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
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
58 #include <sys/mplock2.h>
61 #include <vm/vm_param.h>
62 #include <vm/vm_kern.h>
63 #include <vm/vm_object.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_pager.h>
67 #include <vm/vm_extern.h>
69 #include <machine/stdarg.h>
70 #include <machine/smp.h>
72 #if !defined(KTR_CTXSW)
73 #define KTR_CTXSW KTR_ALL
75 KTR_INFO_MASTER(ctxsw
);
76 KTR_INFO(KTR_CTXSW
, ctxsw
, sw
, 0, "sw %p > %p", 2 * sizeof(struct thread
*));
77 KTR_INFO(KTR_CTXSW
, ctxsw
, pre
, 1, "pre %p > %p", 2 * sizeof(struct thread
*));
78 KTR_INFO(KTR_CTXSW
, ctxsw
, newtd
, 2, "new_td %p %s", sizeof (struct thread
*) +
80 KTR_INFO(KTR_CTXSW
, ctxsw
, deadtd
, 3, "dead_td %p", sizeof (struct thread
*));
82 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
85 static int panic_on_cscount
= 0;
87 static __int64_t switch_count
= 0;
88 static __int64_t preempt_hit
= 0;
89 static __int64_t preempt_miss
= 0;
90 static __int64_t preempt_weird
= 0;
91 static __int64_t token_contention_count __debugvar
= 0;
92 static int lwkt_use_spin_port
;
93 static struct objcache
*thread_cache
;
96 static void lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
);
99 extern void cpu_heavy_restore(void);
100 extern void cpu_lwkt_restore(void);
101 extern void cpu_kthread_restore(void);
102 extern void cpu_idle_restore(void);
107 jg_tos_ok(struct thread
*td
)
115 KKASSERT(td
->td_sp
!= NULL
);
116 tos
= ((void **)td
->td_sp
)[0];
118 if ((tos
== cpu_heavy_restore
) || (tos
== cpu_lwkt_restore
) ||
119 (tos
== cpu_kthread_restore
) || (tos
== cpu_idle_restore
)) {
128 * We can make all thread ports use the spin backend instead of the thread
129 * backend. This should only be set to debug the spin backend.
131 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
134 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
136 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
137 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
138 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
139 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
141 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
142 &token_contention_count
, 0, "spinning due to token contention");
146 * These helper procedures handle the runq, they can only be called from
147 * within a critical section.
149 * WARNING! Prior to SMP being brought up it is possible to enqueue and
150 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
151 * instead of 'mycpu' when referencing the globaldata structure. Once
152 * SMP live enqueuing and dequeueing only occurs on the current cpu.
156 _lwkt_dequeue(thread_t td
)
158 if (td
->td_flags
& TDF_RUNQ
) {
159 int nq
= td
->td_pri
& TDPRI_MASK
;
160 struct globaldata
*gd
= td
->td_gd
;
162 td
->td_flags
&= ~TDF_RUNQ
;
163 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
164 /* runqmask is passively cleaned up by the switcher */
170 _lwkt_enqueue(thread_t td
)
172 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_BLOCKQ
)) == 0) {
173 int nq
= td
->td_pri
& TDPRI_MASK
;
174 struct globaldata
*gd
= td
->td_gd
;
176 td
->td_flags
|= TDF_RUNQ
;
177 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
178 gd
->gd_runqmask
|= 1 << nq
;
183 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
185 struct thread
*td
= (struct thread
*)obj
;
187 td
->td_kstack
= NULL
;
188 td
->td_kstack_size
= 0;
189 td
->td_flags
= TDF_ALLOCATED_THREAD
;
194 _lwkt_thread_dtor(void *obj
, void *privdata
)
196 struct thread
*td
= (struct thread
*)obj
;
198 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
199 ("_lwkt_thread_dtor: not allocated from objcache"));
200 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
201 td
->td_kstack_size
> 0,
202 ("_lwkt_thread_dtor: corrupted stack"));
203 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
207 * Initialize the lwkt s/system.
