enter the kernel).
519 * This function is NOT called if we are switching into a preemption
520 * or returning from a preemption. Typically this causes us to lose
0a3f9b47
MD
521 * our current process designation (if we have one) and become a true
522 * LWKT thread, and may also hand the current process designation to
523 * another process and schedule thread.
cb973d15
MD
524 */
525 if (td->td_release)
526 td->td_release(td);
527
37af14fe 528 crit_enter_gd(gd);
9d265729
MD
529 if (td->td_toks)
530 lwkt_relalltokens(td);
531
532 /*
b02926de
MD
533 * We had better not be holding any spin locks, but don't get into an
534 * endless panic loop.
9d265729 535 */
bbb31c5d
MD
536 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
537 ("lwkt_switch: still holding a shared spinlock %p!",
538 gd->gd_spinlock_rd));
d666840a
MD
539 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
540 ("lwkt_switch: still holding %d exclusive spinlocks!",
541 gd->gd_spinlocks_wr));
9d265729 542
8a8d5d85
MD
543
544#ifdef SMP
545 /*
546 * td_mpcount cannot be used to determine if we currently hold the
547 * MP lock because get_mplock() will increment it prior to attempting
71ef2f5c
MD
548 * to get the lock, and switch out if it can't. Our ownership of
549 * the actual lock will remain stable while we are in a critical section
550 * (but, of course, another cpu may own or release the lock so the
551 * actual value of mp_lock is not stable).
8a8d5d85
MD
552 */
553 mpheld = MP_LOCK_HELD();
0f7a3396
MD
554#ifdef INVARIANTS
555 if (td->td_cscount) {
6ea70f76 556 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
0f7a3396
MD
557 td);
558 if (panic_on_cscount)
559 panic("switching while mastering cpusync");
560 }
561#endif
8a8d5d85 562#endif
99df837e
MD
563 if ((ntd = td->td_preempted) != NULL) {
564 /*
565 * We had preempted another thread on this cpu, resume the preempted
26a0694b
MD
566 * thread. This occurs transparently, whether the preempted thread
567 * was scheduled or not (it may have been preempted after descheduling
8a8d5d85
MD
568 * itself).
569 *
570 * We have to setup the MP lock for the original thread after backing
571 * out the adjustment that was made to curthread when the original
572 * was preempted.
99df837e 573 */
26a0694b 574 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
8a8d5d85 575#ifdef SMP
96728c05 576 if (ntd->td_mpcount && mpheld == 0) {
fc92d4aa 577 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
96728c05
MD
578 td, ntd, td->td_mpcount, ntd->td_mpcount);
579 }
8a8d5d85
MD
580 if (ntd->td_mpcount) {
581 td->td_mpcount -= ntd->td_mpcount;
582 KKASSERT(td->td_mpcount >= 0);
583 }
584#endif
26a0694b 585 ntd->td_flags |= TDF_PREEMPT_DONE;
8ec60c3f
MD
586
587 /*
b9eb1c19
MD
588 * The interrupt may have woken a thread up, we need to properly
589 * set the reschedule flag if the originally interrupted thread is
590 * at a lower priority.
8ec60c3f
MD
591 */
592 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
593 need_lwkt_resched();
8a8d5d85 594 /* YYY release mp lock on switchback if original doesn't need it */
8ad65e08 595 } else {
4b5f931b
MD
596 /*
597 * Priority queue / round-robin at each priority. Note that user
598 * processes run at a fixed, low priority and the user process
599 * scheduler deals with interactions between user processes
600 * by scheduling and descheduling them from the LWKT queue as
601 * necessary.
8a8d5d85
MD
602 *
603 * We have to adjust the MP lock for the target thread. If we
604 * need the MP lock and cannot obtain it we try to locate a
41a01a4d
MD
605 * thread that does not need the MP lock. If we cannot, we spin
606 * instead of HLT.
607 *
608 * A similar issue exists for the tokens held by the target thread.
609 * If we cannot obtain ownership of the tokens we cannot immediately
610 * schedule the thread.
611 */
612
8ec60c3f
MD
613 /*
614 * If an LWKT reschedule was requested, well that is what we are
615 * doing now so clear it.
616 */
617 clear_lwkt_resched();
4b5f931b
MD
618again:
619 if (gd->gd_runqmask) {
620 int nq = bsrl(gd->gd_runqmask);
621 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
622 gd->gd_runqmask &= ~(1 << nq);
623 goto again;
624 }
8a8d5d85 625#ifdef SMP
41a01a4d 626 /*
df6b8ba0
MD
627 * THREAD SELECTION FOR AN SMP MACHINE BUILD
628 *
41a01a4d
MD
629 * If the target needs the MP lock and we couldn't get it,
630 * or if the target is holding tokens and we could not
631 * gain ownership of the tokens, continue looking for a
632 * thread to schedule and spin instead of HLT if we can't.
a453459d
MD
633 *
634 * NOTE: the mpheld variable invalid after this conditional, it
635 * can change due to both cpu_try_mplock() returning success
9d265729 636 * AND interactions in lwkt_getalltokens() due to the fact that
a453459d
MD
637 * we are trying to check the mpcount of a thread other then
638 * the current thread. Because of this, if the current thread
639 * is not holding td_mpcount, an IPI indirectly run via
9d265729 640 * lwkt_getalltokens() can obtain and release the MP lock and
a453459d 641 * cause the core MP lock to be released.
