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 1.119 2008/09/11 01:11:42 y0netan1 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.
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
49 #include <sys/rtprio.h>
50 #include <sys/queue.h>
51 #include <sys/sysctl.h>
52 #include <sys/kthread.h>
53 #include <machine/cpu.h>
56 #include <sys/spinlock.h>
59 #include <sys/thread2.h>
60 #include <sys/spinlock2.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_extern.h>
71 #include <machine/stdarg.h>
72 #include <machine/smp.h>
78 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
80 static int untimely_switch
= 0;
82 static int panic_on_cscount
= 0;
84 static __int64_t switch_count
= 0;
85 static __int64_t preempt_hit
= 0;
86 static __int64_t preempt_miss
= 0;
87 static __int64_t preempt_weird
= 0;
88 static __int64_t token_contention_count
= 0;
89 static __int64_t mplock_contention_count
= 0;
90 static int lwkt_use_spin_port
;
91 static struct objcache
*thread_cache
;
94 * We can make all thread ports use the spin backend instead of the thread
95 * backend. This should only be set to debug the spin backend.
97 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
99 SYSCTL_INT(_lwkt
, OID_AUTO
, untimely_switch
, CTLFLAG_RW
, &untimely_switch
, 0, "");
101 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
103 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
104 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
105 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
106 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
108 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
109 &token_contention_count
, 0, "spinning due to token contention");
110 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
111 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
117 #if !defined(KTR_GIANT_CONTENTION)
118 #define KTR_GIANT_CONTENTION KTR_ALL
121 KTR_INFO_MASTER(giant
);
122 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
123 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
125 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
128 * These helper procedures handle the runq, they can only be called from
129 * within a critical section.
131 * WARNING! Prior to SMP being brought up it is possible to enqueue and
132 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
133 * instead of 'mycpu' when referencing the globaldata structure. Once
134 * SMP live enqueuing and dequeueing only occurs on the current cpu.
138 _lwkt_dequeue(thread_t td
)
140 if (td
->td_flags
& TDF_RUNQ
) {
141 int nq
= td
->td_pri
& TDPRI_MASK
;
142 struct globaldata
*gd
= td
->td_gd
;
144 td
->td_flags
&= ~TDF_RUNQ
;
145 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
146 /* runqmask is passively cleaned up by the switcher */
152 _lwkt_enqueue(thread_t td
)
154 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_TSLEEPQ
|TDF_BLOCKQ
)) == 0) {
155 int nq
= td
->td_pri
& TDPRI_MASK
;
156 struct globaldata
*gd
= td
->td_gd
;
158 td
->td_flags
|= TDF_RUNQ
;
159 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
160 gd
->gd_runqmask
|= 1 << nq
;
165 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
167 struct thread
*td
= (struct thread
*)obj
;
169 td
->td_kstack
= NULL
;
170 td
->td_kstack_size
= 0;
171 td
->td_flags
= TDF_ALLOCATED_THREAD
;
176 _lwkt_thread_dtor(void *obj
, void *privdata
)
178 struct thread
*td
= (struct thread
*)obj
;
180 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
181 ("_lwkt_thread_dtor: not allocated from objcache"));
182 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
183 td
->td_kstack_size
> 0,
184 ("_lwkt_thread_dtor: corrupted stack"));
185 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
189 * Initialize the lwkt s/system.
194 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
195 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
), 0,
196 CACHE_NTHREADS
/2, _lwkt_thread_ctor
, _lwkt_thread_dtor
,
201 * Schedule a thread to run. As the current thread we can always safely
202 * schedule ourselves, and a shortcut procedure is provided for that
205 * (non-blocking, self contained on a per cpu basis)
208 lwkt_schedule_self(thread_t td
)
210 crit_enter_quick(td
);
211 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
212 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
218 * Deschedule a thread.
220 * (non-blocking, self contained on a per cpu basis)
223 lwkt_deschedule_self(thread_t td
)
225 crit_enter_quick(td
);
231 * LWKTs operate on a per-cpu basis
233 * WARNING! Called from early boot, 'mycpu' may not work yet.
236 lwkt_gdinit(struct globaldata
*gd
)
240 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
241 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
243 TAILQ_INIT(&gd
->gd_tdallq
);
247 * Create a new thread. The thread must be associated with a process context
248 * or LWKT start address before it can be scheduled. If the target cpu is
249 * -1 the thread will be created on the current cpu.
