2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
60 #include <vm/vm_param.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_pager.h>
66 #include <vm/vm_extern.h>
68 #include <machine/stdarg.h>
69 #include <machine/smp.h>
72 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
75 static int mplock_countx
= 0;
78 static int panic_on_cscount
= 0;
80 static __int64_t switch_count
= 0;
81 static __int64_t preempt_hit
= 0;
82 static __int64_t preempt_miss
= 0;
83 static __int64_t preempt_weird
= 0;
84 static __int64_t token_contention_count
= 0;
85 static __int64_t mplock_contention_count
= 0;
86 static int lwkt_use_spin_port
;
88 static int chain_mplock
= 0;
89 static int bgl_yield
= 10;
91 static struct objcache
*thread_cache
;
93 volatile cpumask_t mp_lock_contention_mask
;
95 static void lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
);
97 extern void cpu_heavy_restore(void);
98 extern void cpu_lwkt_restore(void);
99 extern void cpu_kthread_restore(void);
100 extern void cpu_idle_restore(void);
105 jg_tos_ok(struct thread
*td
)
113 KKASSERT(td
->td_sp
!= NULL
);
114 tos
= ((void **)td
->td_sp
)[0];
116 if ((tos
== cpu_heavy_restore
) || (tos
== cpu_lwkt_restore
) ||
117 (tos
== cpu_kthread_restore
) || (tos
== cpu_idle_restore
)) {
126 * We can make all thread ports use the spin backend instead of the thread
127 * backend. This should only be set to debug the spin backend.
129 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
132 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
135 SYSCTL_INT(_lwkt
, OID_AUTO
, chain_mplock
, CTLFLAG_RW
, &chain_mplock
, 0, "");
136 SYSCTL_INT(_lwkt
, OID_AUTO
, bgl_yield_delay
, CTLFLAG_RW
, &bgl_yield
, 0, "");
138 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
139 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
140 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
141 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
143 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
144 &token_contention_count
, 0, "spinning due to token contention");
145 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
146 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
152 #if !defined(KTR_GIANT_CONTENTION)
153 #define KTR_GIANT_CONTENTION KTR_ALL
156 KTR_INFO_MASTER(giant
);
157 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
158 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
160 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
163 * These helper procedures handle the runq, they can only be called from
164 * within a critical section.
166 * WARNING! Prior to SMP being brought up it is possible to enqueue and
167 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
168 * instead of 'mycpu' when referencing the globaldata structure. Once
169 * SMP live enqueuing and dequeueing only occurs on the current cpu.
173 _lwkt_dequeue(thread_t td
)
175 if (td
->td_flags
& TDF_RUNQ
) {
176 int nq
= td
->td_pri
& TDPRI_MASK
;
177 struct globaldata
*gd
= td
->td_gd
;
179 td
->td_flags
&= ~TDF_RUNQ
;
180 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
181 /* runqmask is passively cleaned up by the switcher */
187 _lwkt_enqueue(thread_t td
)
189 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_BLOCKQ
)) == 0) {
190 int nq
= td
->td_pri
& TDPRI_MASK
;
191 struct globaldata
*gd
= td
->td_gd
;
193 td
->td_flags
|= TDF_RUNQ
;
194 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
195 gd
->gd_runqmask
|= 1 << nq
;
200 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
202 struct thread
*td
= (struct thread
*)obj
;
204 td
->td_kstack
= NULL
;
205 td
->td_kstack_size
= 0;
206 td
->td_flags
= TDF_ALLOCATED_THREAD
;
211 _lwkt_thread_dtor(void *obj
, void *privdata
)
213 struct thread
*td
= (struct thread
*)obj
;
215 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
216 ("_lwkt_thread_dtor: not allocated from objcache"));
217 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
218 td
->td_kstack_size
> 0,
219 ("_lwkt_thread_dtor: corrupted stack"));
220 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
224 * Initialize the lwkt s/system.
229 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
230 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
),
231 NULL
, CACHE_NTHREADS
/2,
232 _lwkt_thread_ctor
, _lwkt_thread_dtor
, NULL
);
236 * Schedule a thread to run. As the current thread we can always safely
237 * schedule ourselves, and a shortcut procedure is provided for that
240 * (non-blocking, self contained on a per cpu basis)
243 lwkt_schedule_self(thread_t td
)
245 crit_enter_quick(td
);
246 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
247 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
253 * Deschedule a thread.
