2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
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.
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
34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe Exp $
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.
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
48 #include <sys/rtprio.h>
49 #include <sys/queue.h>
50 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
55 #include <sys/spinlock.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.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>
70 #include <machine/stdarg.h>
71 #include <machine/smp.h>
74 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
76 static int untimely_switch
= 0;
78 static int panic_on_cscount
= 0;
80 static __int64_t switch_count
= 0;
81 static __int64_t preempt_hit
= 0;
82 static __int64_t preempt_miss
= 0;
83 static __int64_t preempt_weird
= 0;
84 static __int64_t token_contention_count
= 0;
85 static __int64_t mplock_contention_count
= 0;
86 static int lwkt_use_spin_port
;
88 static int chain_mplock
= 0;
90 static struct objcache
*thread_cache
;
92 volatile cpumask_t mp_lock_contention_mask
;
94 extern void cpu_heavy_restore(void);
95 extern void cpu_lwkt_restore(void);
96 extern void cpu_kthread_restore(void);
97 extern void cpu_idle_restore(void);
102 jg_tos_ok(struct thread
*td
)
110 KKASSERT(td
->td_sp
!= NULL
);
111 tos
= ((void **)td
->td_sp
)[0];
113 if ((tos
== cpu_heavy_restore
) || (tos
== cpu_lwkt_restore
) ||
114 (tos
== cpu_kthread_restore
) || (tos
== cpu_idle_restore
)) {
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.
126 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
128 SYSCTL_INT(_lwkt
, OID_AUTO
, untimely_switch
, CTLFLAG_RW
, &untimely_switch
, 0, "");
130 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
133 SYSCTL_INT(_lwkt
, OID_AUTO
, chain_mplock
, CTLFLAG_RW
, &chain_mplock
, 0, "");
135 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
136 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
137 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
138 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
140 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
141 &token_contention_count
, 0, "spinning due to token contention");
142 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
143 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
149 #if !defined(KTR_GIANT_CONTENTION)
150 #define KTR_GIANT_CONTENTION KTR_ALL
153 KTR_INFO_MASTER(giant
);
154 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
155 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
157 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
160 * These helper procedures handle the runq, they can only be called from
161 * within a critical section.
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.
170 _lwkt_dequeue(thread_t td
)
172 if (td
->td_flags
& TDF_RUNQ
) {
173 int nq
= td
->td_pri
& TDPRI_MASK
;
174 struct globaldata
*gd
= td
->td_gd
;
176 td
->td_flags
&= ~TDF_RUNQ
;
177 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
178 /* runqmask is passively cleaned up by the switcher */
184 _lwkt_enqueue(thread_t td
)
186 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_TSLEEPQ
|TDF_BLOCKQ
)) == 0) {
187 int nq
= td
->td_pri
& TDPRI_MASK
;
188 struct globaldata
*gd
= td
->td_gd
;
190 td
->td_flags
|= TDF_RUNQ
;
191 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
192 gd
->gd_runqmask
|= 1 << nq
;
197 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
199 struct thread
*td
= (struct thread
*)obj
;
201 td
->td_kstack
= NULL
;
202 td
->td_kstack_size
= 0;
203 td
->td_flags
= TDF_ALLOCATED_THREAD
;
208 _lwkt_thread_dtor(void *obj
, void *privdata
)
210 struct thread
*td
= (struct thread
*)obj
;
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
);
221 * Initialize the lwkt s/system.
226 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
227 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
),
228 NULL
, CACHE_NTHREADS
/2,
229 _lwkt_thread_ctor
, _lwkt_thread_dtor
, NULL
);
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
237 * (non-blocking, self contained on a per cpu basis)
240 lwkt_schedule_self(thread_t td
)
242 crit_enter_quick(td
);
243 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
244 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
250 * Deschedule a thread.
252 * (non-blocking, self contained on a per cpu basis)
255 lwkt_deschedule_self(thread_t td
)
257 crit_enter_quick(td
);
263 * LWKTs operate on a per-cpu basis
265 * WARNING! Called from early boot, 'mycpu' may not work yet.
268 lwkt_gdinit(struct globaldata
*gd
)
272 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
273 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
275 TAILQ_INIT(&gd
->gd_tdallq
);
279 * Create a new thread. The thread must be associated with a process context
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.
