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 1.116 2008/06/16 02:00:05 dillon 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>
73 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
75 static int untimely_switch
= 0;
77 static int panic_on_cscount
= 0;
79 static __int64_t switch_count
= 0;
80 static __int64_t preempt_hit
= 0;
81 static __int64_t preempt_miss
= 0;
82 static __int64_t preempt_weird
= 0;
83 static __int64_t token_contention_count
= 0;
84 static __int64_t mplock_contention_count
= 0;
85 static int lwkt_use_spin_port
;
86 static struct objcache
*thread_cache
;
89 * We can make all thread ports use the spin backend instead of the thread
90 * backend. This should only be set to debug the spin backend.
92 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
94 SYSCTL_INT(_lwkt
, OID_AUTO
, untimely_switch
, CTLFLAG_RW
, &untimely_switch
, 0, "");
96 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
98 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
99 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
100 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
101 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
103 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
104 &token_contention_count
, 0, "spinning due to token contention");
105 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
106 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
112 #if !defined(KTR_GIANT_CONTENTION)
113 #define KTR_GIANT_CONTENTION KTR_ALL
116 KTR_INFO_MASTER(giant
);
117 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
118 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
120 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
133 _lwkt_dequeue(thread_t td
)
135 if (td
->td_flags
& TDF_RUNQ
) {
136 int nq
= td
->td_pri
& TDPRI_MASK
;
137 struct globaldata
*gd
= td
->td_gd
;
139 td
->td_flags
&= ~TDF_RUNQ
;
140 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
141 /* runqmask is passively cleaned up by the switcher */
147 _lwkt_enqueue(thread_t td
)
149 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_TSLEEPQ
|TDF_BLOCKQ
)) == 0) {
150 int nq
= td
->td_pri
& TDPRI_MASK
;
151 struct globaldata
*gd
= td
->td_gd
;
153 td
->td_flags
|= TDF_RUNQ
;
154 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
155 gd
->gd_runqmask
|= 1 << nq
;
160 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
162 struct thread
*td
= (struct thread
*)obj
;
164 td
->td_kstack
= NULL
;
165 td
->td_kstack_size
= 0;
166 td
->td_flags
= TDF_ALLOCATED_THREAD
;
171 _lwkt_thread_dtor(void *obj
, void *privdata
)
173 struct thread
*td
= (struct thread
*)obj
;
175 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
176 ("_lwkt_thread_dtor: not allocated from objcache"));
177 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
178 td
->td_kstack_size
> 0,
179 ("_lwkt_thread_dtor: corrupted stack"));
180 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
184 * Initialize the lwkt s/system.
189 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
190 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
), 0,
191 CACHE_NTHREADS
/2, _lwkt_thread_ctor
, _lwkt_thread_dtor
,
196 * Schedule a thread to run. As the current thread we can always safely
197 * schedule ourselves, and a shortcut procedure is provided for that
200 * (non-blocking, self contained on a per cpu basis)
203 lwkt_schedule_self(thread_t td
)
205 crit_enter_quick(td
);
206 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
207 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
213 * Deschedule a thread.
215 * (non-blocking, self contained on a per cpu basis)
218 lwkt_deschedule_self(thread_t td
)
220 crit_enter_quick(td
);
226 * LWKTs operate on a per-cpu basis
228 * WARNING! Called from early boot, 'mycpu' may not work yet.
231 lwkt_gdinit(struct globaldata
*gd
)
235 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
236 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
238 TAILQ_INIT(&gd
->gd_tdallq
);
242 * Create a new thread. The thread must be associated with a process context
243 * or LWKT start address before it can be scheduled. If the target cpu is
244 * -1 the thread will be created on the current cpu.
246 * If you intend to create a thread without a process context this function
247 * does everything except load the startup and switcher function.
250 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
252 globaldata_t gd
= mycpu
;
256 * If static thread storage is not supplied allocate a thread. Reuse
257 * a cached free thread if possible. gd_freetd is used to keep an exiting
258 * thread intact through the exit.
261 if ((td
= gd
->gd_freetd
) != NULL
)
262 gd
->gd_freetd
= NULL
;
264 td
= objcache_get(thread_cache
, M_WAITOK
);
265 KASSERT((td
->td_flags
&
266 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
267 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
268 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
272 * Try to reuse cached stack.