41a01a4d
MD
642 */
643 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
9d265729 644 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
41a01a4d 645 ) {
8a8d5d85 646 u_int32_t rqmask = gd->gd_runqmask;
a453459d
MD
647
648 mpheld = MP_LOCK_HELD();
649 ntd = NULL;
8a8d5d85
MD
650 while (rqmask) {
651 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
38717797 652 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
a453459d 653 /* spinning due to MP lock being held */
38717797 654#ifdef INVARIANTS
a453459d 655 ++mplock_contention_count;
38717797 656#endif
a453459d 657 /* mplock still not held, 'mpheld' still valid */
41a01a4d 658 continue;
38717797 659 }
a453459d
MD
660
661 /*
9d265729 662 * mpheld state invalid after getalltokens call returns
a453459d
MD
663 * failure, but the variable is only needed for
664 * the loop.
665 */
9d265729 666 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
a453459d 667 /* spinning due to token contention */
38717797 668#ifdef INVARIANTS
a453459d 669 ++token_contention_count;
38717797 670#endif
a453459d 671 mpheld = MP_LOCK_HELD();
41a01a4d 672 continue;
38717797 673 }
41a01a4d 674 break;
8a8d5d85
MD
675 }
676 if (ntd)
677 break;
678 rqmask &= ~(1 << nq);
679 nq = bsrl(rqmask);
b9eb1c19
MD
680
681 /*
682 * We have two choices. We can either refuse to run a
683 * user thread when a kernel thread needs the MP lock
684 * but could not get it, or we can allow it to run but
685 * then expect an IPI (hopefully) later on to force a
686 * reschedule when the MP lock might become available.
687 */
688 if (nq < TDPRI_KERN_LPSCHED) {
689 if (chain_mplock == 0)
690 break;
691 atomic_set_int(&mp_lock_contention_mask,
692 gd->gd_cpumask);
693 /* continue loop, allow user threads to be scheduled */
694 }
8a8d5d85
MD
695 }
696 if (ntd == NULL) {
b402c633 697 cpu_mplock_contested();
a2a5ad0d
MD
698 ntd = &gd->gd_idlethread;
699 ntd->td_flags |= TDF_IDLE_NOHLT;
df6b8ba0 700 goto using_idle_thread;
8a8d5d85 701 } else {
344ad853 702 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
703 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
704 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 }
706 } else {
344ad853 707 ++gd->gd_cnt.v_swtch;
8a8d5d85
MD
708 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
709 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 }
711#else
df6b8ba0
MD
712 /*
713 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
7eb611ef
MD
714 * worry about tokens or the BGL. However, we still have
715 * to call lwkt_getalltokens() in order to properly detect
716 * stale tokens. This call cannot fail for a UP build!
df6b8ba0 717 */
7eb611ef 718 lwkt_getalltokens(ntd);
344ad853 719 ++gd->gd_cnt.v_swtch;
4b5f931b
MD
720 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
721 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
8a8d5d85 722#endif
4b5f931b 723 } else {
3c23a41a 724 /*
60f945af
MD
725 * We have nothing to run but only let the idle loop halt
726 * the cpu if there are no pending interrupts.
3c23a41a 727 */
a2a5ad0d 728 ntd = &gd->gd_idlethread;
60f945af 729 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
3c23a41a 730 ntd->td_flags |= TDF_IDLE_NOHLT;
a453459d 731#ifdef SMP
df6b8ba0
MD
732using_idle_thread:
733 /*
734 * The idle thread should not be holding the MP lock unless we
735 * are trapping in the kernel or in a panic. Since we select the
736 * idle thread unconditionally when no other thread is available,
737 * if the MP lock is desired during a panic or kernel trap, we
738 * have to loop in the scheduler until we get it.
739 */
740 if (ntd->td_mpcount) {
741 mpheld = MP_LOCK_HELD();
b402c633 742 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
df6b8ba0 743 panic("Idle thread %p was holding the BGL!", ntd);
b402c633
MD
744 } else if (mpheld == 0) {
745 cpu_mplock_contested();
df6b8ba0 746 goto again;
b402c633 747 }
df6b8ba0 748 }
a453459d 749#endif
4b5f931b 750 }
f1d1c3fa 751 }
26a0694b
MD
752 KASSERT(ntd->td_pri >= TDPRI_CRIT,
753 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
8a8d5d85
MD
754
755 /*
756 * Do the actual switch. If the new target does not need the MP lock
757 * and we are holding it, release the MP lock. If the new target requires
758 * the MP lock we have already acquired it for the target.