212 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
213 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
),
214 NULL
, CACHE_NTHREADS
/2,
215 _lwkt_thread_ctor
, _lwkt_thread_dtor
, NULL
);
219 * Schedule a thread to run. As the current thread we can always safely
220 * schedule ourselves, and a shortcut procedure is provided for that
223 * (non-blocking, self contained on a per cpu basis)
226 lwkt_schedule_self(thread_t td
)
228 crit_enter_quick(td
);
229 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
230 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
236 * Deschedule a thread.
238 * (non-blocking, self contained on a per cpu basis)
241 lwkt_deschedule_self(thread_t td
)
243 crit_enter_quick(td
);
249 * LWKTs operate on a per-cpu basis
251 * WARNING! Called from early boot, 'mycpu' may not work yet.
254 lwkt_gdinit(struct globaldata
*gd
)
258 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
259 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
261 TAILQ_INIT(&gd
->gd_tdallq
);
265 * Create a new thread. The thread must be associated with a process context
266 * or LWKT start address before it can be scheduled. If the target cpu is
267 * -1 the thread will be created on the current cpu.
269 * If you intend to create a thread without a process context this function
270 * does everything except load the startup and switcher function.
273 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
275 globaldata_t gd
= mycpu
;
279 * If static thread storage is not supplied allocate a thread. Reuse
280 * a cached free thread if possible. gd_freetd is used to keep an exiting
281 * thread intact through the exit.
284 if ((td
= gd
->gd_freetd
) != NULL
)
285 gd
->gd_freetd
= NULL
;
287 td
= objcache_get(thread_cache
, M_WAITOK
);
288 KASSERT((td
->td_flags
&
289 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
290 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
291 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
295 * Try to reuse cached stack.
297 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
298 if (flags
& TDF_ALLOCATED_STACK
) {
299 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
304 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
305 flags
|= TDF_ALLOCATED_STACK
;
308 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
310 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
315 * Initialize a preexisting thread structure. This function is used by
316 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
318 * All threads start out in a critical section at a priority of
319 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
320 * appropriate. This function may send an IPI message when the
321 * requested cpu is not the current cpu and consequently gd_tdallq may
322 * not be initialized synchronously from the point of view of the originating
325 * NOTE! we have to be careful in regards to creating threads for other cpus
326 * if SMP has not yet been activated.
331 lwkt_init_thread_remote(void *arg
)
336 * Protected by critical section held by IPI dispatch
338 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
344 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
345 struct globaldata
*gd
)
347 globaldata_t mygd
= mycpu
;
349 bzero(td
, sizeof(struct thread
));
350 td
->td_kstack
= stack
;
351 td
->td_kstack_size
= stksize
;
352 td
->td_flags
= flags
;
354 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
356 if ((flags
& TDF_MPSAFE
) == 0)
359 if (lwkt_use_spin_port
)
360 lwkt_initport_spin(&td
->td_msgport
);
362 lwkt_initport_thread(&td
->td_msgport
, td
);
363 pmap_init_thread(td
);
366 * Normally initializing a thread for a remote cpu requires sending an
367 * IPI. However, the idlethread is setup before the other cpus are
368 * activated so we have to treat it as a special case. XXX manipulation
369 * of gd_tdallq requires the BGL.
371 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
373 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
376 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
380 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
386 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
391 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
393 KTR_LOG(ctxsw_newtd
, td
, &td
->td_comm
[0]);
397 lwkt_hold(thread_t td
)
403 lwkt_rele(thread_t td
)
405 KKASSERT(td
->td_refs
> 0);
410 lwkt_wait_free(thread_t td
)
413 tsleep(td
, 0, "tdreap", hz
);
417 lwkt_free_thread(thread_t td
)
419 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
420 ("lwkt_free_thread: did not exit! %p", td
));
422 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
423 objcache_put(thread_cache
, td
);
424 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
425 /* client-allocated struct with internally allocated stack */
426 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
427 ("lwkt_free_thread: corrupted stack"));
428 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
429 td
->td_kstack
= NULL
;
430 td
->td_kstack_size
= 0;
432 KTR_LOG(ctxsw_deadtd
, td
);
437 * Switch to the next runnable lwkt. If no LWKTs are runnable then
438 * switch to the idlethread. Switching must occur within a critical
439 * section to avoid races with the scheduling queue.
441 * We always have full control over our cpu's run queue. Other cpus
442 * that wish to manipulate our queue must use the cpu_*msg() calls to
443 * talk to our cpu, so a critical section is all that is needed and
444 * the result is very, very fast thread switching.