283 * If you intend to create a thread without a process context this function
284 * does everything except load the startup and switcher function.
287 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
289 globaldata_t gd
= mycpu
;
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.
298 if ((td
= gd
->gd_freetd
) != NULL
)
299 gd
->gd_freetd
= NULL
;
301 td
= objcache_get(thread_cache
, M_WAITOK
);
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
);
309 * Try to reuse cached stack.
311 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
312 if (flags
& TDF_ALLOCATED_STACK
) {
313 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
318 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
319 flags
|= TDF_ALLOCATED_STACK
;
322 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
324 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
329 * Initialize a preexisting thread structure. This function is used by
330 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
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
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
339 * NOTE! we have to be careful in regards to creating threads for other cpus
340 * if SMP has not yet been activated.
345 lwkt_init_thread_remote(void *arg
)
350 * Protected by critical section held by IPI dispatch
352 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
358 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
359 struct globaldata
*gd
)
361 globaldata_t mygd
= mycpu
;
363 bzero(td
, sizeof(struct thread
));
364 td
->td_kstack
= stack
;
365 td
->td_kstack_size
= stksize
;
366 td
->td_flags
= flags
;
368 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
370 if ((flags
& TDF_MPSAFE
) == 0)
373 if (lwkt_use_spin_port
)
374 lwkt_initport_spin(&td
->td_msgport
);
376 lwkt_initport_thread(&td
->td_msgport
, td
);
377 pmap_init_thread(td
);
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.
385 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
387 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
390 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
394 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
400 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
405 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
410 lwkt_hold(thread_t td
)
416 lwkt_rele(thread_t td
)
418 KKASSERT(td
->td_refs
> 0);
423 lwkt_wait_free(thread_t td
)
426 tsleep(td
, 0, "tdreap", hz
);
430 lwkt_free_thread(thread_t td
)
432 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
433 ("lwkt_free_thread: did not exit! %p", td
));
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;
449 * Switch to the next runnable lwkt. If no LWKTs are runnable then
450 * switch to the idlethread. Switching must occur within a critical
451 * section to avoid races with the scheduling queue.
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.
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.
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
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.
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).
477 globaldata_t gd
= mycpu
;
478 thread_t td
= gd
->gd_curthread
;
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.
489 * If this case occurs save and restore the interrupt nesting level.
491 if (gd
->gd_intr_nesting_level
) {
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
);
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;
503 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
504 td
->td_flags
|= TDF_PANICWARN
;
505 kprintf("Warning: thread switch from interrupt or IPI, "
506 "thread %p (%s)\n", td
, td
->td_comm
);
510 gd
->gd_intr_nesting_level
= savegdnest
;
511 gd
->gd_trap_nesting_level
= savegdtrap
;
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
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.
530 lwkt_relalltokens(td
);
533 * We had better not be holding any spin locks, but don't get into an
534 * endless panic loop.
536 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
537 ("lwkt_switch: still holding a shared spinlock %p!",
538 gd
->gd_spinlock_rd
));
539 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
540 ("lwkt_switch: still holding %d exclusive spinlocks!",
541 gd
->gd_spinlocks_wr
));
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
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).
553 mpheld
= MP_LOCK_HELD();
555 if (td
->td_cscount
) {
556 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
558 if (panic_on_cscount
)
559 panic("switching while mastering cpusync");
563 if ((ntd
= td
->td_preempted
) != NULL
) {
565 * We had preempted another thread on this cpu, resume the preempted
566 * thread. This occurs transparently, whether the preempted thread
567 * was scheduled or not (it may have been preempted after descheduling
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
574 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
576 if (ntd
->td_mpcount
&& mpheld
== 0) {
577 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
578 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
580 if (ntd
->td_mpcount
) {
581 td
->td_mpcount
-= ntd
->td_mpcount
;
582 KKASSERT(td
->td_mpcount
>= 0);
585 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
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.
592 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
594 /* YYY release mp lock on switchback if original doesn't need it */
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
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
605 * thread that does not need the MP lock. If we cannot, we spin
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.
614 * If an LWKT reschedule was requested, well that is what we are
615 * doing now so clear it.