759 */
760#ifdef SMP
761 if (ntd->td_mpcount == 0 ) {
762 if (MP_LOCK_HELD())
763 cpu_rel_mplock();
764 } else {
a453459d 765 ASSERT_MP_LOCK_HELD(ntd);
8a8d5d85
MD
766 }
767#endif
94f6d86e
MD
768 if (td != ntd) {
769 ++switch_count;
85514115 770#ifdef __amd64__
0855a2af 771 KKASSERT(jg_tos_ok(ntd));
85514115 772#endif
f1d1c3fa 773 td->td_switch(ntd);
94f6d86e 774 }
37af14fe
MD
775 /* NOTE: current cpu may have changed after switch */
776 crit_exit_quick(td);
8ad65e08
MD
777}
778
b68b7282 779/*
96728c05
MD
780 * Request that the target thread preempt the current thread. Preemption
781 * only works under a specific set of conditions:
b68b7282 782 *
96728c05
MD
783 * - We are not preempting ourselves
784 * - The target thread is owned by the current cpu
785 * - We are not currently being preempted
786 * - The target is not currently being preempted
d3d1cbc8
MD
787 * - We are not holding any spin locks
788 * - The target thread is not holding any tokens
96728c05
MD
789 * - We are able to satisfy the target's MP lock requirements (if any).
790 *
791 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
792 * this is called via lwkt_schedule() through the td_preemptable callback.
793 * critpri is the managed critical priority that we should ignore in order
794 * to determine whether preemption is possible (aka usually just the crit
795 * priority of lwkt_schedule() itself).
b68b7282 796 *
26a0694b
MD
797 * XXX at the moment we run the target thread in a critical section during
798 * the preemption in order to prevent the target from taking interrupts
799 * that *WE* can't. Preemption is strictly limited to interrupt threads
800 * and interrupt-like threads, outside of a critical section, and the
801 * preempted source thread will be resumed the instant the target blocks
802 * whether or not the source is scheduled (i.e. preemption is supposed to
803 * be as transparent as possible).
4b5f931b 804 *
8a8d5d85
MD
805 * The target thread inherits our MP count (added to its own) for the
806 * duration of the preemption in order to preserve the atomicy of the
96728c05
MD
807 * MP lock during the preemption. Therefore, any preempting targets must be
808 * careful in regards to MP assertions. Note that the MP count may be
71ef2f5c
MD
809 * out of sync with the physical mp_lock, but we do not have to preserve
810 * the original ownership of the lock if it was out of synch (that is, we
811 * can leave it synchronized on return).
b68b7282
MD
812 */
813void
96728c05 814lwkt_preempt(thread_t ntd, int critpri)
b68b7282 815{
46a3f46d 816 struct globaldata *gd = mycpu;
0a3f9b47 817 thread_t td;
8a8d5d85
MD
818#ifdef SMP
819 int mpheld;
57c254db 820 int savecnt;
8a8d5d85 821#endif
b68b7282 822
26a0694b 823 /*
96728c05
MD
824 * The caller has put us in a critical section. We can only preempt
825 * if the caller of the caller was not in a critical section (basically
d666840a 826 * a local interrupt), as determined by the 'critpri' parameter. We
47737962 827 * also can't preempt if the caller is holding any spinlocks (even if
d666840a 828 * he isn't in a critical section). This also handles the tokens test.
96728c05
MD
829 *
830 * YYY The target thread must be in a critical section (else it must
831 * inherit our critical section? I dunno yet).
41a01a4d 832 *
0a3f9b47 833 * Set need_lwkt_resched() unconditionally for now YYY.
26a0694b
MD
834 */
835 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
26a0694b 836
0a3f9b47 837 td = gd->gd_curthread;
0a3f9b47 838 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
57c254db
MD
839 ++preempt_miss;
840 return;
841 }
96728c05
MD
842 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
843 ++preempt_miss;
8ec60c3f 844 need_lwkt_resched();
96728c05
MD
845 return;
846 }
847#ifdef SMP
46a3f46d 848 if (ntd->td_gd != gd) {
96728c05 849 ++preempt_miss;
8ec60c3f 850 need_lwkt_resched();
96728c05
MD
851 return;
852 }
853#endif
41a01a4d 854 /*
d3d1cbc8 855 * Take the easy way out and do not preempt if we are holding
d666840a 856 * any spinlocks. We could test whether the thread(s) being
41a01a4d
MD
857 * preempted interlock against the target thread's tokens and whether
858 * we can get all the target thread's tokens, but this situation
859 * should not occur very often so its easier to simply not preempt.
d666840a
MD
860 * Also, plain spinlocks are impossible to figure out at this point so
861 * just don't preempt.
d3d1cbc8
MD
862 *
863 * Do not try to preempt if the target thread is holding any tokens.