274 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
275 if (flags
& TDF_ALLOCATED_STACK
) {
276 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
281 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
282 flags
|= TDF_ALLOCATED_STACK
;
285 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
287 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
292 * Initialize a preexisting thread structure. This function is used by
293 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
295 * All threads start out in a critical section at a priority of
296 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
297 * appropriate. This function may send an IPI message when the
298 * requested cpu is not the current cpu and consequently gd_tdallq may
299 * not be initialized synchronously from the point of view of the originating
302 * NOTE! we have to be careful in regards to creating threads for other cpus
303 * if SMP has not yet been activated.
308 lwkt_init_thread_remote(void *arg
)
313 * Protected by critical section held by IPI dispatch
315 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
321 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
322 struct globaldata
*gd
)
324 globaldata_t mygd
= mycpu
;
326 bzero(td
, sizeof(struct thread
));
327 td
->td_kstack
= stack
;
328 td
->td_kstack_size
= stksize
;
329 td
->td_flags
= flags
;
331 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
333 if ((flags
& TDF_MPSAFE
) == 0)
336 if (lwkt_use_spin_port
)
337 lwkt_initport_spin(&td
->td_msgport
);
339 lwkt_initport_thread(&td
->td_msgport
, td
);
340 pmap_init_thread(td
);
343 * Normally initializing a thread for a remote cpu requires sending an
344 * IPI. However, the idlethread is setup before the other cpus are
345 * activated so we have to treat it as a special case. XXX manipulation
346 * of gd_tdallq requires the BGL.
348 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
350 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
353 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
357 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
363 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
368 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
373 lwkt_hold(thread_t td
)
379 lwkt_rele(thread_t td
)
381 KKASSERT(td
->td_refs
> 0);
386 lwkt_wait_free(thread_t td
)
389 tsleep(td
, 0, "tdreap", hz
);
393 lwkt_free_thread(thread_t td
)
395 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
396 ("lwkt_free_thread: did not exit! %p", td
));
398 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
399 objcache_put(thread_cache
, td
);
400 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
401 /* client-allocated struct with internally allocated stack */
402 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
403 ("lwkt_free_thread: corrupted stack"));
404 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
405 td
->td_kstack
= NULL
;
406 td
->td_kstack_size
= 0;
412 * Switch to the next runnable lwkt. If no LWKTs are runnable then
413 * switch to the idlethread. Switching must occur within a critical
414 * section to avoid races with the scheduling queue.
416 * We always have full control over our cpu's run queue. Other cpus
417 * that wish to manipulate our queue must use the cpu_*msg() calls to
418 * talk to our cpu, so a critical section is all that is needed and
419 * the result is very, very fast thread switching.
421 * The LWKT scheduler uses a fixed priority model and round-robins at
422 * each priority level. User process scheduling is a totally
423 * different beast and LWKT priorities should not be confused with
424 * user process priorities.
426 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
427 * cleans it up. Note that the td_switch() function cannot do anything that
428 * requires the MP lock since the MP lock will have already been setup for
429 * the target thread (not the current thread). It's nice to have a scheduler
430 * that does not need the MP lock to work because it allows us to do some
431 * really cool high-performance MP lock optimizations.
433 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
434 * is not called by the current thread in the preemption case, only when
435 * the preempting thread blocks (in order to return to the original thread).
440 globaldata_t gd
= mycpu
;
441 thread_t td
= gd
->gd_curthread
;
448 * Switching from within a 'fast' (non thread switched) interrupt or IPI
449 * is illegal. However, we may have to do it anyway if we hit a fatal
450 * kernel trap or we have paniced.
452 * If this case occurs save and restore the interrupt nesting level.