255 * (non-blocking, self contained on a per cpu basis)
258 lwkt_deschedule_self(thread_t td
)
260 crit_enter_quick(td
);
266 * LWKTs operate on a per-cpu basis
268 * WARNING! Called from early boot, 'mycpu' may not work yet.
271 lwkt_gdinit(struct globaldata
*gd
)
275 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
276 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
278 TAILQ_INIT(&gd
->gd_tdallq
);
282 * Create a new thread. The thread must be associated with a process context
283 * or LWKT start address before it can be scheduled. If the target cpu is
284 * -1 the thread will be created on the current cpu.
286 * If you intend to create a thread without a process context this function
287 * does everything except load the startup and switcher function.
290 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
292 globaldata_t gd
= mycpu
;
296 * If static thread storage is not supplied allocate a thread. Reuse
297 * a cached free thread if possible. gd_freetd is used to keep an exiting
298 * thread intact through the exit.
301 if ((td
= gd
->gd_freetd
) != NULL
)
302 gd
->gd_freetd
= NULL
;
304 td
= objcache_get(thread_cache
, M_WAITOK
);
305 KASSERT((td
->td_flags
&
306 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
307 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
308 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
312 * Try to reuse cached stack.
314 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
315 if (flags
& TDF_ALLOCATED_STACK
) {
316 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
321 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
322 flags
|= TDF_ALLOCATED_STACK
;
325 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
327 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
332 * Initialize a preexisting thread structure. This function is used by
333 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
335 * All threads start out in a critical section at a priority of
336 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
337 * appropriate. This function may send an IPI message when the
338 * requested cpu is not the current cpu and consequently gd_tdallq may
339 * not be initialized synchronously from the point of view of the originating
342 * NOTE! we have to be careful in regards to creating threads for other cpus
343 * if SMP has not yet been activated.
348 lwkt_init_thread_remote(void *arg
)
353 * Protected by critical section held by IPI dispatch
355 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
361 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
362 struct globaldata
*gd
)
364 globaldata_t mygd
= mycpu
;
366 bzero(td
, sizeof(struct thread
));
367 td
->td_kstack
= stack
;
368 td
->td_kstack_size
= stksize
;
369 td
->td_flags
= flags
;
371 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
373 if ((flags
& TDF_MPSAFE
) == 0)
376 if (lwkt_use_spin_port
)
377 lwkt_initport_spin(&td
->td_msgport
);
379 lwkt_initport_thread(&td
->td_msgport
, td
);
380 pmap_init_thread(td
);
383 * Normally initializing a thread for a remote cpu requires sending an
384 * IPI. However, the idlethread is setup before the other cpus are
385 * activated so we have to treat it as a special case. XXX manipulation
386 * of gd_tdallq requires the BGL.
388 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
390 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
393 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
397 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
403 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
408 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
413 lwkt_hold(thread_t td
)
419 lwkt_rele(thread_t td
)
421 KKASSERT(td
->td_refs
> 0);
426 lwkt_wait_free(thread_t td
)
429 tsleep(td
, 0, "tdreap", hz
);
433 lwkt_free_thread(thread_t td
)
435 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
436 ("lwkt_free_thread: did not exit! %p", td
));
438 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
439 objcache_put(thread_cache
, td
);
440 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
441 /* client-allocated struct with internally allocated stack */
442 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
443 ("lwkt_free_thread: corrupted stack"));
444 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
445 td
->td_kstack
= NULL
;
446 td
->td_kstack_size
= 0;
452 * Switch to the next runnable lwkt. If no LWKTs are runnable then
453 * switch to the idlethread. Switching must occur within a critical
454 * section to avoid races with the scheduling queue.
456 * We always have full control over our cpu's run queue. Other cpus
457 * that wish to manipulate our queue must use the cpu_*msg() calls to
458 * talk to our cpu, so a critical section is all that is needed and
459 * the result is very, very fast thread switching.
461 * The LWKT scheduler uses a fixed priority model and round-robins at
462 * each priority level. User process scheduling is a totally
463 * different beast and LWKT priorities should not be confused with
464 * user process priorities.
466 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
467 * cleans it up. Note that the td_switch() function cannot do anything that
468 * requires the MP lock since the MP lock will have already been setup for
469 * the target thread (not the current thread). It's nice to have a scheduler
470 * that does not need the MP lock to work because it allows us to do some
471 * really cool high-performance MP lock optimizations.