864 * We could try to acquire the tokens but this case is so rare there
865 * is no need to support it.
41a01a4d 866 */
bbb31c5d 867 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
41a01a4d 868 ++preempt_miss;
8ec60c3f 869 need_lwkt_resched();
41a01a4d
MD
870 return;
871 }
d3d1cbc8
MD
872 if (ntd->td_toks) {
873 ++preempt_miss;
874 need_lwkt_resched();
875 return;
876 }
26a0694b
MD
877 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
878 ++preempt_weird;
8ec60c3f 879 need_lwkt_resched();
26a0694b
MD
880 return;
881 }
882 if (ntd->td_preempted) {
4b5f931b 883 ++preempt_hit;
8ec60c3f 884 need_lwkt_resched();
26a0694b 885 return;
b68b7282 886 }
8a8d5d85 887#ifdef SMP
a2a5ad0d
MD
888 /*
889 * note: an interrupt might have occured just as we were transitioning
71ef2f5c
MD
890 * to or from the MP lock. In this case td_mpcount will be pre-disposed
891 * (non-zero) but not actually synchronized with the actual state of the
892 * lock. We can use it to imply an MP lock requirement for the
893 * preemption but we cannot use it to test whether we hold the MP lock
894 * or not.
a2a5ad0d 895 */
96728c05 896 savecnt = td->td_mpcount;
71ef2f5c 897 mpheld = MP_LOCK_HELD();
8a8d5d85
MD
898 ntd->td_mpcount += td->td_mpcount;
899 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
900 ntd->td_mpcount -= td->td_mpcount;
901 ++preempt_miss;
8ec60c3f 902 need_lwkt_resched();
8a8d5d85
MD
903 return;
904 }
905#endif
26a0694b 906
8ec60c3f
MD
907 /*
908 * Since we are able to preempt the current thread, there is no need to
909 * call need_lwkt_resched().
910 */
26a0694b
MD
911 ++preempt_hit;
912 ntd->td_preempted = td;
913 td->td_flags |= TDF_PREEMPT_LOCK;
914 td->td_switch(ntd);
b9eb1c19 915
26a0694b 916 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
96728c05
MD
917#ifdef SMP
918 KKASSERT(savecnt == td->td_mpcount);
71ef2f5c
MD
919 mpheld = MP_LOCK_HELD();
920 if (mpheld && td->td_mpcount == 0)
96728c05 921 cpu_rel_mplock();
71ef2f5c 922 else if (mpheld == 0 && td->td_mpcount)
96728c05
MD
923 panic("lwkt_preempt(): MP lock was not held through");
924#endif
26a0694b
MD
925 ntd->td_preempted = NULL;
926 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
b68b7282
MD
927}
928
f1d1c3fa
MD
929/*
930 * Yield our thread while higher priority threads are pending. This is
931 * typically called when we leave a critical section but it can be safely
932 * called while we are in a critical section.
933 *
934 * This function will not generally yield to equal priority threads but it
935 * can occur as a side effect. Note that lwkt_switch() is called from
46a3f46d 936 * inside the critical section to prevent its own crit_exit() from reentering
f1d1c3fa
MD
937 * lwkt_yield_quick().
938 *
235957ed 939 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
ef0fdad1
MD
940 * came along but was blocked and made pending.
941 *
f1d1c3fa
MD
942 * (self contained on a per cpu basis)
943 */
944void
945lwkt_yield_quick(void)
946{
7966cb69
MD
947 globaldata_t gd = mycpu;
948 thread_t td = gd->gd_curthread;
ef0fdad1 949
a2a5ad0d 950 /*
235957ed 951 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
a2a5ad0d
MD
952 * it with a non-zero cpl then we might not wind up calling splz after
953 * a task switch when the critical section is exited even though the
46a3f46d 954 * new task could accept the interrupt.
a2a5ad0d
MD
955 *
956 * XXX from crit_exit() only called after last crit section is released.
957 * If called directly will run splz() even if in a critical section.
46a3f46d
MD
958 *
959 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
960 * except for this special case, we MUST call splz() here to handle any
961 * pending ints, particularly after we switch, or we might accidently
962 * halt the cpu with interrupts pending.
a2a5ad0d 963 */
46a3f46d 964 if (gd->gd_reqflags && td->td_nest_count < 2)
f1d1c3fa 965 splz();
f1d1c3fa
MD
966
967 /*
968 * YYY enabling will cause wakeup() to task-switch, which really
969 * confused the old 4.x code. This is a good way to simulate
7d0bac62
MD
970 * preemption and MP without actually doing preemption or MP, because a
971 * lot of code assumes that wakeup() does not block.
f1d1c3fa 972 */
46a3f46d
MD
973 if (untimely_switch && td->td_nest_count == 0 &&
974 gd->gd_intr_nesting_level == 0
975 ) {
37af14fe 976 crit_enter_quick(td);
f1d1c3fa
MD
977 /*
978 * YYY temporary hacks until we disassociate the userland scheduler
979 * from the LWKT scheduler.
