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 1.106 2007/02/18 16:16:11 corecode 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.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
50 #include <sys/rtprio.h>
51 #include <sys/queue.h>
52 #include <sys/sysctl.h>
53 #include <sys/kthread.h>
54 #include <machine/cpu.h>
57 #include <sys/spinlock.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_zone.h>
73 #include <machine/stdarg.h>
74 #include <machine/smp.h>
78 #include <sys/stdint.h>
79 #include <libcaps/thread.h>
80 #include <sys/thread.h>
81 #include <sys/msgport.h>
82 #include <sys/errno.h>
83 #include <libcaps/globaldata.h>
84 #include <machine/cpufunc.h>
85 #include <sys/thread2.h>
86 #include <sys/msgport2.h>
90 #include <machine/lock.h>
91 #include <machine/atomic.h>
92 #include <machine/cpu.h>
96 static int untimely_switch
= 0;
98 static int panic_on_cscount
= 0;
100 static __int64_t switch_count
= 0;
101 static __int64_t preempt_hit
= 0;
102 static __int64_t preempt_miss
= 0;
103 static __int64_t preempt_weird
= 0;
104 static __int64_t token_contention_count
= 0;
105 static __int64_t mplock_contention_count
= 0;
109 SYSCTL_INT(_lwkt
, OID_AUTO
, untimely_switch
, CTLFLAG_RW
, &untimely_switch
, 0, "");
111 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
113 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
114 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
115 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
116 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
118 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
119 &token_contention_count
, 0, "spinning due to token contention");
120 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
121 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
130 #if !defined(KTR_GIANT_CONTENTION)
131 #define KTR_GIANT_CONTENTION KTR_ALL
134 KTR_INFO_MASTER(giant
);
135 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
136 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
138 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
143 * These helper procedures handle the runq, they can only be called from
144 * within a critical section.
146 * WARNING! Prior to SMP being brought up it is possible to enqueue and
147 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
148 * instead of 'mycpu' when referencing the globaldata structure. Once
149 * SMP live enqueuing and dequeueing only occurs on the current cpu.
153 _lwkt_dequeue(thread_t td
)
155 if (td
->td_flags
& TDF_RUNQ
) {
156 int nq
= td
->td_pri
& TDPRI_MASK
;
157 struct globaldata
*gd
= td
->td_gd
;
159 td
->td_flags
&= ~TDF_RUNQ
;
160 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
161 /* runqmask is passively cleaned up by the switcher */
167 _lwkt_enqueue(thread_t td
)
169 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_TSLEEPQ
|TDF_BLOCKQ
)) == 0) {
170 int nq
= td
->td_pri
& TDPRI_MASK
;
171 struct globaldata
*gd
= td
->td_gd
;
173 td
->td_flags
|= TDF_RUNQ
;
174 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
175 gd
->gd_runqmask
|= 1 << nq
;
180 * Schedule a thread to run. As the current thread we can always safely
181 * schedule ourselves, and a shortcut procedure is provided for that
184 * (non-blocking, self contained on a per cpu basis)
187 lwkt_schedule_self(thread_t td
)
189 crit_enter_quick(td
);
190 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
191 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
197 * Deschedule a thread.
199 * (non-blocking, self contained on a per cpu basis)
202 lwkt_deschedule_self(thread_t td
)
204 crit_enter_quick(td
);
212 * LWKTs operate on a per-cpu basis
214 * WARNING! Called from early boot, 'mycpu' may not work yet.
217 lwkt_gdinit(struct globaldata
*gd
)
221 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
222 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
224 TAILQ_INIT(&gd
->gd_tdallq
);
230 * Create a new thread. The thread must be associated with a process context
231 * or LWKT start address before it can be scheduled. If the target cpu is
232 * -1 the thread will be created on the current cpu.
234 * If you intend to create a thread without a process context this function
235 * does everything except load the startup and switcher function.