473 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
474 * is not called by the current thread in the preemption case, only when
475 * the preempting thread blocks (in order to return to the original thread).
480 globaldata_t gd
= mycpu
;
481 thread_t td
= gd
->gd_curthread
;
488 * Switching from within a 'fast' (non thread switched) interrupt or IPI
489 * is illegal. However, we may have to do it anyway if we hit a fatal
490 * kernel trap or we have paniced.
492 * If this case occurs save and restore the interrupt nesting level.
494 if (gd
->gd_intr_nesting_level
) {
498 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
499 panic("lwkt_switch: cannot switch from within "
500 "a fast interrupt, yet, td %p\n", td
);
502 savegdnest
= gd
->gd_intr_nesting_level
;
503 savegdtrap
= gd
->gd_trap_nesting_level
;
504 gd
->gd_intr_nesting_level
= 0;
505 gd
->gd_trap_nesting_level
= 0;
506 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
507 td
->td_flags
|= TDF_PANICWARN
;
508 kprintf("Warning: thread switch from interrupt or IPI, "
509 "thread %p (%s)\n", td
, td
->td_comm
);
513 gd
->gd_intr_nesting_level
= savegdnest
;
514 gd
->gd_trap_nesting_level
= savegdtrap
;
520 * Passive release (used to transition from user to kernel mode
521 * when we block or switch rather then when we enter the kernel).
522 * This function is NOT called if we are switching into a preemption
523 * or returning from a preemption. Typically this causes us to lose
524 * our current process designation (if we have one) and become a true
525 * LWKT thread, and may also hand the current process designation to
526 * another process and schedule thread.
533 lwkt_relalltokens(td
);
536 * We had better not be holding any spin locks, but don't get into an
537 * endless panic loop.
539 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
540 ("lwkt_switch: still holding a shared spinlock %p!",
541 gd
->gd_spinlock_rd
));
542 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
543 ("lwkt_switch: still holding %d exclusive spinlocks!",
544 gd
->gd_spinlocks_wr
));
549 * td_mpcount cannot be used to determine if we currently hold the
550 * MP lock because get_mplock() will increment it prior to attempting
551 * to get the lock, and switch out if it can't. Our ownership of
552 * the actual lock will remain stable while we are in a critical section
553 * (but, of course, another cpu may own or release the lock so the
554 * actual value of mp_lock is not stable).
556 mpheld
= MP_LOCK_HELD();
558 if (td
->td_cscount
) {
559 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
561 if (panic_on_cscount
)
562 panic("switching while mastering cpusync");
566 if ((ntd
= td
->td_preempted
) != NULL
) {
568 * We had preempted another thread on this cpu, resume the preempted
569 * thread. This occurs transparently, whether the preempted thread
570 * was scheduled or not (it may have been preempted after descheduling
573 * We have to setup the MP lock for the original thread after backing
574 * out the adjustment that was made to curthread when the original
577 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
579 if (ntd
->td_mpcount
&& mpheld
== 0) {
580 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
581 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
583 if (ntd
->td_mpcount
) {
584 td
->td_mpcount
-= ntd
->td_mpcount
;
585 KKASSERT(td
->td_mpcount
>= 0);
588 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
591 * The interrupt may have woken a thread up, we need to properly
592 * set the reschedule flag if the originally interrupted thread is
593 * at a lower priority.
595 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
597 /* YYY release mp lock on switchback if original doesn't need it */
600 * Priority queue / round-robin at each priority. Note that user
601 * processes run at a fixed, low priority and the user process
602 * scheduler deals with interactions between user processes
603 * by scheduling and descheduling them from the LWKT queue as
606 * We have to adjust the MP lock for the target thread. If we
607 * need the MP lock and cannot obtain it we try to locate a
608 * thread that does not need the MP lock. If we cannot, we spin
611 * A similar issue exists for the tokens held by the target thread.
612 * If we cannot obtain ownership of the tokens we cannot immediately
613 * schedule the thread.
617 * If an LWKT reschedule was requested, well that is what we are
618 * doing now so clear it.