980 */
981 if (td->td_flags & TDF_RUNQ) {
982 lwkt_switch(); /* will not reenter yield function */
983 } else {
37af14fe 984 lwkt_schedule_self(td); /* make sure we are scheduled */
f1d1c3fa 985 lwkt_switch(); /* will not reenter yield function */
37af14fe 986 lwkt_deschedule_self(td); /* make sure we are descheduled */
f1d1c3fa 987 }
7966cb69 988 crit_exit_noyield(td);
f1d1c3fa 989 }
f1d1c3fa
MD
990}
991
8ad65e08 992/*
f1d1c3fa 993 * This implements a normal yield which, unlike _quick, will yield to equal
235957ed 994 * priority threads as well. Note that gd_reqflags tests will be handled by
f1d1c3fa
MD
995 * the crit_exit() call in lwkt_switch().
996 *
997 * (self contained on a per cpu basis)
8ad65e08
MD
998 */
999void
f1d1c3fa 1000lwkt_yield(void)
8ad65e08 1001{
37af14fe 1002 lwkt_schedule_self(curthread);
f1d1c3fa
MD
1003 lwkt_switch();
1004}
1005
b9eb1c19
MD
1006/*
1007 * Return 0 if no runnable threads are pending at the same or higher
1008 * priority as the passed thread.
1009 *
1010 * Return 1 if runnable threads are pending at the same priority.
1011 *
1012 * Return 2 if runnable threads are pending at a higher priority.
1013 */
1014int
1015lwkt_check_resched(thread_t td)
1016{
1017 int pri = td->td_pri & TDPRI_MASK;
1018
1019 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1020 return(2);
1021 if (TAILQ_NEXT(td, td_threadq))
1022 return(1);
1023 return(0);
1024}
1025
8ad65e08 1026/*
f1d1c3fa
MD
1027 * Generic schedule. Possibly schedule threads belonging to other cpus and
1028 * deal with threads that might be blocked on a wait queue.
1029 *
0a3f9b47
MD
1030 * We have a little helper inline function which does additional work after
1031 * the thread has been enqueued, including dealing with preemption and
1032 * setting need_lwkt_resched() (which prevents the kernel from returning
1033 * to userland until it has processed higher priority threads).
6330a558
MD
1034 *
1035 * It is possible for this routine to be called after a failed _enqueue
1036 * (due to the target thread migrating, sleeping, or otherwise blocked).
1037 * We have to check that the thread is actually on the run queue!
361d01dd
MD
1038 *
1039 * reschedok is an optimized constant propagated from lwkt_schedule() or
1040 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1041 * reschedule to be requested if the target thread has a higher priority.
1042 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1043 * be 0, prevented undesired reschedules.
8ad65e08 1044 */
0a3f9b47
MD
1045static __inline
1046void
361d01dd 1047_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
0a3f9b47 1048{
b9eb1c19 1049 thread_t otd;
c730be20 1050
6330a558 1051 if (ntd->td_flags & TDF_RUNQ) {
361d01dd 1052 if (ntd->td_preemptable && reschedok) {
6330a558 1053 ntd->td_preemptable(ntd, cpri); /* YYY +token */
361d01dd 1054 } else if (reschedok) {
b9eb1c19
MD
1055 otd = curthread;
1056 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
c730be20 1057 need_lwkt_resched();
6330a558 1058 }
0a3f9b47
MD
1059 }
1060}
1061
361d01dd 1062static __inline
8ad65e08 1063void
361d01dd 1064_lwkt_schedule(thread_t td, int reschedok)
8ad65e08 1065{
37af14fe
MD
1066 globaldata_t mygd = mycpu;
1067
41a01a4d 1068 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
37af14fe 1069 crit_enter_gd(mygd);
9388413d 1070 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
37af14fe 1071 if (td == mygd->gd_curthread) {
f1d1c3fa
MD
1072 _lwkt_enqueue(td);
1073 } else {
f1d1c3fa 1074 /*
7cd8d145
MD
1075 * If we own the thread, there is no race (since we are in a
1076 * critical section). If we do not own the thread there might
1077 * be a race but the target cpu will deal with it.
f1d1c3fa 1078 */
0f7a3396 1079#ifdef SMP
7cd8d145 1080 if (td->td_gd == mygd) {
9d265729 1081 _lwkt_enqueue(td);
361d01dd 1082 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
f1d1c3fa 1083 } else {
7cd8d145
MD
1084 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1085 }
0f7a3396 1086#else
7cd8d145 1087 _lwkt_enqueue(td);
361d01dd 1088 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
0f7a3396 1089#endif
8ad65e08 1090 }
37af14fe 1091 crit_exit_gd(mygd);
8ad65e08
MD
1092}
1093
361d01dd
MD
1094void
1095lwkt_schedule(thread_t td)
1096{
1097 _lwkt_schedule(td, 1);
1098}
1099
1100void
1101lwkt_schedule_noresched(thread_t td)
1102{
1103 _lwkt_schedule(td, 0);
1104}
1105
52eedfb5
MD
1106#ifdef SMP
1107
d9eea1a5 1108/*
52eedfb5
MD
1109 * Thread migration using a 'Pull' method. The thread may or may not be
1110 * the current thread. It MUST be descheduled and in a stable state.