251 * If you intend to create a thread without a process context this function
252 * does everything except load the startup and switcher function.
255 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
257 globaldata_t gd
= mycpu
;
261 * If static thread storage is not supplied allocate a thread. Reuse
262 * a cached free thread if possible. gd_freetd is used to keep an exiting
263 * thread intact through the exit.
266 if ((td
= gd
->gd_freetd
) != NULL
)
267 gd
->gd_freetd
= NULL
;
269 td
= objcache_get(thread_cache
, M_WAITOK
);
270 KASSERT((td
->td_flags
&
271 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
272 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
273 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
277 * Try to reuse cached stack.
279 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
280 if (flags
& TDF_ALLOCATED_STACK
) {
281 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
286 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
287 flags
|= TDF_ALLOCATED_STACK
;
290 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
292 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
297 * Initialize a preexisting thread structure. This function is used by
298 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
300 * All threads start out in a critical section at a priority of
301 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
302 * appropriate. This function may send an IPI message when the
303 * requested cpu is not the current cpu and consequently gd_tdallq may
304 * not be initialized synchronously from the point of view of the originating
307 * NOTE! we have to be careful in regards to creating threads for other cpus
308 * if SMP has not yet been activated.
313 lwkt_init_thread_remote(void *arg
)
318 * Protected by critical section held by IPI dispatch
320 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
326 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
327 struct globaldata
*gd
)
329 globaldata_t mygd
= mycpu
;
331 bzero(td
, sizeof(struct thread
));
332 td
->td_kstack
= stack
;
333 td
->td_kstack_size
= stksize
;
334 td
->td_flags
= flags
;
336 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
338 if ((flags
& TDF_MPSAFE
) == 0)
341 if (lwkt_use_spin_port
)
342 lwkt_initport_spin(&td
->td_msgport
);
344 lwkt_initport_thread(&td
->td_msgport
, td
);
345 pmap_init_thread(td
);
348 * Normally initializing a thread for a remote cpu requires sending an
349 * IPI. However, the idlethread is setup before the other cpus are
350 * activated so we have to treat it as a special case. XXX manipulation
351 * of gd_tdallq requires the BGL.
353 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
355 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
358 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
362 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
368 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
373 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
378 lwkt_hold(thread_t td
)
384 lwkt_rele(thread_t td
)
386 KKASSERT(td
->td_refs
> 0);
391 lwkt_wait_free(thread_t td
)
394 tsleep(td
, 0, "tdreap", hz
);
398 lwkt_free_thread(thread_t td
)
400 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
401 ("lwkt_free_thread: did not exit! %p", td
));
403 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
404 objcache_put(thread_cache
, td
);
405 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
406 /* client-allocated struct with internally allocated stack */
407 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
408 ("lwkt_free_thread: corrupted stack"));
409 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
410 td
->td_kstack
= NULL
;
411 td
->td_kstack_size
= 0;
417 * Switch to the next runnable lwkt. If no LWKTs are runnable then
418 * switch to the idlethread. Switching must occur within a critical
419 * section to avoid races with the scheduling queue.
421 * We always have full control over our cpu's run queue. Other cpus
422 * that wish to manipulate our queue must use the cpu_*msg() calls to
423 * talk to our cpu, so a critical section is all that is needed and
424 * the result is very, very fast thread switching.
426 * The LWKT scheduler uses a fixed priority model and round-robins at
427 * each priority level. User process scheduling is a totally
428 * different beast and LWKT priorities should not be confused with
429 * user process priorities.
431 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
432 * cleans it up. Note that the td_switch() function cannot do anything that
433 * requires the MP lock since the MP lock will have already been setup for
434 * the target thread (not the current thread). It's nice to have a scheduler
435 * that does not need the MP lock to work because it allows us to do some
436 * really cool high-performance MP lock optimizations.
438 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
439 * is not called by the current thread in the preemption case, only when
440 * the preempting thread blocks (in order to return to the original thread).
445 globaldata_t gd
= mycpu
;
446 thread_t td
= gd
->gd_curthread
;
453 * Switching from within a 'fast' (non thread switched) interrupt or IPI
454 * is illegal. However, we may have to do it anyway if we hit a fatal
455 * kernel trap or we have paniced.
457 * If this case occurs save and restore the interrupt nesting level.