446 * The LWKT scheduler uses a fixed priority model and round-robins at
447 * each priority level. User process scheduling is a totally
448 * different beast and LWKT priorities should not be confused with
449 * user process priorities.
451 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
452 * cleans it up. Note that the td_switch() function cannot do anything that
453 * requires the MP lock since the MP lock will have already been setup for
454 * the target thread (not the current thread). It's nice to have a scheduler
455 * that does not need the MP lock to work because it allows us to do some
456 * really cool high-performance MP lock optimizations.
458 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
459 * is not called by the current thread in the preemption case, only when
460 * the preempting thread blocks (in order to return to the original thread).
465 globaldata_t gd
= mycpu
;
466 thread_t td
= gd
->gd_curthread
;
473 * Switching from within a 'fast' (non thread switched) interrupt or IPI
474 * is illegal. However, we may have to do it anyway if we hit a fatal
475 * kernel trap or we have paniced.
477 * If this case occurs save and restore the interrupt nesting level.
479 if (gd
->gd_intr_nesting_level
) {
483 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
484 panic("lwkt_switch: cannot switch from within "
485 "a fast interrupt, yet, td %p\n", td
);
487 savegdnest
= gd
->gd_intr_nesting_level
;
488 savegdtrap
= gd
->gd_trap_nesting_level
;
489 gd
->gd_intr_nesting_level
= 0;
490 gd
->gd_trap_nesting_level
= 0;
491 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
492 td
->td_flags
|= TDF_PANICWARN
;
493 kprintf("Warning: thread switch from interrupt or IPI, "
494 "thread %p (%s)\n", td
, td
->td_comm
);
498 gd
->gd_intr_nesting_level
= savegdnest
;
499 gd
->gd_trap_nesting_level
= savegdtrap
;
505 * Passive release (used to transition from user to kernel mode
506 * when we block or switch rather then when we enter the kernel).
507 * This function is NOT called if we are switching into a preemption
508 * or returning from a preemption. Typically this causes us to lose
509 * our current process designation (if we have one) and become a true
510 * LWKT thread, and may also hand the current process designation to
511 * another process and schedule thread.
518 lwkt_relalltokens(td
);
521 * We had better not be holding any spin locks, but don't get into an
522 * endless panic loop.
524 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
525 ("lwkt_switch: still holding a shared spinlock %p!",
526 gd
->gd_spinlock_rd
));
527 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
528 ("lwkt_switch: still holding %d exclusive spinlocks!",
529 gd
->gd_spinlocks_wr
));
534 * td_mpcount cannot be used to determine if we currently hold the
535 * MP lock because get_mplock() will increment it prior to attempting
536 * to get the lock, and switch out if it can't. Our ownership of
537 * the actual lock will remain stable while we are in a critical section
538 * (but, of course, another cpu may own or release the lock so the
539 * actual value of mp_lock is not stable).
541 mpheld
= MP_LOCK_HELD();
543 if (td
->td_cscount
) {
544 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
546 if (panic_on_cscount
)
547 panic("switching while mastering cpusync");
551 if ((ntd
= td
->td_preempted
) != NULL
) {
553 * We had preempted another thread on this cpu, resume the preempted
554 * thread. This occurs transparently, whether the preempted thread
555 * was scheduled or not (it may have been preempted after descheduling
558 * We have to setup the MP lock for the original thread after backing
559 * out the adjustment that was made to curthread when the original
562 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
564 if (ntd
->td_mpcount
&& mpheld
== 0) {
565 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
566 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
568 if (ntd
->td_mpcount
) {
569 td
->td_mpcount
-= ntd
->td_mpcount
;
570 KKASSERT(td
->td_mpcount
>= 0);
573 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
576 * The interrupt may have woken a thread up, we need to properly
577 * set the reschedule flag if the originally interrupted thread is
578 * at a lower priority.
580 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
582 /* YYY release mp lock on switchback if original doesn't need it */
585 * Priority queue / round-robin at each priority. Note that user
586 * processes run at a fixed, low priority and the user process
587 * scheduler deals with interactions between user processes
588 * by scheduling and descheduling them from the LWKT queue as
591 * We have to adjust the MP lock for the target thread. If we
592 * need the MP lock and cannot obtain it we try to locate a
593 * thread that does not need the MP lock. If we cannot, we spin
596 * A similar issue exists for the tokens held by the target thread.