617 clear_lwkt_resched();
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
);
627 * THREAD SELECTION FOR AN SMP MACHINE BUILD
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.
634 * NOTE: the mpheld variable invalid after this conditional, it
635 * can change due to both cpu_try_mplock() returning success
636 * AND interactions in lwkt_getalltokens() due to the fact that
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
640 * lwkt_getalltokens() can obtain and release the MP lock and
641 * cause the core MP lock to be released.
643 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
644 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
646 u_int32_t rqmask
= gd
->gd_runqmask
;
648 mpheld
= MP_LOCK_HELD();
651 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
652 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
653 /* spinning due to MP lock being held */
655 ++mplock_contention_count
;
657 /* mplock still not held, 'mpheld' still valid */
662 * mpheld state invalid after getalltokens call returns
663 * failure, but the variable is only needed for
666 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
667 /* spinning due to token contention */
669 ++token_contention_count
;
671 mpheld
= MP_LOCK_HELD();
678 rqmask
&= ~(1 << nq
);
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.
688 if (nq
< TDPRI_KERN_LPSCHED
) {
689 if (chain_mplock
== 0)
691 atomic_set_int(&mp_lock_contention_mask
,
693 /* continue loop, allow user threads to be scheduled */
697 cpu_mplock_contested();
698 ntd
= &gd
->gd_idlethread
;
699 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
700 goto using_idle_thread
;
702 ++gd
->gd_cnt
.v_swtch
;
703 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
704 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
707 ++gd
->gd_cnt
.v_swtch
;
708 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
709 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
713 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
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!
718 lwkt_getalltokens(ntd
);
719 ++gd
->gd_cnt
.v_swtch
;
720 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
721 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
725 * We have nothing to run but only let the idle loop halt
726 * the cpu if there are no pending interrupts.
728 ntd
= &gd
->gd_idlethread
;
729 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
730 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
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.
740 if (ntd
->td_mpcount
) {
741 mpheld
= MP_LOCK_HELD();
742 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
743 panic("Idle thread %p was holding the BGL!", ntd
);
744 } else if (mpheld
== 0) {
745 cpu_mplock_contested();
752 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
753 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
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.
761 if (ntd
->td_mpcount
== 0 ) {
765 ASSERT_MP_LOCK_HELD(ntd
);
771 KKASSERT(jg_tos_ok(ntd
));
775 /* NOTE: current cpu may have changed after switch */
780 * Request that the target thread preempt the current thread. Preemption
781 * only works under a specific set of conditions:
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
787 * - We are not holding any spin locks
788 * - The target thread is not holding any tokens
789 * - We are able to satisfy the target's MP lock requirements (if any).
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).
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).
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
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
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).
814 lwkt_preempt(thread_t ntd
, int critpri
)
816 struct globaldata
*gd
= mycpu
;
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
826 * a local interrupt), as determined by the 'critpri' parameter. We
827 * also can't preempt if the caller is holding any spinlocks (even if
828 * he isn't in a critical section). This also handles the tokens test.
830 * YYY The target thread must be in a critical section (else it must
831 * inherit our critical section? I dunno yet).
833 * Set need_lwkt_resched() unconditionally for now YYY.
835 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
837 td
= gd
->gd_curthread
;
838 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
842 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
848 if (ntd
->td_gd
!= gd
) {
855 * Take the easy way out and do not preempt if we are holding
856 * any spinlocks. We could test whether the thread(s) being
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.
860 * Also, plain spinlocks are impossible to figure out at this point so
861 * just don't preempt.
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.
867 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
877 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
882 if (ntd
->td_preempted
) {
889 * note: an interrupt might have occured just as we were transitioning
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
896 savecnt
= td
->td_mpcount
;
897 mpheld
= MP_LOCK_HELD();
898 ntd
->td_mpcount
+= td
->td_mpcount
;
899 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
900 ntd
->td_mpcount
-= td
->td_mpcount
;
908 * Since we are able to preempt the current thread, there is no need to
909 * call need_lwkt_resched().