238 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
241 globaldata_t gd
= mycpu
;
245 if (gd
->gd_tdfreecount
> 0) {
246 --gd
->gd_tdfreecount
;
247 td
= TAILQ_FIRST(&gd
->gd_tdfreeq
);
248 KASSERT(td
!= NULL
&& (td
->td_flags
& TDF_RUNNING
) == 0,
249 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
250 TAILQ_REMOVE(&gd
->gd_tdfreeq
, td
, td_threadq
);
252 flags
|= td
->td_flags
& (TDF_ALLOCATED_STACK
|TDF_ALLOCATED_THREAD
);
256 td
= zalloc(thread_zone
);
258 td
= malloc(sizeof(struct thread
));
260 td
->td_kstack
= NULL
;
261 td
->td_kstack_size
= 0;
262 flags
|= TDF_ALLOCATED_THREAD
;
265 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
266 if (flags
& TDF_ALLOCATED_STACK
) {
268 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
270 libcaps_free_stack(stack
, td
->td_kstack_size
);
277 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
279 stack
= libcaps_alloc_stack(stksize
);
281 flags
|= TDF_ALLOCATED_STACK
;
284 lwkt_init_thread(td
, stack
, stksize
, flags
, mycpu
);
286 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
293 * Initialize a preexisting thread structure. This function is used by
294 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
296 * All threads start out in a critical section at a priority of
297 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
298 * appropriate. This function may send an IPI message when the
299 * requested cpu is not the current cpu and consequently gd_tdallq may
300 * not be initialized synchronously from the point of view of the originating
303 * NOTE! we have to be careful in regards to creating threads for other cpus
304 * if SMP has not yet been activated.
309 lwkt_init_thread_remote(void *arg
)
314 * Protected by critical section held by IPI dispatch
316 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
322 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
323 struct globaldata
*gd
)
325 globaldata_t mygd
= mycpu
;
327 bzero(td
, sizeof(struct thread
));
328 td
->td_kstack
= stack
;
329 td
->td_kstack_size
= stksize
;
330 td
->td_flags
= flags
;
332 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
334 if ((flags
& TDF_MPSAFE
) == 0)
337 lwkt_initport(&td
->td_msgport
, td
);
338 pmap_init_thread(td
);
341 * Normally initializing a thread for a remote cpu requires sending an
342 * IPI. However, the idlethread is setup before the other cpus are
343 * activated so we have to treat it as a special case. XXX manipulation
344 * of gd_tdallq requires the BGL.
346 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
348 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
351 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
355 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);
388 lwkt_wait_free(thread_t td
)
391 tsleep(td
, 0, "tdreap", hz
);
397 lwkt_free_thread(thread_t td
)
399 struct globaldata
*gd
= mycpu
;
401 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
402 ("lwkt_free_thread: did not exit! %p", td
));
405 if (gd
->gd_tdfreecount
< CACHE_NTHREADS
&&
406 (td
->td_flags
& TDF_ALLOCATED_THREAD
)
408 ++gd
->gd_tdfreecount
;
409 TAILQ_INSERT_HEAD(&gd
->gd_tdfreeq
, td
, td_threadq
);
413 if (td
->td_kstack
&& (td
->td_flags
& TDF_ALLOCATED_STACK
)) {
415 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
417 libcaps_free_stack(td
->td_kstack
, td
->td_kstack_size
);
420 td
->td_kstack
= NULL
;
421 td
->td_kstack_size
= 0;
423 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
425 zfree(thread_zone
, td
);
435 * Switch to the next runnable lwkt. If no LWKTs are runnable then
436 * switch to the idlethread. Switching must occur within a critical
437 * section to avoid races with the scheduling queue.
439 * We always have full control over our cpu's run queue. Other cpus
440 * that wish to manipulate our queue must use the cpu_*msg() calls to
441 * talk to our cpu, so a critical section is all that is needed and
442 * the result is very, very fast thread switching.
444 * The LWKT scheduler uses a fixed priority model and round-robins at
445 * each priority level. User process scheduling is a totally
446 * different beast and LWKT priorities should not be confused with
447 * user process priorities.