597 * If we cannot obtain ownership of the tokens we cannot immediately
598 * schedule the thread.
602 * If an LWKT reschedule was requested, well that is what we are
603 * doing now so clear it.
605 clear_lwkt_resched();
607 if (gd
->gd_runqmask
) {
608 int nq
= bsrl(gd
->gd_runqmask
);
609 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
610 gd
->gd_runqmask
&= ~(1 << nq
);
615 * THREAD SELECTION FOR AN SMP MACHINE BUILD
617 * If the target needs the MP lock and we couldn't get it,
618 * or if the target is holding tokens and we could not
619 * gain ownership of the tokens, continue looking for a
620 * thread to schedule and spin instead of HLT if we can't.
622 * NOTE: the mpheld variable invalid after this conditional, it
623 * can change due to both cpu_try_mplock() returning success
624 * AND interactions in lwkt_getalltokens() due to the fact that
625 * we are trying to check the mpcount of a thread other then
626 * the current thread. Because of this, if the current thread
627 * is not holding td_mpcount, an IPI indirectly run via
628 * lwkt_getalltokens() can obtain and release the MP lock and
629 * cause the core MP lock to be released.
631 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
632 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
634 u_int32_t rqmask
= gd
->gd_runqmask
;
636 mpheld
= MP_LOCK_HELD();
639 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
640 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
641 /* spinning due to MP lock being held */
646 * mpheld state invalid after getalltokens call returns
647 * failure, but the variable is only needed for
650 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
651 /* spinning due to token contention */
653 ++token_contention_count
;
655 mpheld
= MP_LOCK_HELD();
662 rqmask
&= ~(1 << nq
);
666 * We have two choices. We can either refuse to run a
667 * user thread when a kernel thread needs the MP lock
668 * but could not get it, or we can allow it to run but
669 * then expect an IPI (hopefully) later on to force a
670 * reschedule when the MP lock might become available.
672 if (nq
< TDPRI_KERN_LPSCHED
) {
673 break; /* for now refuse to run */
675 if (chain_mplock
== 0)
677 /* continue loop, allow user threads to be scheduled */
683 * Case where a (kernel) thread needed the MP lock and could
684 * not get one, and we may or may not have found another
685 * thread which does not need the MP lock to run while
689 ntd
= &gd
->gd_idlethread
;
690 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
691 set_mplock_contention_mask(gd
);
692 cpu_mplock_contested();
693 goto using_idle_thread
;
695 clr_mplock_contention_mask(gd
);
696 ++gd
->gd_cnt
.v_swtch
;
697 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
698 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
701 clr_mplock_contention_mask(gd
);
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
);
708 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
709 * worry about tokens or the BGL. However, we still have
710 * to call lwkt_getalltokens() in order to properly detect
711 * stale tokens. This call cannot fail for a UP build!
713 lwkt_getalltokens(ntd
);
714 ++gd
->gd_cnt
.v_swtch
;
715 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
716 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
720 * We have nothing to run but only let the idle loop halt
721 * the cpu if there are no pending interrupts.
723 ntd
= &gd
->gd_idlethread
;
724 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
725 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
729 * The idle thread should not be holding the MP lock unless we
730 * are trapping in the kernel or in a panic. Since we select the
731 * idle thread unconditionally when no other thread is available,
732 * if the MP lock is desired during a panic or kernel trap, we
733 * have to loop in the scheduler until we get it.
735 if (ntd
->td_mpcount
) {
736 mpheld
= MP_LOCK_HELD();
737 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
)
738 panic("Idle thread %p was holding the BGL!", ntd
);
745 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
746 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
749 * Do the actual switch. If the new target does not need the MP lock
750 * and we are holding it, release the MP lock. If the new target requires
751 * the MP lock we have already acquired it for the target.