912 ntd
->td_preempted
= td
;
913 td
->td_flags
|= TDF_PREEMPT_LOCK
;
916 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
918 KKASSERT(savecnt
== td
->td_mpcount
);
919 mpheld
= MP_LOCK_HELD();
920 if (mpheld
&& td
->td_mpcount
== 0)
922 else if (mpheld
== 0 && td
->td_mpcount
)
923 panic("lwkt_preempt(): MP lock was not held through");
925 ntd
->td_preempted
= NULL
;
926 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
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.
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
936 * inside the critical section to prevent its own crit_exit() from reentering
937 * lwkt_yield_quick().
939 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
940 * came along but was blocked and made pending.
942 * (self contained on a per cpu basis)
945 lwkt_yield_quick(void)
947 globaldata_t gd
= mycpu
;
948 thread_t td
= gd
->gd_curthread
;
951 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
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
954 * new task could accept the interrupt.
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.
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.
964 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
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
970 * preemption and MP without actually doing preemption or MP, because a
971 * lot of code assumes that wakeup() does not block.
973 if (untimely_switch
&& td
->td_nest_count
== 0 &&
974 gd
->gd_intr_nesting_level
== 0
976 crit_enter_quick(td
);
978 * YYY temporary hacks until we disassociate the userland scheduler
979 * from the LWKT scheduler.
981 if (td
->td_flags
& TDF_RUNQ
) {
982 lwkt_switch(); /* will not reenter yield function */
984 lwkt_schedule_self(td
); /* make sure we are scheduled */
985 lwkt_switch(); /* will not reenter yield function */
986 lwkt_deschedule_self(td
); /* make sure we are descheduled */
988 crit_exit_noyield(td
);
993 * This implements a normal yield which, unlike _quick, will yield to equal
994 * priority threads as well. Note that gd_reqflags tests will be handled by
995 * the crit_exit() call in lwkt_switch().
997 * (self contained on a per cpu basis)
1002 lwkt_schedule_self(curthread
);
1007 * Return 0 if no runnable threads are pending at the same or higher
1008 * priority as the passed thread.
1010 * Return 1 if runnable threads are pending at the same priority.
1012 * Return 2 if runnable threads are pending at a higher priority.
1015 lwkt_check_resched(thread_t td
)
1017 int pri
= td
->td_pri
& TDPRI_MASK
;
1019 if (td
->td_gd
->gd_runqmask
> (2 << pri
) - 1)
1021 if (TAILQ_NEXT(td
, td_threadq
))
1027 * Generic schedule. Possibly schedule threads belonging to other cpus and
1028 * deal with threads that might be blocked on a wait queue.
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).
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!
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.
1047 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
, int reschedok
)
1051 if (ntd
->td_flags
& TDF_RUNQ
) {
1052 if (ntd
->td_preemptable
&& reschedok
) {
1053 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
1054 } else if (reschedok
) {
1056 if ((ntd
->td_pri
& TDPRI_MASK
) > (otd
->td_pri
& TDPRI_MASK
))
1057 need_lwkt_resched();
1064 _lwkt_schedule(thread_t td
, int reschedok
)
1066 globaldata_t mygd
= mycpu
;
1068 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1069 crit_enter_gd(mygd
);
1070 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1071 if (td
== mygd
->gd_curthread
) {
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.
1080 if (td
->td_gd
== mygd
) {
1082 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1084 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_schedule
, td
);
1088 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1095 lwkt_schedule(thread_t td
)
1097 _lwkt_schedule(td
, 1);
1101 lwkt_schedule_noresched(thread_t td
)
1103 _lwkt_schedule(td
, 0);
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.
1113 * At any point after lwkt_giveaway() is called, the target cpu may
1114 * 'pull' the thread by calling lwkt_acquire().
1116 * MPSAFE - must be called under very specific conditions.
1119 lwkt_giveaway(thread_t td
)
1121 globaldata_t gd
= mycpu
;
1124 KKASSERT(td
->td_gd
== gd
);
1125 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1126 td
->td_flags
|= TDF_MIGRATING
;
1131 lwkt_acquire(thread_t td
)
1136 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1141 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1142 crit_enter_gd(mygd
);
1143 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1145 lwkt_process_ipiq();
1150 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1151 td
->td_flags
&= ~TDF_MIGRATING
;
1154 crit_enter_gd(mygd
);
1155 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1156 td
->td_flags
&= ~TDF_MIGRATING
;
1164 * Generic deschedule. Descheduling threads other then your own should be
1165 * done only in carefully controlled circumstances. Descheduling is
1168 * This function may block if the cpu has run out of messages.