449 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
450 * cleans it up. Note that the td_switch() function cannot do anything that
451 * requires the MP lock since the MP lock will have already been setup for
452 * the target thread (not the current thread). It's nice to have a scheduler
453 * that does not need the MP lock to work because it allows us to do some
454 * really cool high-performance MP lock optimizations.
456 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
457 * is not called by the current thread in the preemption case, only when
458 * the preempting thread blocks (in order to return to the original thread).
463 globaldata_t gd
= mycpu
;
464 thread_t td
= gd
->gd_curthread
;
471 * Switching from within a 'fast' (non thread switched) interrupt or IPI
472 * is illegal. However, we may have to do it anyway if we hit a fatal
473 * kernel trap or we have paniced.
475 * If this case occurs save and restore the interrupt nesting level.
477 if (gd
->gd_intr_nesting_level
) {
481 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
482 panic("lwkt_switch: cannot switch from within "
483 "a fast interrupt, yet, td %p\n", td
);
485 savegdnest
= gd
->gd_intr_nesting_level
;
486 savegdtrap
= gd
->gd_trap_nesting_level
;
487 gd
->gd_intr_nesting_level
= 0;
488 gd
->gd_trap_nesting_level
= 0;
489 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
490 td
->td_flags
|= TDF_PANICWARN
;
491 kprintf("Warning: thread switch from interrupt or IPI, "
492 "thread %p (%s)\n", td
, td
->td_comm
);
494 db_print_backtrace();
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.
519 lwkt_relalltokens(td
);
523 * We had better not be holding any spin locks, but don't get into an
524 * endless panic loop.
526 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
527 ("lwkt_switch: still holding a shared spinlock %p!",
528 gd
->gd_spinlock_rd
));
529 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
530 ("lwkt_switch: still holding %d exclusive spinlocks!",
531 gd
->gd_spinlocks_wr
));
536 * td_mpcount cannot be used to determine if we currently hold the
537 * MP lock because get_mplock() will increment it prior to attempting
538 * to get the lock, and switch out if it can't. Our ownership of
539 * the actual lock will remain stable while we are in a critical section
540 * (but, of course, another cpu may own or release the lock so the
541 * actual value of mp_lock is not stable).
543 mpheld
= MP_LOCK_HELD();
545 if (td
->td_cscount
) {
546 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
548 if (panic_on_cscount
)
549 panic("switching while mastering cpusync");
553 if ((ntd
= td
->td_preempted
) != NULL
) {
555 * We had preempted another thread on this cpu, resume the preempted
556 * thread. This occurs transparently, whether the preempted thread
557 * was scheduled or not (it may have been preempted after descheduling
560 * We have to setup the MP lock for the original thread after backing
561 * out the adjustment that was made to curthread when the original
564 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
566 if (ntd
->td_mpcount
&& mpheld
== 0) {
567 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
568 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
570 if (ntd
->td_mpcount
) {
571 td
->td_mpcount
-= ntd
->td_mpcount
;
572 KKASSERT(td
->td_mpcount
>= 0);
575 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
578 * XXX. The interrupt may have woken a thread up, we need to properly
579 * set the reschedule flag if the originally interrupted thread is at
582 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
584 /* YYY release mp lock on switchback if original doesn't need it */
587 * Priority queue / round-robin at each priority. Note that user
588 * processes run at a fixed, low priority and the user process
589 * scheduler deals with interactions between user processes
590 * by scheduling and descheduling them from the LWKT queue as
593 * We have to adjust the MP lock for the target thread. If we
594 * need the MP lock and cannot obtain it we try to locate a
595 * thread that does not need the MP lock. If we cannot, we spin
598 * A similar issue exists for the tokens held by the target thread.
599 * If we cannot obtain ownership of the tokens we cannot immediately
600 * schedule the thread.
604 * If an LWKT reschedule was requested, well that is what we are
605 * doing now so clear it.
607 clear_lwkt_resched();
609 if (gd
->gd_runqmask
) {
610 int nq
= bsrl(gd
->gd_runqmask
);
611 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
612 gd
->gd_runqmask
&= ~(1 << nq
);
617 * THREAD SELECTION FOR AN SMP MACHINE BUILD
619 * If the target needs the MP lock and we couldn't get it,
620 * or if the target is holding tokens and we could not
621 * gain ownership of the tokens, continue looking for a
622 * thread to schedule and spin instead of HLT if we can't.