454 if (gd
->gd_intr_nesting_level
) {
458 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
459 panic("lwkt_switch: cannot switch from within "
460 "a fast interrupt, yet, td %p\n", td
);
462 savegdnest
= gd
->gd_intr_nesting_level
;
463 savegdtrap
= gd
->gd_trap_nesting_level
;
464 gd
->gd_intr_nesting_level
= 0;
465 gd
->gd_trap_nesting_level
= 0;
466 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
467 td
->td_flags
|= TDF_PANICWARN
;
468 kprintf("Warning: thread switch from interrupt or IPI, "
469 "thread %p (%s)\n", td
, td
->td_comm
);
471 db_print_backtrace();
475 gd
->gd_intr_nesting_level
= savegdnest
;
476 gd
->gd_trap_nesting_level
= savegdtrap
;
482 * Passive release (used to transition from user to kernel mode
483 * when we block or switch rather then when we enter the kernel).
484 * This function is NOT called if we are switching into a preemption
485 * or returning from a preemption. Typically this causes us to lose
486 * our current process designation (if we have one) and become a true
487 * LWKT thread, and may also hand the current process designation to
488 * another process and schedule thread.
495 lwkt_relalltokens(td
);
498 * We had better not be holding any spin locks, but don't get into an
499 * endless panic loop.
501 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
502 ("lwkt_switch: still holding a shared spinlock %p!",
503 gd
->gd_spinlock_rd
));
504 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
505 ("lwkt_switch: still holding %d exclusive spinlocks!",
506 gd
->gd_spinlocks_wr
));
511 * td_mpcount cannot be used to determine if we currently hold the
512 * MP lock because get_mplock() will increment it prior to attempting
513 * to get the lock, and switch out if it can't. Our ownership of
514 * the actual lock will remain stable while we are in a critical section
515 * (but, of course, another cpu may own or release the lock so the
516 * actual value of mp_lock is not stable).
518 mpheld
= MP_LOCK_HELD();
520 if (td
->td_cscount
) {
521 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
523 if (panic_on_cscount
)
524 panic("switching while mastering cpusync");
528 if ((ntd
= td
->td_preempted
) != NULL
) {
530 * We had preempted another thread on this cpu, resume the preempted
531 * thread. This occurs transparently, whether the preempted thread
532 * was scheduled or not (it may have been preempted after descheduling
535 * We have to setup the MP lock for the original thread after backing
536 * out the adjustment that was made to curthread when the original
539 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
541 if (ntd
->td_mpcount
&& mpheld
== 0) {
542 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
543 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
545 if (ntd
->td_mpcount
) {
546 td
->td_mpcount
-= ntd
->td_mpcount
;
547 KKASSERT(td
->td_mpcount
>= 0);
550 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
553 * XXX. The interrupt may have woken a thread up, we need to properly
554 * set the reschedule flag if the originally interrupted thread is at
557 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
559 /* YYY release mp lock on switchback if original doesn't need it */
562 * Priority queue / round-robin at each priority. Note that user
563 * processes run at a fixed, low priority and the user process
564 * scheduler deals with interactions between user processes
565 * by scheduling and descheduling them from the LWKT queue as
568 * We have to adjust the MP lock for the target thread. If we
569 * need the MP lock and cannot obtain it we try to locate a
570 * thread that does not need the MP lock. If we cannot, we spin
573 * A similar issue exists for the tokens held by the target thread.
574 * If we cannot obtain ownership of the tokens we cannot immediately
575 * schedule the thread.
579 * If an LWKT reschedule was requested, well that is what we are
580 * doing now so clear it.
582 clear_lwkt_resched();
584 if (gd
->gd_runqmask
) {
585 int nq
= bsrl(gd
->gd_runqmask
);
586 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
587 gd
->gd_runqmask
&= ~(1 << nq
);
592 * THREAD SELECTION FOR AN SMP MACHINE BUILD
594 * If the target needs the MP lock and we couldn't get it,
595 * or if the target is holding tokens and we could not
596 * gain ownership of the tokens, continue looking for a
597 * thread to schedule and spin instead of HLT if we can't.
599 * NOTE: the mpheld variable invalid after this conditional, it
600 * can change due to both cpu_try_mplock() returning success
601 * AND interactions in lwkt_getalltokens() due to the fact that
602 * we are trying to check the mpcount of a thread other then
603 * the current thread. Because of this, if the current thread
604 * is not holding td_mpcount, an IPI indirectly run via
605 * lwkt_getalltokens() can obtain and release the MP lock and
606 * cause the core MP lock to be released.