754 if (ntd
->td_mpcount
== 0 ) {
758 ASSERT_MP_LOCK_HELD(ntd
);
765 int tos_ok __debugvar
= jg_tos_ok(ntd
);
769 KTR_LOG(ctxsw_sw
, td
, ntd
);
772 /* NOTE: current cpu may have changed after switch */
777 * Request that the target thread preempt the current thread. Preemption
778 * only works under a specific set of conditions:
780 * - We are not preempting ourselves
781 * - The target thread is owned by the current cpu
782 * - We are not currently being preempted
783 * - The target is not currently being preempted
784 * - We are not holding any spin locks
785 * - The target thread is not holding any tokens
786 * - We are able to satisfy the target's MP lock requirements (if any).
788 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
789 * this is called via lwkt_schedule() through the td_preemptable callback.
790 * critpri is the managed critical priority that we should ignore in order
791 * to determine whether preemption is possible (aka usually just the crit
792 * priority of lwkt_schedule() itself).
794 * XXX at the moment we run the target thread in a critical section during
795 * the preemption in order to prevent the target from taking interrupts
796 * that *WE* can't. Preemption is strictly limited to interrupt threads
797 * and interrupt-like threads, outside of a critical section, and the
798 * preempted source thread will be resumed the instant the target blocks
799 * whether or not the source is scheduled (i.e. preemption is supposed to
800 * be as transparent as possible).
802 * The target thread inherits our MP count (added to its own) for the
803 * duration of the preemption in order to preserve the atomicy of the
804 * MP lock during the preemption. Therefore, any preempting targets must be
805 * careful in regards to MP assertions. Note that the MP count may be
806 * out of sync with the physical mp_lock, but we do not have to preserve
807 * the original ownership of the lock if it was out of synch (that is, we
808 * can leave it synchronized on return).
811 lwkt_preempt(thread_t ntd
, int critpri
)
813 struct globaldata
*gd
= mycpu
;
821 * The caller has put us in a critical section. We can only preempt
822 * if the caller of the caller was not in a critical section (basically
823 * a local interrupt), as determined by the 'critpri' parameter. We
824 * also can't preempt if the caller is holding any spinlocks (even if
825 * he isn't in a critical section). This also handles the tokens test.
827 * YYY The target thread must be in a critical section (else it must
828 * inherit our critical section? I dunno yet).
830 * Set need_lwkt_resched() unconditionally for now YYY.
832 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
834 td
= gd
->gd_curthread
;
835 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
839 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
845 if (ntd
->td_gd
!= gd
) {
852 * Take the easy way out and do not preempt if we are holding
853 * any spinlocks. We could test whether the thread(s) being
854 * preempted interlock against the target thread's tokens and whether
855 * we can get all the target thread's tokens, but this situation
856 * should not occur very often so its easier to simply not preempt.
857 * Also, plain spinlocks are impossible to figure out at this point so
858 * just don't preempt.
860 * Do not try to preempt if the target thread is holding any tokens.
861 * We could try to acquire the tokens but this case is so rare there
862 * is no need to support it.
864 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
874 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
879 if (ntd
->td_preempted
) {
886 * note: an interrupt might have occured just as we were transitioning
887 * to or from the MP lock. In this case td_mpcount will be pre-disposed
888 * (non-zero) but not actually synchronized with the actual state of the
889 * lock. We can use it to imply an MP lock requirement for the
890 * preemption but we cannot use it to test whether we hold the MP lock
893 savecnt
= td
->td_mpcount
;
894 mpheld
= MP_LOCK_HELD();
895 ntd
->td_mpcount
+= td
->td_mpcount
;
896 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
897 ntd
->td_mpcount
-= td
->td_mpcount
;
905 * Since we are able to preempt the current thread, there is no need to
906 * call need_lwkt_resched().
909 ntd
->td_preempted
= td
;
910 td
->td_flags
|= TDF_PREEMPT_LOCK
;
911 KTR_LOG(ctxsw_pre
, td
, ntd
);
914 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
916 KKASSERT(savecnt
== td
->td_mpcount
);
917 mpheld
= MP_LOCK_HELD();
918 if (mpheld
&& td
->td_mpcount
== 0)
920 else if (mpheld
== 0 && td
->td_mpcount
)
921 panic("lwkt_preempt(): MP lock was not held through");
923 ntd
->td_preempted
= NULL
;
924 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
928 * Conditionally call splz() if gd_reqflags indicates work is pending.