624 * NOTE: the mpheld variable invalid after this conditional, it
625 * can change due to both cpu_try_mplock() returning success
626 * AND interactions in lwkt_getalltokens() due to the fact that
627 * we are trying to check the mpcount of a thread other then
628 * the current thread. Because of this, if the current thread
629 * is not holding td_mpcount, an IPI indirectly run via
630 * lwkt_getalltokens() can obtain and release the MP lock and
631 * cause the core MP lock to be released.
633 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
634 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
636 u_int32_t rqmask
= gd
->gd_runqmask
;
638 mpheld
= MP_LOCK_HELD();
641 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
642 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
643 /* spinning due to MP lock being held */
645 ++mplock_contention_count
;
647 /* mplock still not held, 'mpheld' still valid */
652 * mpheld state invalid after getalltokens call returns
653 * failure, but the variable is only needed for
656 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
657 /* spinning due to token contention */
659 ++token_contention_count
;
661 mpheld
= MP_LOCK_HELD();
668 rqmask
&= ~(1 << nq
);
672 ntd
= &gd
->gd_idlethread
;
673 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
674 goto using_idle_thread
;
676 ++gd
->gd_cnt
.v_swtch
;
677 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
678 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
681 ++gd
->gd_cnt
.v_swtch
;
682 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
683 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
687 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
688 * worry about tokens or the BGL.
690 ++gd
->gd_cnt
.v_swtch
;
691 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
692 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
696 * We have nothing to run but only let the idle loop halt
697 * the cpu if there are no pending interrupts.
699 ntd
= &gd
->gd_idlethread
;
700 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
701 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
705 * The idle thread should not be holding the MP lock unless we
706 * are trapping in the kernel or in a panic. Since we select the
707 * idle thread unconditionally when no other thread is available,
708 * if the MP lock is desired during a panic or kernel trap, we
709 * have to loop in the scheduler until we get it.
711 if (ntd
->td_mpcount
) {
712 mpheld
= MP_LOCK_HELD();
713 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
)
714 panic("Idle thread %p was holding the BGL!", ntd
);
715 else if (mpheld
== 0)
721 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
722 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
725 * Do the actual switch. If the new target does not need the MP lock
726 * and we are holding it, release the MP lock. If the new target requires
727 * the MP lock we have already acquired it for the target.
730 if (ntd
->td_mpcount
== 0 ) {
734 ASSERT_MP_LOCK_HELD(ntd
);
741 /* NOTE: current cpu may have changed after switch */
746 * Request that the target thread preempt the current thread. Preemption
747 * only works under a specific set of conditions:
749 * - We are not preempting ourselves
750 * - The target thread is owned by the current cpu
751 * - We are not currently being preempted
752 * - The target is not currently being preempted
753 * - We are able to satisfy the target's MP lock requirements (if any).
755 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
756 * this is called via lwkt_schedule() through the td_preemptable callback.
757 * critpri is the managed critical priority that we should ignore in order
758 * to determine whether preemption is possible (aka usually just the crit
759 * priority of lwkt_schedule() itself).
761 * XXX at the moment we run the target thread in a critical section during
762 * the preemption in order to prevent the target from taking interrupts
763 * that *WE* can't. Preemption is strictly limited to interrupt threads
764 * and interrupt-like threads, outside of a critical section, and the
765 * preempted source thread will be resumed the instant the target blocks
766 * whether or not the source is scheduled (i.e. preemption is supposed to
767 * be as transparent as possible).
769 * The target thread inherits our MP count (added to its own) for the
770 * duration of the preemption in order to preserve the atomicy of the
771 * MP lock during the preemption. Therefore, any preempting targets must be
772 * careful in regards to MP assertions. Note that the MP count may be
773 * out of sync with the physical mp_lock, but we do not have to preserve
774 * the original ownership of the lock if it was out of synch (that is, we
775 * can leave it synchronized on return).