608 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
609 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
611 u_int32_t rqmask
= gd
->gd_runqmask
;
613 mpheld
= MP_LOCK_HELD();
616 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
617 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
618 /* spinning due to MP lock being held */
620 ++mplock_contention_count
;
622 /* mplock still not held, 'mpheld' still valid */
627 * mpheld state invalid after getalltokens call returns
628 * failure, but the variable is only needed for
631 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
632 /* spinning due to token contention */
634 ++token_contention_count
;
636 mpheld
= MP_LOCK_HELD();
643 rqmask
&= ~(1 << nq
);
647 cpu_mplock_contested();
648 ntd
= &gd
->gd_idlethread
;
649 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
650 goto using_idle_thread
;
652 ++gd
->gd_cnt
.v_swtch
;
653 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
654 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
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
);
663 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
664 * worry about tokens or the BGL. However, we still have
665 * to call lwkt_getalltokens() in order to properly detect
666 * stale tokens. This call cannot fail for a UP build!
668 lwkt_getalltokens(ntd
);
669 ++gd
->gd_cnt
.v_swtch
;
670 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
671 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
675 * We have nothing to run but only let the idle loop halt
676 * the cpu if there are no pending interrupts.
678 ntd
= &gd
->gd_idlethread
;
679 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
680 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
684 * The idle thread should not be holding the MP lock unless we
685 * are trapping in the kernel or in a panic. Since we select the
686 * idle thread unconditionally when no other thread is available,
687 * if the MP lock is desired during a panic or kernel trap, we
688 * have to loop in the scheduler until we get it.
690 if (ntd
->td_mpcount
) {
691 mpheld
= MP_LOCK_HELD();
692 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
693 panic("Idle thread %p was holding the BGL!", ntd
);
694 } else if (mpheld
== 0) {
695 cpu_mplock_contested();
702 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
703 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
706 * Do the actual switch. If the new target does not need the MP lock
707 * and we are holding it, release the MP lock. If the new target requires
708 * the MP lock we have already acquired it for the target.
711 if (ntd
->td_mpcount
== 0 ) {
715 ASSERT_MP_LOCK_HELD(ntd
);
722 /* NOTE: current cpu may have changed after switch */
727 * Request that the target thread preempt the current thread. Preemption
728 * only works under a specific set of conditions:
730 * - We are not preempting ourselves
731 * - The target thread is owned by the current cpu
732 * - We are not currently being preempted
733 * - The target is not currently being preempted
734 * - We are not holding any spin locks
735 * - The target thread is not holding any tokens
736 * - We are able to satisfy the target's MP lock requirements (if any).
738 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
739 * this is called via lwkt_schedule() through the td_preemptable callback.
740 * critpri is the managed critical priority that we should ignore in order
741 * to determine whether preemption is possible (aka usually just the crit
742 * priority of lwkt_schedule() itself).
744 * XXX at the moment we run the target thread in a critical section during
745 * the preemption in order to prevent the target from taking interrupts
746 * that *WE* can't. Preemption is strictly limited to interrupt threads
747 * and interrupt-like threads, outside of a critical section, and the
748 * preempted source thread will be resumed the instant the target blocks
749 * whether or not the source is scheduled (i.e. preemption is supposed to
750 * be as transparent as possible).
752 * The target thread inherits our MP count (added to its own) for the
753 * duration of the preemption in order to preserve the atomicy of the
754 * MP lock during the preemption. Therefore, any preempting targets must be
755 * careful in regards to MP assertions. Note that the MP count may be
756 * out of sync with the physical mp_lock, but we do not have to preserve
757 * the original ownership of the lock if it was out of synch (that is, we
758 * can leave it synchronized on return).
761 lwkt_preempt(thread_t ntd
, int critpri
)
763 struct globaldata
*gd
= mycpu
;
771 * The caller has put us in a critical section. We can only preempt
772 * if the caller of the caller was not in a critical section (basically
773 * a local interrupt), as determined by the 'critpri' parameter. We
774 * also can't preempt if the caller is holding any spinlocks (even if
775 * he isn't in a critical section). This also handles the tokens test.
777 * YYY The target thread must be in a critical section (else it must
778 * inherit our critical section? I dunno yet).
780 * Set need_lwkt_resched() unconditionally for now YYY.