930 * td_nest_count prevents deep nesting via splz() or doreti() which
931 * might otherwise blow out the kernel stack. Note that except for
932 * this special case, we MUST call splz() here to handle any
933 * pending ints, particularly after we switch, or we might accidently
934 * halt the cpu with interrupts pending.
936 * (self contained on a per cpu basis)
941 globaldata_t gd
= mycpu
;
942 thread_t td
= gd
->gd_curthread
;
944 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
949 * This implements a normal yield which will yield to equal priority
950 * threads as well as higher priority threads. Note that gd_reqflags
951 * tests will be handled by the crit_exit() call in lwkt_switch().
953 * (self contained on a per cpu basis)
958 lwkt_schedule_self(curthread
);
963 * This function is used along with the lwkt_passive_recover() inline
964 * by the trap code to negotiate a passive release of the current
965 * process/lwp designation with the user scheduler.
968 lwkt_passive_release(struct thread
*td
)
970 struct lwp
*lp
= td
->td_lwp
;
972 td
->td_release
= NULL
;
973 lwkt_setpri_self(TDPRI_KERN_USER
);
974 lp
->lwp_proc
->p_usched
->release_curproc(lp
);
978 * Make a kernel thread act as if it were in user mode with regards
979 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
980 * loops which may be potentially cpu-bound can call lwkt_user_yield().
982 * The lwkt_user_yield() function is designed to have very low overhead
983 * if no yield is determined to be needed.
986 lwkt_user_yield(void)
988 thread_t td
= curthread
;
989 struct lwp
*lp
= td
->td_lwp
;
993 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
994 * kernel can prevent other cpus from servicing interrupt threads
995 * which still require the MP lock (which is a lot of them). This
996 * has a chaining effect since if the interrupt is blocked, so is
997 * the event, so normal scheduling will not pick up on the problem.
999 if (mp_lock_contention_mask
&& td
->td_mpcount
) {
1005 * Another kernel thread wants the cpu
1007 if (lwkt_resched_wanted())
1011 * If the user scheduler has asynchronously determined that the current
1012 * process (when running in user mode) needs to lose the cpu then make
1013 * sure we are released.
1015 if (user_resched_wanted()) {
1021 * If we are released reduce our priority
1023 if (td
->td_release
== NULL
) {
1024 if (lwkt_check_resched(td
) > 0)
1027 lp
->lwp_proc
->p_usched
->acquire_curproc(lp
);
1028 td
->td_release
= lwkt_passive_release
;
1029 lwkt_setpri_self(TDPRI_USER_NORM
);
1035 * Return 0 if no runnable threads are pending at the same or higher
1036 * priority as the passed thread.
1038 * Return 1 if runnable threads are pending at the same priority.
1040 * Return 2 if runnable threads are pending at a higher priority.
1043 lwkt_check_resched(thread_t td
)
1045 int pri
= td
->td_pri
& TDPRI_MASK
;
1047 if (td
->td_gd
->gd_runqmask
> (2 << pri
) - 1)
1049 if (TAILQ_NEXT(td
, td_threadq
))
1055 * Generic schedule. Possibly schedule threads belonging to other cpus and
1056 * deal with threads that might be blocked on a wait queue.
1058 * We have a little helper inline function which does additional work after
1059 * the thread has been enqueued, including dealing with preemption and
1060 * setting need_lwkt_resched() (which prevents the kernel from returning
1061 * to userland until it has processed higher priority threads).
1063 * It is possible for this routine to be called after a failed _enqueue
1064 * (due to the target thread migrating, sleeping, or otherwise blocked).
1065 * We have to check that the thread is actually on the run queue!
1067 * reschedok is an optimized constant propagated from lwkt_schedule() or
1068 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1069 * reschedule to be requested if the target thread has a higher priority.
1070 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1071 * be 0, prevented undesired reschedules.
1075 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
, int reschedok
)
1079 if (ntd
->td_flags
& TDF_RUNQ
) {
1080 if (ntd
->td_preemptable
&& reschedok
) {
1081 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
1082 } else if (reschedok
) {
1084 if ((ntd
->td_pri
& TDPRI_MASK
) > (otd
->td_pri
& TDPRI_MASK
))
1085 need_lwkt_resched();
1092 _lwkt_schedule(thread_t td
, int reschedok
)
1094 globaldata_t mygd
= mycpu
;
1096 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1097 crit_enter_gd(mygd
);
1098 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1099 if (td
== mygd
->gd_curthread
) {
1103 * If we own the thread, there is no race (since we are in a
1104 * critical section). If we do not own the thread there might
1105 * be a race but the target cpu will deal with it.