459 if (gd
->gd_intr_nesting_level
) {
463 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
464 panic("lwkt_switch: cannot switch from within "
465 "a fast interrupt, yet, td %p\n", td
);
467 savegdnest
= gd
->gd_intr_nesting_level
;
468 savegdtrap
= gd
->gd_trap_nesting_level
;
469 gd
->gd_intr_nesting_level
= 0;
470 gd
->gd_trap_nesting_level
= 0;
471 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
472 td
->td_flags
|= TDF_PANICWARN
;
473 kprintf("Warning: thread switch from interrupt or IPI, "
474 "thread %p (%s)\n", td
, td
->td_comm
);
476 db_print_backtrace();
480 gd
->gd_intr_nesting_level
= savegdnest
;
481 gd
->gd_trap_nesting_level
= savegdtrap
;
487 * Passive release (used to transition from user to kernel mode
488 * when we block or switch rather then when we enter the kernel).
489 * This function is NOT called if we are switching into a preemption
490 * or returning from a preemption. Typically this causes us to lose
491 * our current process designation (if we have one) and become a true
492 * LWKT thread, and may also hand the current process designation to
493 * another process and schedule thread.
500 lwkt_relalltokens(td
);
503 * We had better not be holding any spin locks, but don't get into an
504 * endless panic loop.
506 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
507 ("lwkt_switch: still holding a shared spinlock %p!",
508 gd
->gd_spinlock_rd
));
509 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
510 ("lwkt_switch: still holding %d exclusive spinlocks!",
511 gd
->gd_spinlocks_wr
));
516 * td_mpcount cannot be used to determine if we currently hold the
517 * MP lock because get_mplock() will increment it prior to attempting
518 * to get the lock, and switch out if it can't. Our ownership of
519 * the actual lock will remain stable while we are in a critical section
520 * (but, of course, another cpu may own or release the lock so the
521 * actual value of mp_lock is not stable).
523 mpheld
= MP_LOCK_HELD();
525 if (td
->td_cscount
) {
526 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
528 if (panic_on_cscount
)
529 panic("switching while mastering cpusync");
533 if ((ntd
= td
->td_preempted
) != NULL
) {
535 * We had preempted another thread on this cpu, resume the preempted
536 * thread. This occurs transparently, whether the preempted thread
537 * was scheduled or not (it may have been preempted after descheduling
540 * We have to setup the MP lock for the original thread after backing
541 * out the adjustment that was made to curthread when the original
544 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
546 if (ntd
->td_mpcount
&& mpheld
== 0) {
547 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
548 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
550 if (ntd
->td_mpcount
) {
551 td
->td_mpcount
-= ntd
->td_mpcount
;
552 KKASSERT(td
->td_mpcount
>= 0);
555 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
558 * XXX. The interrupt may have woken a thread up, we need to properly
559 * set the reschedule flag if the originally interrupted thread is at
562 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
564 /* YYY release mp lock on switchback if original doesn't need it */
567 * Priority queue / round-robin at each priority. Note that user
568 * processes run at a fixed, low priority and the user process
569 * scheduler deals with interactions between user processes
570 * by scheduling and descheduling them from the LWKT queue as
573 * We have to adjust the MP lock for the target thread. If we
574 * need the MP lock and cannot obtain it we try to locate a
575 * thread that does not need the MP lock. If we cannot, we spin
578 * A similar issue exists for the tokens held by the target thread.
579 * If we cannot obtain ownership of the tokens we cannot immediately
580 * schedule the thread.
584 * If an LWKT reschedule was requested, well that is what we are
585 * doing now so clear it.
587 clear_lwkt_resched();
589 if (gd
->gd_runqmask
) {
590 int nq
= bsrl(gd
->gd_runqmask
);
591 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
592 gd
->gd_runqmask
&= ~(1 << nq
);
597 * THREAD SELECTION FOR AN SMP MACHINE BUILD
599 * If the target needs the MP lock and we couldn't get it,
600 * or if the target is holding tokens and we could not
601 * gain ownership of the tokens, continue looking for a
602 * thread to schedule and spin instead of HLT if we can't.
604 * NOTE: the mpheld variable invalid after this conditional, it
605 * can change due to both cpu_try_mplock() returning success
606 * AND interactions in lwkt_getalltokens() due to the fact that
607 * we are trying to check the mpcount of a thread other then
608 * the current thread. Because of this, if the current thread
609 * is not holding td_mpcount, an IPI indirectly run via
610 * lwkt_getalltokens() can obtain and release the MP lock and
611 * cause the core MP lock to be released.