620 clear_lwkt_resched();
622 if (gd
->gd_runqmask
) {
623 int nq
= bsrl(gd
->gd_runqmask
);
624 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
625 gd
->gd_runqmask
&= ~(1 << nq
);
630 * THREAD SELECTION FOR AN SMP MACHINE BUILD
632 * If the target needs the MP lock and we couldn't get it,
633 * or if the target is holding tokens and we could not
634 * gain ownership of the tokens, continue looking for a
635 * thread to schedule and spin instead of HLT if we can't.
637 * NOTE: the mpheld variable invalid after this conditional, it
638 * can change due to both cpu_try_mplock() returning success
639 * AND interactions in lwkt_getalltokens() due to the fact that
640 * we are trying to check the mpcount of a thread other then
641 * the current thread. Because of this, if the current thread
642 * is not holding td_mpcount, an IPI indirectly run via
643 * lwkt_getalltokens() can obtain and release the MP lock and
644 * cause the core MP lock to be released.
646 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
647 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
649 u_int32_t rqmask
= gd
->gd_runqmask
;
651 mpheld
= MP_LOCK_HELD();
654 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
655 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
656 /* spinning due to MP lock being held */
658 ++mplock_contention_count
;
660 /* mplock still not held, 'mpheld' still valid */
665 * mpheld state invalid after getalltokens call returns
666 * failure, but the variable is only needed for
669 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
670 /* spinning due to token contention */
672 ++token_contention_count
;
674 mpheld
= MP_LOCK_HELD();
681 rqmask
&= ~(1 << nq
);
685 * We have two choices. We can either refuse to run a
686 * user thread when a kernel thread needs the MP lock
687 * but could not get it, or we can allow it to run but
688 * then expect an IPI (hopefully) later on to force a
689 * reschedule when the MP lock might become available.
691 if (nq
< TDPRI_KERN_LPSCHED
) {
692 if (chain_mplock
== 0)
694 atomic_set_int(&mp_lock_contention_mask
,
696 /* continue loop, allow user threads to be scheduled */
700 cpu_mplock_contested();
701 ntd
= &gd
->gd_idlethread
;
702 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
703 goto using_idle_thread
;
705 ++gd
->gd_cnt
.v_swtch
;
706 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
707 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
712 ++gd
->gd_cnt
.v_swtch
;
713 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
714 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
718 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
719 * worry about tokens or the BGL. However, we still have
720 * to call lwkt_getalltokens() in order to properly detect
721 * stale tokens. This call cannot fail for a UP build!
723 lwkt_getalltokens(ntd
);
724 ++gd
->gd_cnt
.v_swtch
;
725 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
726 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
730 * We have nothing to run but only let the idle loop halt
731 * the cpu if there are no pending interrupts.
733 ntd
= &gd
->gd_idlethread
;
734 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
735 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
739 * The idle thread should not be holding the MP lock unless we
740 * are trapping in the kernel or in a panic. Since we select the
741 * idle thread unconditionally when no other thread is available,
742 * if the MP lock is desired during a panic or kernel trap, we
743 * have to loop in the scheduler until we get it.
745 if (ntd
->td_mpcount
) {
746 mpheld
= MP_LOCK_HELD();
747 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
748 panic("Idle thread %p was holding the BGL!", ntd
);
749 } else if (mpheld
== 0) {
750 cpu_mplock_contested();
757 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
758 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
761 * Do the actual switch. If the new target does not need the MP lock
762 * and we are holding it, release the MP lock. If the new target requires
763 * the MP lock we have already acquired it for the target.
766 if (ntd
->td_mpcount
== 0 ) {
770 ASSERT_MP_LOCK_HELD(ntd
);
776 KKASSERT(jg_tos_ok(ntd
));
780 /* NOTE: current cpu may have changed after switch */
785 * Request that the target thread preempt the current thread. Preemption
786 * only works under a specific set of conditions:
788 * - We are not preempting ourselves
789 * - The target thread is owned by the current cpu
790 * - We are not currently being preempted
791 * - The target is not currently being preempted
792 * - We are not holding any spin locks
793 * - The target thread is not holding any tokens
794 * - We are able to satisfy the target's MP lock requirements (if any).
796 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
797 * this is called via lwkt_schedule() through the td_preemptable callback.
798 * critpri is the managed critical priority that we should ignore in order
799 * to determine whether preemption is possible (aka usually just the crit
800 * priority of lwkt_schedule() itself).