1108 if (td
->td_gd
== mygd
) {
1110 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1112 lwkt_send_ipiq3(td
->td_gd
, lwkt_schedule_remote
, td
, 0);
1116 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1123 lwkt_schedule(thread_t td
)
1125 _lwkt_schedule(td
, 1);
1129 lwkt_schedule_noresched(thread_t td
)
1131 _lwkt_schedule(td
, 0);
1137 * When scheduled remotely if frame != NULL the IPIQ is being
1138 * run via doreti or an interrupt then preemption can be allowed.
1140 * To allow preemption we have to drop the critical section so only
1141 * one is present in _lwkt_schedule_post.
1144 lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
)
1146 thread_t td
= curthread
;
1149 if (frame
&& ntd
->td_preemptable
) {
1150 crit_exit_noyield(td
);
1151 _lwkt_schedule(ntd
, 1);
1152 crit_enter_quick(td
);
1154 _lwkt_schedule(ntd
, 1);
1159 * Thread migration using a 'Pull' method. The thread may or may not be
1160 * the current thread. It MUST be descheduled and in a stable state.
1161 * lwkt_giveaway() must be called on the cpu owning the thread.
1163 * At any point after lwkt_giveaway() is called, the target cpu may
1164 * 'pull' the thread by calling lwkt_acquire().
1166 * We have to make sure the thread is not sitting on a per-cpu tsleep
1167 * queue or it will blow up when it moves to another cpu.
1169 * MPSAFE - must be called under very specific conditions.
1172 lwkt_giveaway(thread_t td
)
1174 globaldata_t gd
= mycpu
;
1177 if (td
->td_flags
& TDF_TSLEEPQ
)
1179 KKASSERT(td
->td_gd
== gd
);
1180 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1181 td
->td_flags
|= TDF_MIGRATING
;
1186 lwkt_acquire(thread_t td
)
1191 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1196 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1197 crit_enter_gd(mygd
);
1198 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1200 lwkt_process_ipiq();
1205 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1206 td
->td_flags
&= ~TDF_MIGRATING
;
1209 crit_enter_gd(mygd
);
1210 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1211 td
->td_flags
&= ~TDF_MIGRATING
;
1219 * Generic deschedule. Descheduling threads other then your own should be
1220 * done only in carefully controlled circumstances. Descheduling is
1223 * This function may block if the cpu has run out of messages.
1226 lwkt_deschedule(thread_t td
)
1230 if (td
== curthread
) {
1233 if (td
->td_gd
== mycpu
) {
1236 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1246 * Set the target thread's priority. This routine does not automatically
1247 * switch to a higher priority thread, LWKT threads are not designed for
1248 * continuous priority changes. Yield if you want to switch.
1250 * We have to retain the critical section count which uses the high bits
1251 * of the td_pri field. The specified priority may also indicate zero or
1252 * more critical sections by adding TDPRI_CRIT*N.
1254 * Note that we requeue the thread whether it winds up on a different runq
1255 * or not. uio_yield() depends on this and the routine is not normally
1256 * called with the same priority otherwise.
1259 lwkt_setpri(thread_t td
, int pri
)
1262 KKASSERT(td
->td_gd
== mycpu
);
1264 if (td
->td_flags
& TDF_RUNQ
) {
1266 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1269 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1275 * Set the initial priority for a thread prior to it being scheduled for
1276 * the first time. The thread MUST NOT be scheduled before or during
1277 * this call. The thread may be assigned to a cpu other then the current
1280 * Typically used after a thread has been created with TDF_STOPPREQ,
1281 * and before the thread is initially scheduled.
1284 lwkt_setpri_initial(thread_t td
, int pri
)
1287 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1288 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1292 lwkt_setpri_self(int pri
)
1294 thread_t td
= curthread
;
1296 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1298 if (td
->td_flags
& TDF_RUNQ
) {
1300 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1303 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1309 * Migrate the current thread to the specified cpu.