613 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
614 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
616 u_int32_t rqmask
= gd
->gd_runqmask
;
618 mpheld
= MP_LOCK_HELD();
621 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
622 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
623 /* spinning due to MP lock being held */
625 ++mplock_contention_count
;
627 /* mplock still not held, 'mpheld' still valid */
632 * mpheld state invalid after getalltokens call returns
633 * failure, but the variable is only needed for
636 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
637 /* spinning due to token contention */
639 ++token_contention_count
;
641 mpheld
= MP_LOCK_HELD();
648 rqmask
&= ~(1 << nq
);
652 cpu_mplock_contested();
653 ntd
= &gd
->gd_idlethread
;
654 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
655 goto using_idle_thread
;
657 ++gd
->gd_cnt
.v_swtch
;
658 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
659 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
662 ++gd
->gd_cnt
.v_swtch
;
663 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
664 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
668 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
669 * worry about tokens or the BGL. However, we still have
670 * to call lwkt_getalltokens() in order to properly detect
671 * stale tokens. This call cannot fail for a UP build!
673 lwkt_getalltokens(ntd
);
674 ++gd
->gd_cnt
.v_swtch
;
675 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
676 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
680 * We have nothing to run but only let the idle loop halt
681 * the cpu if there are no pending interrupts.
683 ntd
= &gd
->gd_idlethread
;
684 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
685 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
689 * The idle thread should not be holding the MP lock unless we
690 * are trapping in the kernel or in a panic. Since we select the
691 * idle thread unconditionally when no other thread is available,
692 * if the MP lock is desired during a panic or kernel trap, we
693 * have to loop in the scheduler until we get it.
695 if (ntd
->td_mpcount
) {
696 mpheld
= MP_LOCK_HELD();
697 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
698 panic("Idle thread %p was holding the BGL!", ntd
);
699 } else if (mpheld
== 0) {
700 cpu_mplock_contested();
707 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
708 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
711 * Do the actual switch. If the new target does not need the MP lock
712 * and we are holding it, release the MP lock. If the new target requires
713 * the MP lock we have already acquired it for the target.
716 if (ntd
->td_mpcount
== 0 ) {
720 ASSERT_MP_LOCK_HELD(ntd
);
727 /* NOTE: current cpu may have changed after switch */
732 * Request that the target thread preempt the current thread. Preemption
733 * only works under a specific set of conditions:
735 * - We are not preempting ourselves
736 * - The target thread is owned by the current cpu
737 * - We are not currently being preempted
738 * - The target is not currently being preempted
739 * - We are not holding any spin locks
740 * - The target thread is not holding any tokens
741 * - We are able to satisfy the target's MP lock requirements (if any).
743 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
744 * this is called via lwkt_schedule() through the td_preemptable callback.
745 * critpri is the managed critical priority that we should ignore in order
746 * to determine whether preemption is possible (aka usually just the crit
747 * priority of lwkt_schedule() itself).
749 * XXX at the moment we run the target thread in a critical section during
750 * the preemption in order to prevent the target from taking interrupts
751 * that *WE* can't. Preemption is strictly limited to interrupt threads
752 * and interrupt-like threads, outside of a critical section, and the
753 * preempted source thread will be resumed the instant the target blocks
754 * whether or not the source is scheduled (i.e. preemption is supposed to
755 * be as transparent as possible).
757 * The target thread inherits our MP count (added to its own) for the
758 * duration of the preemption in order to preserve the atomicy of the
759 * MP lock during the preemption. Therefore, any preempting targets must be
760 * careful in regards to MP assertions. Note that the MP count may be
761 * out of sync with the physical mp_lock, but we do not have to preserve
762 * the original ownership of the lock if it was out of synch (that is, we
763 * can leave it synchronized on return).
766 lwkt_preempt(thread_t ntd
, int critpri
)
768 struct globaldata
*gd
= mycpu
;
776 * The caller has put us in a critical section. We can only preempt
777 * if the caller of the caller was not in a critical section (basically
778 * a local interrupt), as determined by the 'critpri' parameter. We
779 * also can't preempt if the caller is holding any spinlocks (even if
780 * he isn't in a critical section). This also handles the tokens test.
782 * YYY The target thread must be in a critical section (else it must
783 * inherit our critical section? I dunno yet).