1171 lwkt_deschedule(thread_t td
)
1175 if (td
== curthread
) {
1178 if (td
->td_gd
== mycpu
) {
1181 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
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.
1195 * We have to retain the critical section count which uses the high bits
1196 * of the td_pri field. The specified priority may also indicate zero or
1197 * more critical sections by adding TDPRI_CRIT*N.
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.
1204 lwkt_setpri(thread_t td
, int pri
)
1207 KKASSERT(td
->td_gd
== mycpu
);
1209 if (td
->td_flags
& TDF_RUNQ
) {
1211 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1214 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1220 lwkt_setpri_self(int pri
)
1222 thread_t td
= curthread
;
1224 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1226 if (td
->td_flags
& TDF_RUNQ
) {
1228 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1231 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1237 * Migrate the current thread to the specified cpu.
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.
1245 static void lwkt_setcpu_remote(void *arg
);
1249 lwkt_setcpu_self(globaldata_t rgd
)
1252 thread_t td
= curthread
;
1254 if (td
->td_gd
!= rgd
) {
1255 crit_enter_quick(td
);
1256 td
->td_flags
|= TDF_MIGRATING
;
1257 lwkt_deschedule_self(td
);
1258 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1259 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1261 /* we are now on the target cpu */
1262 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1263 crit_exit_quick(td
);
1269 lwkt_migratecpu(int cpuid
)
1274 rgd
= globaldata_find(cpuid
);
1275 lwkt_setcpu_self(rgd
);
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.
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.
1293 lwkt_setcpu_remote(void *arg
)
1296 globaldata_t gd
= mycpu
;
1298 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1300 lwkt_process_ipiq();
1306 td
->td_flags
&= ~TDF_MIGRATING
;
1307 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1313 lwkt_preempted_proc(void)
1315 thread_t td
= curthread
;
1316 while (td
->td_preempted
)
1317 td
= td
->td_preempted
;
1322 * Create a kernel process/thread/whatever. It shares it's address space
1323 * with proc0 - ie: kernel only.
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.
1330 lwkt_create(void (*func
)(void *), void *arg
,
1331 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1332 const char *fmt
, ...)
1337 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1341 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1344 * Set up arg0 for 'ps' etc
1346 __va_start(ap
, fmt
);
1347 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1351 * Schedule the thread to run
1353 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1356 td
->td_flags
&= ~TDF_STOPREQ
;
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.
1368 thread_t td
= curthread
;
1372 if (td
->td_flags
& TDF_VERBOSE
)
1373 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1377 * Get us into a critical section to interlock gd_freetd and loop
1378 * until we can get it freed.
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.
1384 crit_enter_quick(td
);
1385 while ((std
= gd
->gd_freetd
) != NULL
) {
1386 gd
->gd_freetd
= NULL
;
1387 objcache_put(thread_cache
, std
);
1389 lwkt_deschedule_self(td
);
1390 lwkt_remove_tdallq(td
);
1391 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1397 lwkt_remove_tdallq(thread_t td
)
1399 KKASSERT(td
->td_gd
== mycpu
);
1400 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1406 thread_t td
= curthread
;
1407 int lpri
= td
->td_pri
;
1410 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
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
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
1425 lwkt_smp_stopped(void)
1427 globaldata_t gd
= mycpu
;
1431 lwkt_process_ipiq();
1434 lwkt_process_ipiq();
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.
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.
1448 lwkt_mp_lock_contested(void)
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.
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.
1463 static void lwkt_mp_lock_uncontested_remote(void *arg __unused
);
1466 lwkt_mp_lock_uncontested(void)
1476 atomic_clear_int(&mp_lock_contention_mask
, gd
->gd_cpumask
);
1477 mask
= mp_lock_contention_mask
;
1478 tmpmask
= ~((1 << gd
->gd_cpuid
) - 1);
1482 cpuid
= bsfl(mask
& tmpmask
);
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
);
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.
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.
1502 lwkt_mp_lock_uncontested_remote(void *arg __unused
)
1504 need_lwkt_resched();