782 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
784 td
= gd
->gd_curthread
;
785 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
789 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
795 if (ntd
->td_gd
!= gd
) {
802 * Take the easy way out and do not preempt if we are holding
803 * any spinlocks. We could test whether the thread(s) being
804 * preempted interlock against the target thread's tokens and whether
805 * we can get all the target thread's tokens, but this situation
806 * should not occur very often so its easier to simply not preempt.
807 * Also, plain spinlocks are impossible to figure out at this point so
808 * just don't preempt.
810 * Do not try to preempt if the target thread is holding any tokens.
811 * We could try to acquire the tokens but this case is so rare there
812 * is no need to support it.
814 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
824 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
829 if (ntd
->td_preempted
) {
836 * note: an interrupt might have occured just as we were transitioning
837 * to or from the MP lock. In this case td_mpcount will be pre-disposed
838 * (non-zero) but not actually synchronized with the actual state of the
839 * lock. We can use it to imply an MP lock requirement for the
840 * preemption but we cannot use it to test whether we hold the MP lock
843 savecnt
= td
->td_mpcount
;
844 mpheld
= MP_LOCK_HELD();
845 ntd
->td_mpcount
+= td
->td_mpcount
;
846 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
847 ntd
->td_mpcount
-= td
->td_mpcount
;
855 * Since we are able to preempt the current thread, there is no need to
856 * call need_lwkt_resched().
859 ntd
->td_preempted
= td
;
860 td
->td_flags
|= TDF_PREEMPT_LOCK
;
862 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
864 KKASSERT(savecnt
== td
->td_mpcount
);
865 mpheld
= MP_LOCK_HELD();
866 if (mpheld
&& td
->td_mpcount
== 0)
868 else if (mpheld
== 0 && td
->td_mpcount
)
869 panic("lwkt_preempt(): MP lock was not held through");
871 ntd
->td_preempted
= NULL
;
872 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
876 * Yield our thread while higher priority threads are pending. This is
877 * typically called when we leave a critical section but it can be safely
878 * called while we are in a critical section.
880 * This function will not generally yield to equal priority threads but it
881 * can occur as a side effect. Note that lwkt_switch() is called from
882 * inside the critical section to prevent its own crit_exit() from reentering
883 * lwkt_yield_quick().
885 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
886 * came along but was blocked and made pending.
888 * (self contained on a per cpu basis)
891 lwkt_yield_quick(void)
893 globaldata_t gd
= mycpu
;
894 thread_t td
= gd
->gd_curthread
;
897 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
898 * it with a non-zero cpl then we might not wind up calling splz after
899 * a task switch when the critical section is exited even though the
900 * new task could accept the interrupt.
902 * XXX from crit_exit() only called after last crit section is released.
903 * If called directly will run splz() even if in a critical section.
905 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
906 * except for this special case, we MUST call splz() here to handle any
907 * pending ints, particularly after we switch, or we might accidently
908 * halt the cpu with interrupts pending.
910 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
914 * YYY enabling will cause wakeup() to task-switch, which really
915 * confused the old 4.x code. This is a good way to simulate
916 * preemption and MP without actually doing preemption or MP, because a
917 * lot of code assumes that wakeup() does not block.
919 if (untimely_switch
&& td
->td_nest_count
== 0 &&
920 gd
->gd_intr_nesting_level
== 0
922 crit_enter_quick(td
);
924 * YYY temporary hacks until we disassociate the userland scheduler
925 * from the LWKT scheduler.
927 if (td
->td_flags
& TDF_RUNQ
) {
928 lwkt_switch(); /* will not reenter yield function */
930 lwkt_schedule_self(td
); /* make sure we are scheduled */
931 lwkt_switch(); /* will not reenter yield function */
932 lwkt_deschedule_self(td
); /* make sure we are descheduled */
934 crit_exit_noyield(td
);
939 * This implements a normal yield which, unlike _quick, will yield to equal
940 * priority threads as well. Note that gd_reqflags tests will be handled by
941 * the crit_exit() call in lwkt_switch().
943 * (self contained on a per cpu basis)
948 lwkt_schedule_self(curthread
);
953 * Generic schedule. Possibly schedule threads belonging to other cpus and
954 * deal with threads that might be blocked on a wait queue.