785 * Set need_lwkt_resched() unconditionally for now YYY.
787 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
789 td
= gd
->gd_curthread
;
790 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
794 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
800 if (ntd
->td_gd
!= gd
) {
807 * Take the easy way out and do not preempt if we are holding
808 * any spinlocks. We could test whether the thread(s) being
809 * preempted interlock against the target thread's tokens and whether
810 * we can get all the target thread's tokens, but this situation
811 * should not occur very often so its easier to simply not preempt.
812 * Also, plain spinlocks are impossible to figure out at this point so
813 * just don't preempt.
815 * Do not try to preempt if the target thread is holding any tokens.
816 * We could try to acquire the tokens but this case is so rare there
817 * is no need to support it.
819 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
829 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
834 if (ntd
->td_preempted
) {
841 * note: an interrupt might have occured just as we were transitioning
842 * to or from the MP lock. In this case td_mpcount will be pre-disposed
843 * (non-zero) but not actually synchronized with the actual state of the
844 * lock. We can use it to imply an MP lock requirement for the
845 * preemption but we cannot use it to test whether we hold the MP lock
848 savecnt
= td
->td_mpcount
;
849 mpheld
= MP_LOCK_HELD();
850 ntd
->td_mpcount
+= td
->td_mpcount
;
851 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
852 ntd
->td_mpcount
-= td
->td_mpcount
;
860 * Since we are able to preempt the current thread, there is no need to
861 * call need_lwkt_resched().
864 ntd
->td_preempted
= td
;
865 td
->td_flags
|= TDF_PREEMPT_LOCK
;
867 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
869 KKASSERT(savecnt
== td
->td_mpcount
);
870 mpheld
= MP_LOCK_HELD();
871 if (mpheld
&& td
->td_mpcount
== 0)
873 else if (mpheld
== 0 && td
->td_mpcount
)
874 panic("lwkt_preempt(): MP lock was not held through");
876 ntd
->td_preempted
= NULL
;
877 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
881 * Yield our thread while higher priority threads are pending. This is
882 * typically called when we leave a critical section but it can be safely
883 * called while we are in a critical section.
885 * This function will not generally yield to equal priority threads but it
886 * can occur as a side effect. Note that lwkt_switch() is called from
887 * inside the critical section to prevent its own crit_exit() from reentering
888 * lwkt_yield_quick().
890 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
891 * came along but was blocked and made pending.
893 * (self contained on a per cpu basis)
896 lwkt_yield_quick(void)
898 globaldata_t gd
= mycpu
;
899 thread_t td
= gd
->gd_curthread
;
902 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
903 * it with a non-zero cpl then we might not wind up calling splz after
904 * a task switch when the critical section is exited even though the
905 * new task could accept the interrupt.
907 * XXX from crit_exit() only called after last crit section is released.
908 * If called directly will run splz() even if in a critical section.
910 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
911 * except for this special case, we MUST call splz() here to handle any
912 * pending ints, particularly after we switch, or we might accidently
913 * halt the cpu with interrupts pending.
915 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
919 * YYY enabling will cause wakeup() to task-switch, which really
920 * confused the old 4.x code. This is a good way to simulate
921 * preemption and MP without actually doing preemption or MP, because a
922 * lot of code assumes that wakeup() does not block.
924 if (untimely_switch
&& td
->td_nest_count
== 0 &&
925 gd
->gd_intr_nesting_level
== 0
927 crit_enter_quick(td
);
929 * YYY temporary hacks until we disassociate the userland scheduler
930 * from the LWKT scheduler.
932 if (td
->td_flags
& TDF_RUNQ
) {
933 lwkt_switch(); /* will not reenter yield function */
935 lwkt_schedule_self(td
); /* make sure we are scheduled */
936 lwkt_switch(); /* will not reenter yield function */
937 lwkt_deschedule_self(td
); /* make sure we are descheduled */
939 crit_exit_noyield(td
);
944 * This implements a normal yield which, unlike _quick, will yield to equal
945 * priority threads as well. Note that gd_reqflags tests will be handled by
946 * the crit_exit() call in lwkt_switch().
948 * (self contained on a per cpu basis)
953 lwkt_schedule_self(curthread
);
958 * Generic schedule. Possibly schedule threads belonging to other cpus and
959 * deal with threads that might be blocked on a wait queue.