802 * XXX at the moment we run the target thread in a critical section during
803 * the preemption in order to prevent the target from taking interrupts
804 * that *WE* can't. Preemption is strictly limited to interrupt threads
805 * and interrupt-like threads, outside of a critical section, and the
806 * preempted source thread will be resumed the instant the target blocks
807 * whether or not the source is scheduled (i.e. preemption is supposed to
808 * be as transparent as possible).
810 * The target thread inherits our MP count (added to its own) for the
811 * duration of the preemption in order to preserve the atomicy of the
812 * MP lock during the preemption. Therefore, any preempting targets must be
813 * careful in regards to MP assertions. Note that the MP count may be
814 * out of sync with the physical mp_lock, but we do not have to preserve
815 * the original ownership of the lock if it was out of synch (that is, we
816 * can leave it synchronized on return).
819 lwkt_preempt(thread_t ntd
, int critpri
)
821 struct globaldata
*gd
= mycpu
;
829 * The caller has put us in a critical section. We can only preempt
830 * if the caller of the caller was not in a critical section (basically
831 * a local interrupt), as determined by the 'critpri' parameter. We
832 * also can't preempt if the caller is holding any spinlocks (even if
833 * he isn't in a critical section). This also handles the tokens test.
835 * YYY The target thread must be in a critical section (else it must
836 * inherit our critical section? I dunno yet).
838 * Set need_lwkt_resched() unconditionally for now YYY.
840 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
842 td
= gd
->gd_curthread
;
843 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
847 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
853 if (ntd
->td_gd
!= gd
) {
860 * Take the easy way out and do not preempt if we are holding
861 * any spinlocks. We could test whether the thread(s) being
862 * preempted interlock against the target thread's tokens and whether
863 * we can get all the target thread's tokens, but this situation
864 * should not occur very often so its easier to simply not preempt.
865 * Also, plain spinlocks are impossible to figure out at this point so
866 * just don't preempt.
868 * Do not try to preempt if the target thread is holding any tokens.
869 * We could try to acquire the tokens but this case is so rare there
870 * is no need to support it.
872 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
882 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
887 if (ntd
->td_preempted
) {
894 * note: an interrupt might have occured just as we were transitioning
895 * to or from the MP lock. In this case td_mpcount will be pre-disposed
896 * (non-zero) but not actually synchronized with the actual state of the
897 * lock. We can use it to imply an MP lock requirement for the
898 * preemption but we cannot use it to test whether we hold the MP lock
901 savecnt
= td
->td_mpcount
;
902 mpheld
= MP_LOCK_HELD();
903 ntd
->td_mpcount
+= td
->td_mpcount
;
904 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
905 ntd
->td_mpcount
-= td
->td_mpcount
;
913 * Since we are able to preempt the current thread, there is no need to
914 * call need_lwkt_resched().
917 ntd
->td_preempted
= td
;
918 td
->td_flags
|= TDF_PREEMPT_LOCK
;
921 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
923 KKASSERT(savecnt
== td
->td_mpcount
);
924 mpheld
= MP_LOCK_HELD();
925 if (mpheld
&& td
->td_mpcount
== 0)
927 else if (mpheld
== 0 && td
->td_mpcount
)
928 panic("lwkt_preempt(): MP lock was not held through");
930 ntd
->td_preempted
= NULL
;
931 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
935 * Conditionally call splz() if gd_reqflags indicates work is pending.
937 * td_nest_count prevents deep nesting via splz() or doreti() which
938 * might otherwise blow out the kernel stack. Note that except for
939 * this special case, we MUST call splz() here to handle any
940 * pending ints, particularly after we switch, or we might accidently
941 * halt the cpu with interrupts pending.
943 * (self contained on a per cpu basis)
948 globaldata_t gd
= mycpu
;
949 thread_t td
= gd
->gd_curthread
;
951 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
956 * This implements a normal yield which will yield to equal priority
957 * threads as well as higher priority threads. Note that gd_reqflags
958 * tests will be handled by the crit_exit() call in lwkt_switch().
960 * (self contained on a per cpu basis)
965 lwkt_schedule_self(curthread
);
970 * This function is used along with the lwkt_passive_recover() inline
971 * by the trap code to negotiate a passive release of the current
972 * process/lwp designation with the user scheduler.