778 lwkt_preempt(thread_t ntd
, int critpri
)
780 struct globaldata
*gd
= mycpu
;
788 * The caller has put us in a critical section. We can only preempt
789 * if the caller of the caller was not in a critical section (basically
790 * a local interrupt), as determined by the 'critpri' parameter. We
791 * also acn't preempt if the caller is holding any spinlocks (even if
792 * he isn't in a critical section). This also handles the tokens test.
794 * YYY The target thread must be in a critical section (else it must
795 * inherit our critical section? I dunno yet).
797 * Set need_lwkt_resched() unconditionally for now YYY.
799 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
801 td
= gd
->gd_curthread
;
802 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
806 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
812 if (ntd
->td_gd
!= gd
) {
819 * Take the easy way out and do not preempt if the target is holding
820 * any spinlocks. We could test whether the thread(s) being
821 * preempted interlock against the target thread's tokens and whether
822 * we can get all the target thread's tokens, but this situation
823 * should not occur very often so its easier to simply not preempt.
824 * Also, plain spinlocks are impossible to figure out at this point so
825 * just don't preempt.
827 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
832 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
837 if (ntd
->td_preempted
) {
844 * note: an interrupt might have occured just as we were transitioning
845 * to or from the MP lock. In this case td_mpcount will be pre-disposed
846 * (non-zero) but not actually synchronized with the actual state of the
847 * lock. We can use it to imply an MP lock requirement for the
848 * preemption but we cannot use it to test whether we hold the MP lock
851 savecnt
= td
->td_mpcount
;
852 mpheld
= MP_LOCK_HELD();
853 ntd
->td_mpcount
+= td
->td_mpcount
;
854 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
855 ntd
->td_mpcount
-= td
->td_mpcount
;
863 * Since we are able to preempt the current thread, there is no need to
864 * call need_lwkt_resched().
867 ntd
->td_preempted
= td
;
868 td
->td_flags
|= TDF_PREEMPT_LOCK
;
870 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
872 KKASSERT(savecnt
== td
->td_mpcount
);
873 mpheld
= MP_LOCK_HELD();
874 if (mpheld
&& td
->td_mpcount
== 0)
876 else if (mpheld
== 0 && td
->td_mpcount
)
877 panic("lwkt_preempt(): MP lock was not held through");
879 ntd
->td_preempted
= NULL
;
880 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
884 * Yield our thread while higher priority threads are pending. This is
885 * typically called when we leave a critical section but it can be safely
886 * called while we are in a critical section.
888 * This function will not generally yield to equal priority threads but it
889 * can occur as a side effect. Note that lwkt_switch() is called from
890 * inside the critical section to prevent its own crit_exit() from reentering
891 * lwkt_yield_quick().
893 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
894 * came along but was blocked and made pending.
896 * (self contained on a per cpu basis)
899 lwkt_yield_quick(void)
901 globaldata_t gd
= mycpu
;
902 thread_t td
= gd
->gd_curthread
;
905 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
906 * it with a non-zero cpl then we might not wind up calling splz after
907 * a task switch when the critical section is exited even though the
908 * new task could accept the interrupt.
910 * XXX from crit_exit() only called after last crit section is released.
911 * If called directly will run splz() even if in a critical section.
913 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
914 * except for this special case, we MUST call splz() here to handle any
915 * pending ints, particularly after we switch, or we might accidently
916 * halt the cpu with interrupts pending.
918 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
922 * YYY enabling will cause wakeup() to task-switch, which really
923 * confused the old 4.x code. This is a good way to simulate
924 * preemption and MP without actually doing preemption or MP, because a
925 * lot of code assumes that wakeup() does not block.
927 if (untimely_switch
&& td
->td_nest_count
== 0 &&
928 gd
->gd_intr_nesting_level
== 0
930 crit_enter_quick(td
);
932 * YYY temporary hacks until we disassociate the userland scheduler
933 * from the LWKT scheduler.