956 * We have a little helper inline function which does additional work after
957 * the thread has been enqueued, including dealing with preemption and
958 * setting need_lwkt_resched() (which prevents the kernel from returning
959 * to userland until it has processed higher priority threads).
961 * It is possible for this routine to be called after a failed _enqueue
962 * (due to the target thread migrating, sleeping, or otherwise blocked).
963 * We have to check that the thread is actually on the run queue!
967 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
)
969 if (ntd
->td_flags
& TDF_RUNQ
) {
970 if (ntd
->td_preemptable
) {
971 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
972 } else if ((ntd
->td_flags
& TDF_NORESCHED
) == 0 &&
973 (ntd
->td_pri
& TDPRI_MASK
) > (gd
->gd_curthread
->td_pri
& TDPRI_MASK
)
981 lwkt_schedule(thread_t td
)
983 globaldata_t mygd
= mycpu
;
985 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
987 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
988 if (td
== mygd
->gd_curthread
) {
992 * If we own the thread, there is no race (since we are in a
993 * critical section). If we do not own the thread there might
994 * be a race but the target cpu will deal with it.
997 if (td
->td_gd
== mygd
) {
999 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
);
1001 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_schedule
, td
);
1005 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
);
1014 * Thread migration using a 'Pull' method. The thread may or may not be
1015 * the current thread. It MUST be descheduled and in a stable state.
1016 * lwkt_giveaway() must be called on the cpu owning the thread.
1018 * At any point after lwkt_giveaway() is called, the target cpu may
1019 * 'pull' the thread by calling lwkt_acquire().
1021 * MPSAFE - must be called under very specific conditions.
1024 lwkt_giveaway(thread_t td
)
1026 globaldata_t gd
= mycpu
;
1029 KKASSERT(td
->td_gd
== gd
);
1030 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1031 td
->td_flags
|= TDF_MIGRATING
;
1036 lwkt_acquire(thread_t td
)
1041 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1046 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1047 crit_enter_gd(mygd
);
1048 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1050 lwkt_process_ipiq();
1055 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1056 td
->td_flags
&= ~TDF_MIGRATING
;
1059 crit_enter_gd(mygd
);
1060 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1061 td
->td_flags
&= ~TDF_MIGRATING
;
1069 * Generic deschedule. Descheduling threads other then your own should be
1070 * done only in carefully controlled circumstances. Descheduling is
1073 * This function may block if the cpu has run out of messages.
1076 lwkt_deschedule(thread_t td
)
1080 if (td
== curthread
) {
1083 if (td
->td_gd
== mycpu
) {
1086 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1096 * Set the target thread's priority. This routine does not automatically
1097 * switch to a higher priority thread, LWKT threads are not designed for
1098 * continuous priority changes. Yield if you want to switch.
1100 * We have to retain the critical section count which uses the high bits
1101 * of the td_pri field. The specified priority may also indicate zero or
1102 * more critical sections by adding TDPRI_CRIT*N.
1104 * Note that we requeue the thread whether it winds up on a different runq
1105 * or not. uio_yield() depends on this and the routine is not normally
1106 * called with the same priority otherwise.
1109 lwkt_setpri(thread_t td
, int pri
)
1112 KKASSERT(td
->td_gd
== mycpu
);
1114 if (td
->td_flags
& TDF_RUNQ
) {
1116 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1119 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1125 lwkt_setpri_self(int pri
)
1127 thread_t td
= curthread
;
1129 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1131 if (td
->td_flags
& TDF_RUNQ
) {
1133 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1136 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1142 * Determine if there is a runnable thread at a higher priority then
1143 * the current thread. lwkt_setpri() does not check this automatically.
1144 * Return 1 if there is, 0 if there isn't.
1146 * Example: if bit 31 of runqmask is set and the current thread is priority
1147 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1149 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1150 * up comparing against 0xffffffff, a comparison that will always be false.
1153 lwkt_checkpri_self(void)
1155 globaldata_t gd
= mycpu
;
1156 thread_t td
= gd
->gd_curthread
;
1157 int nq
= td
->td_pri
& TDPRI_MASK
;
1159 while (gd
->gd_runqmask
> (__uint32_t
)(2 << nq
) - 1) {
1160 if (TAILQ_FIRST(&gd
->gd_tdrunq
[nq
+ 1]))
1168 * Migrate the current thread to the specified cpu.