1311 * This is accomplished by descheduling ourselves from the current cpu,
1312 * moving our thread to the tdallq of the target cpu, IPI messaging the
1313 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1314 * races while the thread is being migrated.
1316 * We must be sure to remove ourselves from the current cpu's tsleepq
1317 * before potentially moving to another queue. The thread can be on
1318 * a tsleepq due to a left-over tsleep_interlock().
1321 static void lwkt_setcpu_remote(void *arg
);
1325 lwkt_setcpu_self(globaldata_t rgd
)
1328 thread_t td
= curthread
;
1330 if (td
->td_gd
!= rgd
) {
1331 crit_enter_quick(td
);
1332 if (td
->td_flags
& TDF_TSLEEPQ
)
1334 td
->td_flags
|= TDF_MIGRATING
;
1335 lwkt_deschedule_self(td
);
1336 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1337 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1339 /* we are now on the target cpu */
1340 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1341 crit_exit_quick(td
);
1347 lwkt_migratecpu(int cpuid
)
1352 rgd
= globaldata_find(cpuid
);
1353 lwkt_setcpu_self(rgd
);
1358 * Remote IPI for cpu migration (called while in a critical section so we
1359 * do not have to enter another one). The thread has already been moved to
1360 * our cpu's allq, but we must wait for the thread to be completely switched
1361 * out on the originating cpu before we schedule it on ours or the stack
1362 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1363 * change to main memory.
1365 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1366 * against wakeups. It is best if this interface is used only when there
1367 * are no pending events that might try to schedule the thread.
1371 lwkt_setcpu_remote(void *arg
)
1374 globaldata_t gd
= mycpu
;
1376 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1378 lwkt_process_ipiq();
1384 td
->td_flags
&= ~TDF_MIGRATING
;
1385 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1391 lwkt_preempted_proc(void)
1393 thread_t td
= curthread
;
1394 while (td
->td_preempted
)
1395 td
= td
->td_preempted
;
1400 * Create a kernel process/thread/whatever. It shares it's address space
1401 * with proc0 - ie: kernel only.
1403 * NOTE! By default new threads are created with the MP lock held. A
1404 * thread which does not require the MP lock should release it by calling
1405 * rel_mplock() at the start of the new thread.
1408 lwkt_create(void (*func
)(void *), void *arg
,
1409 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1410 const char *fmt
, ...)
1415 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1419 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1422 * Set up arg0 for 'ps' etc
1424 __va_start(ap
, fmt
);
1425 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1429 * Schedule the thread to run
1431 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1434 td
->td_flags
&= ~TDF_STOPREQ
;
1439 * Destroy an LWKT thread. Warning! This function is not called when
1440 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1441 * uses a different reaping mechanism.
1446 thread_t td
= curthread
;
1450 if (td
->td_flags
& TDF_VERBOSE
)
1451 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1455 * Get us into a critical section to interlock gd_freetd and loop
1456 * until we can get it freed.
1458 * We have to cache the current td in gd_freetd because objcache_put()ing
1459 * it would rip it out from under us while our thread is still active.
1462 crit_enter_quick(td
);
1463 while ((std
= gd
->gd_freetd
) != NULL
) {
1464 gd
->gd_freetd
= NULL
;
1465 objcache_put(thread_cache
, std
);
1469 * Remove thread resources from kernel lists and deschedule us for
1472 if (td
->td_flags
& TDF_TSLEEPQ
)
1475 lwkt_deschedule_self(td
);
1476 lwkt_remove_tdallq(td
);
1477 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1483 lwkt_remove_tdallq(thread_t td
)
1485 KKASSERT(td
->td_gd
== mycpu
);
1486 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1492 thread_t td
= curthread
;
1493 int lpri
= td
->td_pri
;
1496 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1502 * Called from debugger/panic on cpus which have been stopped. We must still
1503 * process the IPIQ while stopped, even if we were stopped while in a critical
1506 * If we are dumping also try to process any pending interrupts. This may
1507 * or may not work depending on the state of the cpu at the point it was
1511 lwkt_smp_stopped(void)
1513 globaldata_t gd
= mycpu
;
1517 lwkt_process_ipiq();
1520 lwkt_process_ipiq();