935 if (td
->td_flags
& TDF_RUNQ
) {
936 lwkt_switch(); /* will not reenter yield function */
938 lwkt_schedule_self(td
); /* make sure we are scheduled */
939 lwkt_switch(); /* will not reenter yield function */
940 lwkt_deschedule_self(td
); /* make sure we are descheduled */
942 crit_exit_noyield(td
);
947 * This implements a normal yield which, unlike _quick, will yield to equal
948 * priority threads as well. Note that gd_reqflags tests will be handled by
949 * the crit_exit() call in lwkt_switch().
951 * (self contained on a per cpu basis)
956 lwkt_schedule_self(curthread
);
961 * Generic schedule. Possibly schedule threads belonging to other cpus and
962 * deal with threads that might be blocked on a wait queue.
964 * We have a little helper inline function which does additional work after
965 * the thread has been enqueued, including dealing with preemption and
966 * setting need_lwkt_resched() (which prevents the kernel from returning
967 * to userland until it has processed higher priority threads).
969 * It is possible for this routine to be called after a failed _enqueue
970 * (due to the target thread migrating, sleeping, or otherwise blocked).
971 * We have to check that the thread is actually on the run queue!
975 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
)
977 if (ntd
->td_flags
& TDF_RUNQ
) {
978 if (ntd
->td_preemptable
) {
979 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
980 } else if ((ntd
->td_flags
& TDF_NORESCHED
) == 0 &&
981 (ntd
->td_pri
& TDPRI_MASK
) > (gd
->gd_curthread
->td_pri
& TDPRI_MASK
)
989 lwkt_schedule(thread_t td
)
991 globaldata_t mygd
= mycpu
;
993 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
995 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
996 if (td
== mygd
->gd_curthread
) {
1000 * If we own the thread, there is no race (since we are in a
1001 * critical section). If we do not own the thread there might
1002 * be a race but the target cpu will deal with it.
1005 if (td
->td_gd
== mygd
) {
1007 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
);
1009 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_schedule
, td
);
1013 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
);
1022 * Thread migration using a 'Pull' method. The thread may or may not be
1023 * the current thread. It MUST be descheduled and in a stable state.
1024 * lwkt_giveaway() must be called on the cpu owning the thread.
1026 * At any point after lwkt_giveaway() is called, the target cpu may
1027 * 'pull' the thread by calling lwkt_acquire().
1029 * MPSAFE - must be called under very specific conditions.
1032 lwkt_giveaway(thread_t td
)
1034 globaldata_t gd
= mycpu
;
1037 KKASSERT(td
->td_gd
== gd
);
1038 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1039 td
->td_flags
|= TDF_MIGRATING
;
1044 lwkt_acquire(thread_t td
)
1049 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1054 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1055 crit_enter_gd(mygd
);
1056 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
))
1059 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1060 td
->td_flags
&= ~TDF_MIGRATING
;
1063 crit_enter_gd(mygd
);
1064 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1065 td
->td_flags
&= ~TDF_MIGRATING
;
1073 * Generic deschedule. Descheduling threads other then your own should be
1074 * done only in carefully controlled circumstances. Descheduling is
1077 * This function may block if the cpu has run out of messages.
1080 lwkt_deschedule(thread_t td
)
1084 if (td
== curthread
) {
1087 if (td
->td_gd
== mycpu
) {
1090 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1100 * Set the target thread's priority. This routine does not automatically
1101 * switch to a higher priority thread, LWKT threads are not designed for
1102 * continuous priority changes. Yield if you want to switch.
1104 * We have to retain the critical section count which uses the high bits
1105 * of the td_pri field. The specified priority may also indicate zero or
1106 * more critical sections by adding TDPRI_CRIT*N.
1108 * Note that we requeue the thread whether it winds up on a different runq
1109 * or not. uio_yield() depends on this and the routine is not normally
1110 * called with the same priority otherwise.