975 lwkt_passive_release(struct thread
*td
)
977 struct lwp
*lp
= td
->td_lwp
;
979 td
->td_release
= NULL
;
980 lwkt_setpri_self(TDPRI_KERN_USER
);
981 lp
->lwp_proc
->p_usched
->release_curproc(lp
);
985 * Make a kernel thread act as if it were in user mode with regards
986 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
987 * loops which may be potentially cpu-bound can call lwkt_user_yield().
989 * The lwkt_user_yield() function is designed to have very low overhead
990 * if no yield is determined to be needed.
993 lwkt_user_yield(void)
995 thread_t td
= curthread
;
996 struct lwp
*lp
= td
->td_lwp
;
1000 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1001 * kernel can prevent other cpus from servicing interrupt threads
1002 * which still require the MP lock (which is a lot of them). This
1003 * has a chaining effect since if the interrupt is blocked, so is
1004 * the event, so normal scheduling will not pick up on the problem.
1006 if (mplock_countx
&& td
->td_mpcount
) {
1007 int savecnt
= td
->td_mpcount
;
1014 td
->td_mpcount
= savecnt
;
1019 * Another kernel thread wants the cpu
1021 if (lwkt_resched_wanted())
1025 * If the user scheduler has asynchronously determined that the current
1026 * process (when running in user mode) needs to lose the cpu then make
1027 * sure we are released.
1029 if (user_resched_wanted()) {
1035 * If we are released reduce our priority
1037 if (td
->td_release
== NULL
) {
1038 if (lwkt_check_resched(td
) > 0)
1041 lp
->lwp_proc
->p_usched
->acquire_curproc(lp
);
1042 td
->td_release
= lwkt_passive_release
;
1043 lwkt_setpri_self(TDPRI_USER_NORM
);
1049 * Return 0 if no runnable threads are pending at the same or higher
1050 * priority as the passed thread.
1052 * Return 1 if runnable threads are pending at the same priority.
1054 * Return 2 if runnable threads are pending at a higher priority.
1057 lwkt_check_resched(thread_t td
)
1059 int pri
= td
->td_pri
& TDPRI_MASK
;
1061 if (td
->td_gd
->gd_runqmask
> (2 << pri
) - 1)
1063 if (TAILQ_NEXT(td
, td_threadq
))
1069 * Generic schedule. Possibly schedule threads belonging to other cpus and
1070 * deal with threads that might be blocked on a wait queue.
1072 * We have a little helper inline function which does additional work after
1073 * the thread has been enqueued, including dealing with preemption and
1074 * setting need_lwkt_resched() (which prevents the kernel from returning
1075 * to userland until it has processed higher priority threads).
1077 * It is possible for this routine to be called after a failed _enqueue
1078 * (due to the target thread migrating, sleeping, or otherwise blocked).
1079 * We have to check that the thread is actually on the run queue!
1081 * reschedok is an optimized constant propagated from lwkt_schedule() or
1082 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1083 * reschedule to be requested if the target thread has a higher priority.
1084 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1085 * be 0, prevented undesired reschedules.
1089 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
, int reschedok
)
1093 if (ntd
->td_flags
& TDF_RUNQ
) {
1094 if (ntd
->td_preemptable
&& reschedok
) {
1095 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
1096 } else if (reschedok
) {
1098 if ((ntd
->td_pri
& TDPRI_MASK
) > (otd
->td_pri
& TDPRI_MASK
))
1099 need_lwkt_resched();
1106 _lwkt_schedule(thread_t td
, int reschedok
)
1108 globaldata_t mygd
= mycpu
;
1110 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1111 crit_enter_gd(mygd
);
1112 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1113 if (td
== mygd
->gd_curthread
) {
1117 * If we own the thread, there is no race (since we are in a
1118 * critical section). If we do not own the thread there might
1119 * be a race but the target cpu will deal with it.
1122 if (td
->td_gd
== mygd
) {
1124 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1126 lwkt_send_ipiq3(td
->td_gd
, lwkt_schedule_remote
, td
, 0);
1130 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1137 lwkt_schedule(thread_t td
)
1139 _lwkt_schedule(td
, 1);
1143 lwkt_schedule_noresched(thread_t td
)
1145 _lwkt_schedule(td
, 0);
1149 * When scheduled remotely if frame != NULL the IPIQ is being
1150 * run via doreti or an interrupt then preemption can be allowed.
1152 * To allow preemption we have to drop the critical section so only
1153 * one is present in _lwkt_schedule_post.