961 * We have a little helper inline function which does additional work after
962 * the thread has been enqueued, including dealing with preemption and
963 * setting need_lwkt_resched() (which prevents the kernel from returning
964 * to userland until it has processed higher priority threads).
966 * It is possible for this routine to be called after a failed _enqueue
967 * (due to the target thread migrating, sleeping, or otherwise blocked).
968 * We have to check that the thread is actually on the run queue!
970 * reschedok is an optimized constant propagated from lwkt_schedule() or
971 * lwkt_schedule_noresched(). By default it is non-zero, causing a
972 * reschedule to be requested if the target thread has a higher priority.
973 * The port messaging code will set MSG_NORESCHED and cause reschedok to
974 * be 0, prevented undesired reschedules.
978 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
, int reschedok
)
982 if (ntd
->td_flags
& TDF_RUNQ
) {
983 if (ntd
->td_preemptable
&& reschedok
) {
984 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
985 } else if (reschedok
) {
987 * This is a little sticky. Due to the passive release function
988 * the LWKT priority can wiggle around for threads acting in
989 * the kernel on behalf of a user process. We do not want this
990 * to effect the comparison per-say.
992 * What will happen is that the current user process will be
993 * allowed to run until the next hardclock at which time a
994 * forced need_lwkt_resched() will allow the other kernel mode
995 * threads to get in their two cents. This prevents cavitation.
997 mypri
= gd
->gd_curthread
->td_pri
& TDPRI_MASK
;
998 if (mypri
>= TDPRI_USER_IDLE
&& mypri
<= TDPRI_USER_REAL
)
999 mypri
= TDPRI_KERN_USER
;
1001 if ((ntd
->td_pri
& TDPRI_MASK
) > mypri
)
1002 need_lwkt_resched();
1009 _lwkt_schedule(thread_t td
, int reschedok
)
1011 globaldata_t mygd
= mycpu
;
1013 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1014 crit_enter_gd(mygd
);
1015 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1016 if (td
== mygd
->gd_curthread
) {
1020 * If we own the thread, there is no race (since we are in a
1021 * critical section). If we do not own the thread there might
1022 * be a race but the target cpu will deal with it.
1025 if (td
->td_gd
== mygd
) {
1027 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1029 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_schedule
, td
);
1033 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1040 lwkt_schedule(thread_t td
)
1042 _lwkt_schedule(td
, 1);
1046 lwkt_schedule_noresched(thread_t td
)
1048 _lwkt_schedule(td
, 0);
1054 * Thread migration using a 'Pull' method. The thread may or may not be
1055 * the current thread. It MUST be descheduled and in a stable state.
1056 * lwkt_giveaway() must be called on the cpu owning the thread.
1058 * At any point after lwkt_giveaway() is called, the target cpu may
1059 * 'pull' the thread by calling lwkt_acquire().
1061 * MPSAFE - must be called under very specific conditions.
1064 lwkt_giveaway(thread_t td
)
1066 globaldata_t gd
= mycpu
;
1069 KKASSERT(td
->td_gd
== gd
);
1070 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1071 td
->td_flags
|= TDF_MIGRATING
;
1076 lwkt_acquire(thread_t td
)
1081 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1086 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1087 crit_enter_gd(mygd
);
1088 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1090 lwkt_process_ipiq();
1095 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1096 td
->td_flags
&= ~TDF_MIGRATING
;
1099 crit_enter_gd(mygd
);
1100 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1101 td
->td_flags
&= ~TDF_MIGRATING
;
1109 * Generic deschedule. Descheduling threads other then your own should be
1110 * done only in carefully controlled circumstances. Descheduling is
1113 * This function may block if the cpu has run out of messages.
1116 lwkt_deschedule(thread_t td
)
1120 if (td
== curthread
) {
1123 if (td
->td_gd
== mycpu
) {
1126 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1136 * Set the target thread's priority. This routine does not automatically
1137 * switch to a higher priority thread, LWKT threads are not designed for
1138 * continuous priority changes. Yield if you want to switch.
1140 * We have to retain the critical section count which uses the high bits
1141 * of the td_pri field. The specified priority may also indicate zero or
1142 * more critical sections by adding TDPRI_CRIT*N.
1144 * Note that we requeue the thread whether it winds up on a different runq
1145 * or not. uio_yield() depends on this and the routine is not normally
1146 * called with the same priority otherwise.