1111 * lwkt_giveaway() must be called on the cpu owning the thread.
1112 *
1113 * At any point after lwkt_giveaway() is called, the target cpu may
1114 * 'pull' the thread by calling lwkt_acquire().
1115 *
1116 * MPSAFE - must be called under very specific conditions.
d9eea1a5 1117 */
52eedfb5
MD
1118void
1119lwkt_giveaway(thread_t td)
1120{
1121 globaldata_t gd = mycpu;
1122
1123 crit_enter_gd(gd);
1124 KKASSERT(td->td_gd == gd);
1125 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1126 td->td_flags |= TDF_MIGRATING;
1127 crit_exit_gd(gd);
1128}
1129
a2a5ad0d
MD
1130void
1131lwkt_acquire(thread_t td)
1132{
37af14fe
MD
1133 globaldata_t gd;
1134 globaldata_t mygd;
a2a5ad0d 1135
52eedfb5 1136 KKASSERT(td->td_flags & TDF_MIGRATING);
a2a5ad0d 1137 gd = td->td_gd;
37af14fe 1138 mygd = mycpu;
52eedfb5 1139 if (gd != mycpu) {
35238fa5 1140 cpu_lfence();
52eedfb5 1141 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
37af14fe 1142 crit_enter_gd(mygd);
df910c23
MD
1143 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1144#ifdef SMP
1145 lwkt_process_ipiq();
1146#endif
52eedfb5 1147 cpu_lfence();
df910c23 1148 }
37af14fe 1149 td->td_gd = mygd;
52eedfb5
MD
1150 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1151 td->td_flags &= ~TDF_MIGRATING;
1152 crit_exit_gd(mygd);
1153 } else {
1154 crit_enter_gd(mygd);
1155 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1156 td->td_flags &= ~TDF_MIGRATING;
37af14fe 1157 crit_exit_gd(mygd);
a2a5ad0d
MD
1158 }
1159}
1160
52eedfb5
MD
1161#endif
1162
f1d1c3fa
MD
1163/*
1164 * Generic deschedule. Descheduling threads other then your own should be
1165 * done only in carefully controlled circumstances. Descheduling is
1166 * asynchronous.
1167 *
1168 * This function may block if the cpu has run out of messages.
8ad65e08
MD
1169 */
1170void
1171lwkt_deschedule(thread_t td)
1172{
f1d1c3fa 1173 crit_enter();
b8a98473 1174#ifdef SMP
f1d1c3fa
MD
1175 if (td == curthread) {
1176 _lwkt_dequeue(td);
1177 } else {
a72187e9 1178 if (td->td_gd == mycpu) {
f1d1c3fa
MD
1179 _lwkt_dequeue(td);
1180 } else {
b8a98473 1181 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
f1d1c3fa
MD
1182 }
1183 }
b8a98473
MD
1184#else
1185 _lwkt_dequeue(td);
1186#endif
f1d1c3fa
MD
1187 crit_exit();
1188}
1189
4b5f931b
MD
1190/*
1191 * Set the target thread's priority. This routine does not automatically
1192 * switch to a higher priority thread, LWKT threads are not designed for
1193 * continuous priority changes. Yield if you want to switch.
1194 *
1195 * We have to retain the critical section count which uses the high bits
26a0694b
MD
1196 * of the td_pri field. The specified priority may also indicate zero or
1197 * more critical sections by adding TDPRI_CRIT*N.
18bbe476
MD
1198 *
1199 * Note that we requeue the thread whether it winds up on a different runq
1200 * or not. uio_yield() depends on this and the routine is not normally
1201 * called with the same priority otherwise.
4b5f931b
MD
1202 */
1203void
1204lwkt_setpri(thread_t td, int pri)
1205{
26a0694b 1206 KKASSERT(pri >= 0);
a72187e9 1207 KKASSERT(td->td_gd == mycpu);
26a0694b
MD
1208 crit_enter();
1209 if (td->td_flags & TDF_RUNQ) {
1210 _lwkt_dequeue(td);
1211 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1212 _lwkt_enqueue(td);
1213 } else {
1214 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1215 }
1216 crit_exit();
1217}
1218
1219void
1220lwkt_setpri_self(int pri)
1221{
1222 thread_t td = curthread;
1223
4b5f931b
MD
1224 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1225 crit_enter();
1226 if (td->td_flags & TDF_RUNQ) {
1227 _lwkt_dequeue(td);
1228 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1229 _lwkt_enqueue(td);
1230 } else {
1231 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1232 }
1233 crit_exit();
1234}
1235
5d21b981 1236/*
52eedfb5
MD
1237 * Migrate the current thread to the specified cpu.