1113 lwkt_setpri(thread_t td
, int pri
)
1116 KKASSERT(td
->td_gd
== mycpu
);
1118 if (td
->td_flags
& TDF_RUNQ
) {
1120 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1123 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1129 lwkt_setpri_self(int pri
)
1131 thread_t td
= curthread
;
1133 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1135 if (td
->td_flags
& TDF_RUNQ
) {
1137 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1140 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1146 * Determine if there is a runnable thread at a higher priority then
1147 * the current thread. lwkt_setpri() does not check this automatically.
1148 * Return 1 if there is, 0 if there isn't.
1150 * Example: if bit 31 of runqmask is set and the current thread is priority
1151 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1153 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1154 * up comparing against 0xffffffff, a comparison that will always be false.
1157 lwkt_checkpri_self(void)
1159 globaldata_t gd
= mycpu
;
1160 thread_t td
= gd
->gd_curthread
;
1161 int nq
= td
->td_pri
& TDPRI_MASK
;
1163 while (gd
->gd_runqmask
> (__uint32_t
)(2 << nq
) - 1) {
1164 if (TAILQ_FIRST(&gd
->gd_tdrunq
[nq
+ 1]))
1172 * Migrate the current thread to the specified cpu.
1174 * This is accomplished by descheduling ourselves from the current cpu,
1175 * moving our thread to the tdallq of the target cpu, IPI messaging the
1176 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1177 * races while the thread is being migrated.
1180 static void lwkt_setcpu_remote(void *arg
);
1184 lwkt_setcpu_self(globaldata_t rgd
)
1187 thread_t td
= curthread
;
1189 if (td
->td_gd
!= rgd
) {
1190 crit_enter_quick(td
);
1191 td
->td_flags
|= TDF_MIGRATING
;
1192 lwkt_deschedule_self(td
);
1193 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1194 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1196 /* we are now on the target cpu */
1197 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1198 crit_exit_quick(td
);
1204 lwkt_migratecpu(int cpuid
)
1209 rgd
= globaldata_find(cpuid
);
1210 lwkt_setcpu_self(rgd
);
1215 * Remote IPI for cpu migration (called while in a critical section so we
1216 * do not have to enter another one). The thread has already been moved to
1217 * our cpu's allq, but we must wait for the thread to be completely switched
1218 * out on the originating cpu before we schedule it on ours or the stack
1219 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1220 * change to main memory.
1222 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1223 * against wakeups. It is best if this interface is used only when there
1224 * are no pending events that might try to schedule the thread.
1228 lwkt_setcpu_remote(void *arg
)
1231 globaldata_t gd
= mycpu
;
1233 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
))
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
,
1269 tdflags
| TDF_VERBOSE
);
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 * kthread_* is specific to the kernel and is not needed by userland.
1297 * Destroy an LWKT thread. Warning! This function is not called when
1298 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1299 * uses a different reaping mechanism.
1304 thread_t td
= curthread
;
1307 if (td
->td_flags
& TDF_VERBOSE
)
1308 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1310 crit_enter_quick(td
);
1311 lwkt_deschedule_self(td
);
1313 lwkt_remove_tdallq(td
);
1314 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
1315 ++gd
->gd_tdfreecount
;
1316 TAILQ_INSERT_TAIL(&gd
->gd_tdfreeq
, td
, td_threadq
);
1322 lwkt_remove_tdallq(thread_t td
)
1324 KKASSERT(td
->td_gd
== mycpu
);
1325 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1328 #endif /* _KERNEL */
1333 thread_t td
= curthread
;
1334 int lpri
= td
->td_pri
;
1337 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1343 * Called from debugger/panic on cpus which have been stopped. We must still
1344 * process the IPIQ while stopped, even if we were stopped while in a critical
1347 * If we are dumping also try to process any pending interrupts. This may
1348 * or may not work depending on the state of the cpu at the point it was
1352 lwkt_smp_stopped(void)
1354 globaldata_t gd
= mycpu
;
1358 lwkt_process_ipiq();
1361 lwkt_process_ipiq();
1367 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1368 * get_mplock() has already incremented td_mpcount. We must block and
1369 * not return until giant is held.
1371 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1372 * reschedule the thread until it can obtain the giant lock for it.
1375 lwkt_mp_lock_contested(void)