1149 lwkt_setpri(thread_t td
, int pri
)
1152 KKASSERT(td
->td_gd
== mycpu
);
1154 if (td
->td_flags
& TDF_RUNQ
) {
1156 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1159 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1165 lwkt_setpri_self(int pri
)
1167 thread_t td
= curthread
;
1169 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1171 if (td
->td_flags
& TDF_RUNQ
) {
1173 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1176 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1182 * Migrate the current thread to the specified cpu.
1184 * This is accomplished by descheduling ourselves from the current cpu,
1185 * moving our thread to the tdallq of the target cpu, IPI messaging the
1186 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1187 * races while the thread is being migrated.
1190 static void lwkt_setcpu_remote(void *arg
);
1194 lwkt_setcpu_self(globaldata_t rgd
)
1197 thread_t td
= curthread
;
1199 if (td
->td_gd
!= rgd
) {
1200 crit_enter_quick(td
);
1201 td
->td_flags
|= TDF_MIGRATING
;
1202 lwkt_deschedule_self(td
);
1203 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1204 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1206 /* we are now on the target cpu */
1207 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1208 crit_exit_quick(td
);
1214 lwkt_migratecpu(int cpuid
)
1219 rgd
= globaldata_find(cpuid
);
1220 lwkt_setcpu_self(rgd
);
1225 * Remote IPI for cpu migration (called while in a critical section so we
1226 * do not have to enter another one). The thread has already been moved to
1227 * our cpu's allq, but we must wait for the thread to be completely switched
1228 * out on the originating cpu before we schedule it on ours or the stack
1229 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1230 * change to main memory.
1232 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1233 * against wakeups. It is best if this interface is used only when there
1234 * are no pending events that might try to schedule the thread.
1238 lwkt_setcpu_remote(void *arg
)
1241 globaldata_t gd
= mycpu
;
1243 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1245 lwkt_process_ipiq();
1251 td
->td_flags
&= ~TDF_MIGRATING
;
1252 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1258 lwkt_preempted_proc(void)
1260 thread_t td
= curthread
;
1261 while (td
->td_preempted
)
1262 td
= td
->td_preempted
;
1267 * Create a kernel process/thread/whatever. It shares it's address space
1268 * with proc0 - ie: kernel only.
1270 * NOTE! By default new threads are created with the MP lock held. A
1271 * thread which does not require the MP lock should release it by calling
1272 * rel_mplock() at the start of the new thread.
1275 lwkt_create(void (*func
)(void *), void *arg
,
1276 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1277 const char *fmt
, ...)
1282 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1286 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1289 * Set up arg0 for 'ps' etc
1291 __va_start(ap
, fmt
);
1292 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1296 * Schedule the thread to run
1298 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1301 td
->td_flags
&= ~TDF_STOPREQ
;
1306 * Destroy an LWKT thread. Warning! This function is not called when
1307 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1308 * uses a different reaping mechanism.
1313 thread_t td
= curthread
;
1317 if (td
->td_flags
& TDF_VERBOSE
)
1318 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1322 * Get us into a critical section to interlock gd_freetd and loop
1323 * until we can get it freed.
1325 * We have to cache the current td in gd_freetd because objcache_put()ing
1326 * it would rip it out from under us while our thread is still active.
1329 crit_enter_quick(td
);
1330 while ((std
= gd
->gd_freetd
) != NULL
) {
1331 gd
->gd_freetd
= NULL
;
1332 objcache_put(thread_cache
, std
);
1334 lwkt_deschedule_self(td
);
1335 lwkt_remove_tdallq(td
);
1336 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1342 lwkt_remove_tdallq(thread_t td
)
1344 KKASSERT(td
->td_gd
== mycpu
);
1345 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1351 thread_t td
= curthread
;
1352 int lpri
= td
->td_pri
;
1355 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1361 * Called from debugger/panic on cpus which have been stopped. We must still
1362 * process the IPIQ while stopped, even if we were stopped while in a critical
1365 * If we are dumping also try to process any pending interrupts. This may
1366 * or may not work depending on the state of the cpu at the point it was
1370 lwkt_smp_stopped(void)
1372 globaldata_t gd
= mycpu
;
1376 lwkt_process_ipiq();
1379 lwkt_process_ipiq();
1385 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1386 * get_mplock() has already incremented td_mpcount. We must block and
1387 * not return until giant is held.
1389 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1390 * reschedule the thread until it can obtain the giant lock for it.
1393 lwkt_mp_lock_contested(void)