1238 *
1239 * This is accomplished by descheduling ourselves from the current cpu,
1240 * moving our thread to the tdallq of the target cpu, IPI messaging the
1241 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1242 * races while the thread is being migrated.
5d21b981 1243 */
3d28ff59 1244#ifdef SMP
5d21b981 1245static void lwkt_setcpu_remote(void *arg);
3d28ff59 1246#endif
5d21b981
MD
1247
1248void
1249lwkt_setcpu_self(globaldata_t rgd)
1250{
1251#ifdef SMP
1252 thread_t td = curthread;
1253
1254 if (td->td_gd != rgd) {
1255 crit_enter_quick(td);
1256 td->td_flags |= TDF_MIGRATING;
1257 lwkt_deschedule_self(td);
52eedfb5 1258 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
b8a98473 1259 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
5d21b981
MD
1260 lwkt_switch();
1261 /* we are now on the target cpu */
52eedfb5 1262 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
5d21b981
MD
1263 crit_exit_quick(td);
1264 }
1265#endif
1266}
1267
ecdefdda
MD
1268void
1269lwkt_migratecpu(int cpuid)
1270{
1271#ifdef SMP
1272 globaldata_t rgd;
1273
1274 rgd = globaldata_find(cpuid);
1275 lwkt_setcpu_self(rgd);
1276#endif
1277}
1278
5d21b981
MD
1279/*
1280 * Remote IPI for cpu migration (called while in a critical section so we
1281 * do not have to enter another one). The thread has already been moved to
1282 * our cpu's allq, but we must wait for the thread to be completely switched
1283 * out on the originating cpu before we schedule it on ours or the stack
1284 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1285 * change to main memory.
1286 *
1287 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1288 * against wakeups. It is best if this interface is used only when there
1289 * are no pending events that might try to schedule the thread.
1290 */
3d28ff59 1291#ifdef SMP
5d21b981
MD
1292static void
1293lwkt_setcpu_remote(void *arg)
1294{
1295 thread_t td = arg;
1296 globaldata_t gd = mycpu;
1297
df910c23
MD
1298 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1299#ifdef SMP
1300 lwkt_process_ipiq();
1301#endif
35238fa5 1302 cpu_lfence();
df910c23 1303 }
5d21b981 1304 td->td_gd = gd;
35238fa5 1305 cpu_sfence();
5d21b981 1306 td->td_flags &= ~TDF_MIGRATING;
9388413d 1307 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
5d21b981
MD
1308 _lwkt_enqueue(td);
1309}
3d28ff59 1310#endif
5d21b981 1311
553ea3c8 1312struct lwp *
4b5f931b
MD
1313lwkt_preempted_proc(void)
1314{
73e4f7b9 1315 thread_t td = curthread;
4b5f931b
MD
1316 while (td->td_preempted)
1317 td = td->td_preempted;
553ea3c8 1318 return(td->td_lwp);
4b5f931b
MD
1319}
1320
99df837e
MD
1321/*
1322 * Create a kernel process/thread/whatever. It shares it's address space
1323 * with proc0 - ie: kernel only.
1324 *
365fa13f
MD
1325 * NOTE! By default new threads are created with the MP lock held. A
1326 * thread which does not require the MP lock should release it by calling
1327 * rel_mplock() at the start of the new thread.
99df837e
MD
1328 */
1329int
1330lwkt_create(void (*func)(void *), void *arg,
75cdbe6c 1331 struct thread **tdp, thread_t template, int tdflags, int cpu,
ef0fdad1 1332 const char *fmt, ...)
99df837e 1333{
73e4f7b9 1334 thread_t td;
e2565a42 1335 __va_list ap;
99df837e 1336
d3d32139 1337 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
dbcd0c9b 1338 tdflags);
a2a5ad0d
MD
1339 if (tdp)
1340 *tdp = td;
709799ea 1341 cpu_set_thread_handler(td, lwkt_exit, func, arg);
99df837e
MD
1342
1343 /*
1344 * Set up arg0 for 'ps' etc
1345 */
e2565a42 1346 __va_start(ap, fmt);
379210cb 1347 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
e2565a42 1348 __va_end(ap);
99df837e
MD
1349
1350 /*
1351 * Schedule the thread to run
1352 */
ef0fdad1
MD
1353 if ((td->td_flags & TDF_STOPREQ) == 0)
1354 lwkt_schedule(td);
1355 else
1356 td->td_flags &= ~TDF_STOPREQ;
99df837e
MD
1357 return 0;
1358}
1359
1360/*
1361 * Destroy an LWKT thread. Warning! This function is not called when
1362 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1363 * uses a different reaping mechanism.