1156 lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
)
1158 thread_t td
= curthread
;
1161 if (frame
&& ntd
->td_preemptable
) {
1162 crit_exit_noyield(td
);
1163 _lwkt_schedule(ntd
, 1);
1164 crit_enter_quick(td
);
1166 _lwkt_schedule(ntd
, 1);
1173 * Thread migration using a 'Pull' method. The thread may or may not be
1174 * the current thread. It MUST be descheduled and in a stable state.
1175 * lwkt_giveaway() must be called on the cpu owning the thread.
1177 * At any point after lwkt_giveaway() is called, the target cpu may
1178 * 'pull' the thread by calling lwkt_acquire().
1180 * We have to make sure the thread is not sitting on a per-cpu tsleep
1181 * queue or it will blow up when it moves to another cpu.
1183 * MPSAFE - must be called under very specific conditions.
1186 lwkt_giveaway(thread_t td
)
1188 globaldata_t gd
= mycpu
;
1191 if (td
->td_flags
& TDF_TSLEEPQ
)
1193 KKASSERT(td
->td_gd
== gd
);
1194 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1195 td
->td_flags
|= TDF_MIGRATING
;
1200 lwkt_acquire(thread_t td
)
1205 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1210 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1211 crit_enter_gd(mygd
);
1212 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1214 lwkt_process_ipiq();
1219 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1220 td
->td_flags
&= ~TDF_MIGRATING
;
1223 crit_enter_gd(mygd
);
1224 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1225 td
->td_flags
&= ~TDF_MIGRATING
;
1233 * Generic deschedule. Descheduling threads other then your own should be
1234 * done only in carefully controlled circumstances. Descheduling is
1237 * This function may block if the cpu has run out of messages.
1240 lwkt_deschedule(thread_t td
)
1244 if (td
== curthread
) {
1247 if (td
->td_gd
== mycpu
) {
1250 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1260 * Set the target thread's priority. This routine does not automatically
1261 * switch to a higher priority thread, LWKT threads are not designed for
1262 * continuous priority changes. Yield if you want to switch.
1264 * We have to retain the critical section count which uses the high bits
1265 * of the td_pri field. The specified priority may also indicate zero or
1266 * more critical sections by adding TDPRI_CRIT*N.
1268 * Note that we requeue the thread whether it winds up on a different runq
1269 * or not. uio_yield() depends on this and the routine is not normally
1270 * called with the same priority otherwise.
1273 lwkt_setpri(thread_t td
, int pri
)
1276 KKASSERT(td
->td_gd
== mycpu
);
1278 if (td
->td_flags
& TDF_RUNQ
) {
1280 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1283 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1289 lwkt_setpri_self(int pri
)
1291 thread_t td
= curthread
;
1293 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1295 if (td
->td_flags
& TDF_RUNQ
) {
1297 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1300 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1306 * Migrate the current thread to the specified cpu.
1308 * This is accomplished by descheduling ourselves from the current cpu,
1309 * moving our thread to the tdallq of the target cpu, IPI messaging the
1310 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1311 * races while the thread is being migrated.
1313 * We must be sure to remove ourselves from the current cpu's tsleepq
1314 * before potentially moving to another queue. The thread can be on
1315 * a tsleepq due to a left-over tsleep_interlock().
1318 static void lwkt_setcpu_remote(void *arg
);
1322 lwkt_setcpu_self(globaldata_t rgd
)
1325 thread_t td
= curthread
;
1327 if (td
->td_gd
!= rgd
) {
1328 crit_enter_quick(td
);
1329 if (td
->td_flags
& TDF_TSLEEPQ
)
1331 td
->td_flags
|= TDF_MIGRATING
;
1332 lwkt_deschedule_self(td
);
1333 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1334 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1336 /* we are now on the target cpu */
1337 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1338 crit_exit_quick(td
);
1344 lwkt_migratecpu(int cpuid
)
1349 rgd
= globaldata_find(cpuid
);
1350 lwkt_setcpu_self(rgd
);
1355 * Remote IPI for cpu migration (called while in a critical section so we
1356 * do not have to enter another one). The thread has already been moved to
1357 * our cpu's allq, but we must wait for the thread to be completely switched
1358 * out on the originating cpu before we schedule it on ours or the stack
1359 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1360 * change to main memory.
1362 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1363 * against wakeups. It is best if this interface is used only when there
1364 * are no pending events that might try to schedule the thread.