1170 * This is accomplished by descheduling ourselves from the current cpu,
1171 * moving our thread to the tdallq of the target cpu, IPI messaging the
1172 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1173 * races while the thread is being migrated.
1176 static void lwkt_setcpu_remote(void *arg
);
1180 lwkt_setcpu_self(globaldata_t rgd
)
1183 thread_t td
= curthread
;
1185 if (td
->td_gd
!= rgd
) {
1186 crit_enter_quick(td
);
1187 td
->td_flags
|= TDF_MIGRATING
;
1188 lwkt_deschedule_self(td
);
1189 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1190 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1192 /* we are now on the target cpu */
1193 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1194 crit_exit_quick(td
);
1200 lwkt_migratecpu(int cpuid
)
1205 rgd
= globaldata_find(cpuid
);
1206 lwkt_setcpu_self(rgd
);
1211 * Remote IPI for cpu migration (called while in a critical section so we
1212 * do not have to enter another one). The thread has already been moved to
1213 * our cpu's allq, but we must wait for the thread to be completely switched
1214 * out on the originating cpu before we schedule it on ours or the stack
1215 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1216 * change to main memory.
1218 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1219 * against wakeups. It is best if this interface is used only when there
1220 * are no pending events that might try to schedule the thread.
1224 lwkt_setcpu_remote(void *arg
)
1227 globaldata_t gd
= mycpu
;
1229 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1231 lwkt_process_ipiq();
1237 td
->td_flags
&= ~TDF_MIGRATING
;
1238 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1244 lwkt_preempted_proc(void)
1246 thread_t td
= curthread
;
1247 while (td
->td_preempted
)
1248 td
= td
->td_preempted
;
1253 * Create a kernel process/thread/whatever. It shares it's address space
1254 * with proc0 - ie: kernel only.
1256 * NOTE! By default new threads are created with the MP lock held. A
1257 * thread which does not require the MP lock should release it by calling
1258 * rel_mplock() at the start of the new thread.
1261 lwkt_create(void (*func
)(void *), void *arg
,
1262 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1263 const char *fmt
, ...)
1268 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1272 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1275 * Set up arg0 for 'ps' etc
1277 __va_start(ap
, fmt
);
1278 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1282 * Schedule the thread to run
1284 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1287 td
->td_flags
&= ~TDF_STOPREQ
;
1292 * Destroy an LWKT thread. Warning! This function is not called when
1293 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1294 * uses a different reaping mechanism.
1299 thread_t td
= curthread
;
1303 if (td
->td_flags
& TDF_VERBOSE
)
1304 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1308 * Get us into a critical section to interlock gd_freetd and loop
1309 * until we can get it freed.
1311 * We have to cache the current td in gd_freetd because objcache_put()ing
1312 * it would rip it out from under us while our thread is still active.
1315 crit_enter_quick(td
);
1316 while ((std
= gd
->gd_freetd
) != NULL
) {
1317 gd
->gd_freetd
= NULL
;
1318 objcache_put(thread_cache
, std
);
1320 lwkt_deschedule_self(td
);
1321 lwkt_remove_tdallq(td
);
1322 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1328 lwkt_remove_tdallq(thread_t td
)
1330 KKASSERT(td
->td_gd
== mycpu
);
1331 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1337 thread_t td
= curthread
;
1338 int lpri
= td
->td_pri
;
1341 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1347 * Called from debugger/panic on cpus which have been stopped. We must still
1348 * process the IPIQ while stopped, even if we were stopped while in a critical
1351 * If we are dumping also try to process any pending interrupts. This may
1352 * or may not work depending on the state of the cpu at the point it was
1356 lwkt_smp_stopped(void)
1358 globaldata_t gd
= mycpu
;
1362 lwkt_process_ipiq();
1365 lwkt_process_ipiq();
1371 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1372 * get_mplock() has already incremented td_mpcount. We must block and
1373 * not return until giant is held.
1375 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1376 * reschedule the thread until it can obtain the giant lock for it.
1379 lwkt_mp_lock_contested(void)