1364 */
1365void
1366lwkt_exit(void)
1367{
1368 thread_t td = curthread;
c070746a 1369 thread_t std;
8826f33a 1370 globaldata_t gd;
99df837e
MD
1371
1372 if (td->td_flags & TDF_VERBOSE)
6ea70f76 1373 kprintf("kthread %p %s has exited\n", td, td->td_comm);
f6bf3af1 1374 caps_exit(td);
c070746a
MD
1375
1376 /*
1377 * Get us into a critical section to interlock gd_freetd and loop
1378 * until we can get it freed.
1379 *
1380 * We have to cache the current td in gd_freetd because objcache_put()ing
1381 * it would rip it out from under us while our thread is still active.
1382 */
1383 gd = mycpu;
37af14fe 1384 crit_enter_quick(td);
c070746a
MD
1385 while ((std = gd->gd_freetd) != NULL) {
1386 gd->gd_freetd = NULL;
1387 objcache_put(thread_cache, std);
1388 }
37af14fe 1389 lwkt_deschedule_self(td);
e56e4dea 1390 lwkt_remove_tdallq(td);
c070746a
MD
1391 if (td->td_flags & TDF_ALLOCATED_THREAD)
1392 gd->gd_freetd = td;
99df837e
MD
1393 cpu_thread_exit();
1394}
1395
e56e4dea
MD
1396void
1397lwkt_remove_tdallq(thread_t td)
1398{
1399 KKASSERT(td->td_gd == mycpu);
1400 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1401}
1402
2d93b37a
MD
1403void
1404crit_panic(void)
1405{
1406 thread_t td = curthread;
1407 int lpri = td->td_pri;
1408
1409 td->td_pri = 0;
1410 panic("td_pri is/would-go negative! %p %d", td, lpri);
1411}
1412
d165e668
MD
1413#ifdef SMP
1414
bd8015ca
MD
1415/*
1416 * Called from debugger/panic on cpus which have been stopped. We must still
1417 * process the IPIQ while stopped, even if we were stopped while in a critical
1418 * section (XXX).
1419 *
1420 * If we are dumping also try to process any pending interrupts. This may
1421 * or may not work depending on the state of the cpu at the point it was
1422 * stopped.
1423 */
1424void
1425lwkt_smp_stopped(void)
1426{
1427 globaldata_t gd = mycpu;
1428
1429 crit_enter_gd(gd);
1430 if (dumping) {
1431 lwkt_process_ipiq();
1432 splz();
1433 } else {
1434 lwkt_process_ipiq();
1435 }
1436 crit_exit_gd(gd);
1437}
1438
57aa743c
MD
1439/*
1440 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1441 * get_mplock() has already incremented td_mpcount. We must block and
1442 * not return until giant is held.
1443 *
1444 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1445 * reschedule the thread until it can obtain the giant lock for it.
1446 */
1447void
1448lwkt_mp_lock_contested(void)
1449{
57aa743c 1450 loggiant(beg);
57aa743c 1451 lwkt_switch();
57aa743c 1452 loggiant(end);
57aa743c
MD
1453}
1454
b9eb1c19
MD
1455/*
1456 * The rel_mplock() code will call this function after releasing the
1457 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1458 *
1459 * We then chain an IPI to a single other cpu potentially needing the
1460 * lock. This is a bit heuristical and we can wind up with IPIs flying
1461 * all over the place.
1462 */
1463static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1464
1465void
1466lwkt_mp_lock_uncontested(void)
1467{
1468 globaldata_t gd;
1469 globaldata_t dgd;
1470 cpumask_t mask;
1471 cpumask_t tmpmask;
1472 int cpuid;
1473
1474 if (chain_mplock) {
1475 gd = mycpu;
1476 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1477 mask = mp_lock_contention_mask;
1478 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1479
1480 if (mask) {
1481 if (mask & tmpmask)
1482 cpuid = bsfl(mask & tmpmask);
1483 else
1484 cpuid = bsfl(mask);
1485 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1486 dgd = globaldata_find(cpuid);
1487 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1488 }
1489 }
1490}
1491
1492/*
1493 * The idea is for this IPI to interrupt a potentially lower priority
1494 * thread, such as a user thread, to allow the scheduler to reschedule
1495 * a higher priority kernel thread that needs the MP lock.
1496 *
1497 * For now we set the LWKT reschedule flag which generates an AST in
1498 * doreti, though theoretically it is also possible to possibly preempt
1499 * here if the underlying thread was operating in user mode. Nah.
1500 */
1501static void
1502lwkt_mp_lock_uncontested_remote(void *arg __unused)
1503{
1504 need_lwkt_resched();
1505}
1506
d165e668 1507#endif