1368 lwkt_setcpu_remote(void *arg
)
1371 globaldata_t gd
= mycpu
;
1373 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1375 lwkt_process_ipiq();
1381 td
->td_flags
&= ~TDF_MIGRATING
;
1382 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1388 lwkt_preempted_proc(void)
1390 thread_t td
= curthread
;
1391 while (td
->td_preempted
)
1392 td
= td
->td_preempted
;
1397 * Create a kernel process/thread/whatever. It shares it's address space
1398 * with proc0 - ie: kernel only.
1400 * NOTE! By default new threads are created with the MP lock held. A
1401 * thread which does not require the MP lock should release it by calling
1402 * rel_mplock() at the start of the new thread.
1405 lwkt_create(void (*func
)(void *), void *arg
,
1406 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1407 const char *fmt
, ...)
1412 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1416 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1419 * Set up arg0 for 'ps' etc
1421 __va_start(ap
, fmt
);
1422 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1426 * Schedule the thread to run
1428 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1431 td
->td_flags
&= ~TDF_STOPREQ
;
1436 * Destroy an LWKT thread. Warning! This function is not called when
1437 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1438 * uses a different reaping mechanism.
1443 thread_t td
= curthread
;
1447 if (td
->td_flags
& TDF_VERBOSE
)
1448 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1452 * Get us into a critical section to interlock gd_freetd and loop
1453 * until we can get it freed.
1455 * We have to cache the current td in gd_freetd because objcache_put()ing
1456 * it would rip it out from under us while our thread is still active.
1459 crit_enter_quick(td
);
1460 while ((std
= gd
->gd_freetd
) != NULL
) {
1461 gd
->gd_freetd
= NULL
;
1462 objcache_put(thread_cache
, std
);
1466 * Remove thread resources from kernel lists and deschedule us for
1469 if (td
->td_flags
& TDF_TSLEEPQ
)
1471 lwkt_deschedule_self(td
);
1472 lwkt_remove_tdallq(td
);
1473 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1479 lwkt_remove_tdallq(thread_t td
)
1481 KKASSERT(td
->td_gd
== mycpu
);
1482 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1488 thread_t td
= curthread
;
1489 int lpri
= td
->td_pri
;
1492 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1498 * Called from debugger/panic on cpus which have been stopped. We must still
1499 * process the IPIQ while stopped, even if we were stopped while in a critical
1502 * If we are dumping also try to process any pending interrupts. This may
1503 * or may not work depending on the state of the cpu at the point it was
1507 lwkt_smp_stopped(void)
1509 globaldata_t gd
= mycpu
;
1513 lwkt_process_ipiq();
1516 lwkt_process_ipiq();
1522 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1523 * get_mplock() has already incremented td_mpcount. We must block and
1524 * not return until giant is held.
1526 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1527 * reschedule the thread until it can obtain the giant lock for it.
1530 lwkt_mp_lock_contested(void)
1539 * The rel_mplock() code will call this function after releasing the
1540 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1542 * We then chain an IPI to a single other cpu potentially needing the
1543 * lock. This is a bit heuristical and we can wind up with IPIs flying
1544 * all over the place.
1546 static void lwkt_mp_lock_uncontested_remote(void *arg __unused
);
1549 lwkt_mp_lock_uncontested(void)
1559 atomic_clear_int(&mp_lock_contention_mask
, gd
->gd_cpumask
);
1560 mask
= mp_lock_contention_mask
;
1561 tmpmask
= ~((1 << gd
->gd_cpuid
) - 1);
1565 cpuid
= bsfl(mask
& tmpmask
);
1568 atomic_clear_int(&mp_lock_contention_mask
, 1 << cpuid
);
1569 dgd
= globaldata_find(cpuid
);
1570 lwkt_send_ipiq(dgd
, lwkt_mp_lock_uncontested_remote
, NULL
);
1576 * The idea is for this IPI to interrupt a potentially lower priority
1577 * thread, such as a user thread, to allow the scheduler to reschedule
1578 * a higher priority kernel thread that needs the MP lock.
1580 * For now we set the LWKT reschedule flag which generates an AST in
1581 * doreti, though theoretically it is also possible to possibly preempt
1582 * here if the underlying thread was operating in user mode. Nah.
1585 lwkt_mp_lock_uncontested_remote(void *arg __unused
)
1587 need_lwkt_resched();