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 1.120 2008/10/26 04:29:19 sephe 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>
74 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
76 static int untimely_switch
= 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;
90 static struct objcache
*thread_cache
;
92 volatile cpumask_t mp_lock_contention_mask
;
94 extern void cpu_heavy_restore(void);
95 extern void cpu_lwkt_restore(void);
96 extern void cpu_kthread_restore(void);
97 extern void cpu_idle_restore(void);
100 jg_tos_ok(struct thread
*td
)
105 KKASSERT(td
->td_sp
!= NULL
);
106 unsigned long tos
= ((unsigned long *)td
->td_sp
)[0];
108 if ((tos
== cpu_heavy_restore
) || (tos
== cpu_lwkt_restore
)
109 || (tos
== cpu_kthread_restore
) || (tos
== cpu_idle_restore
)) {
116 * We can make all thread ports use the spin backend instead of the thread
117 * backend. This should only be set to debug the spin backend.
119 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
121 SYSCTL_INT(_lwkt
, OID_AUTO
, untimely_switch
, CTLFLAG_RW
, &untimely_switch
, 0, "");
123 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
126 SYSCTL_INT(_lwkt
, OID_AUTO
, chain_mplock
, CTLFLAG_RW
, &chain_mplock
, 0, "");
128 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
129 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0, "");
130 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0, "");
131 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
133 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
134 &token_contention_count
, 0, "spinning due to token contention");
135 SYSCTL_QUAD(_lwkt
, OID_AUTO
, mplock_contention_count
, CTLFLAG_RW
,
136 &mplock_contention_count
, 0, "spinning due to MPLOCK contention");
142 #if !defined(KTR_GIANT_CONTENTION)
143 #define KTR_GIANT_CONTENTION KTR_ALL
146 KTR_INFO_MASTER(giant
);
147 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, beg
, 0, "thread=%p", sizeof(void *));
148 KTR_INFO(KTR_GIANT_CONTENTION
, giant
, end
, 1, "thread=%p", sizeof(void *));
150 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
153 * These helper procedures handle the runq, they can only be called from
154 * within a critical section.
156 * WARNING! Prior to SMP being brought up it is possible to enqueue and
157 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
158 * instead of 'mycpu' when referencing the globaldata structure. Once
159 * SMP live enqueuing and dequeueing only occurs on the current cpu.
163 _lwkt_dequeue(thread_t td
)
165 if (td
->td_flags
& TDF_RUNQ
) {
166 int nq
= td
->td_pri
& TDPRI_MASK
;
167 struct globaldata
*gd
= td
->td_gd
;
169 td
->td_flags
&= ~TDF_RUNQ
;
170 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
171 /* runqmask is passively cleaned up by the switcher */
177 _lwkt_enqueue(thread_t td
)
179 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_TSLEEPQ
|TDF_BLOCKQ
)) == 0) {
180 int nq
= td
->td_pri
& TDPRI_MASK
;
181 struct globaldata
*gd
= td
->td_gd
;
183 td
->td_flags
|= TDF_RUNQ
;
184 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], td
, td_threadq
);
185 gd
->gd_runqmask
|= 1 << nq
;
190 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
192 struct thread
*td
= (struct thread
*)obj
;
194 td
->td_kstack
= NULL
;
195 td
->td_kstack_size
= 0;
196 td
->td_flags
= TDF_ALLOCATED_THREAD
;
201 _lwkt_thread_dtor(void *obj
, void *privdata
)
203 struct thread
*td
= (struct thread
*)obj
;
205 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
206 ("_lwkt_thread_dtor: not allocated from objcache"));
207 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
208 td
->td_kstack_size
> 0,
209 ("_lwkt_thread_dtor: corrupted stack"));
210 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
214 * Initialize the lwkt s/system.
219 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
220 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
),
221 NULL
, CACHE_NTHREADS
/2,
222 _lwkt_thread_ctor
, _lwkt_thread_dtor
, NULL
);
226 * Schedule a thread to run. As the current thread we can always safely
227 * schedule ourselves, and a shortcut procedure is provided for that
230 * (non-blocking, self contained on a per cpu basis)
233 lwkt_schedule_self(thread_t td
)
235 crit_enter_quick(td
);
236 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
237 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
243 * Deschedule a thread.
245 * (non-blocking, self contained on a per cpu basis)
248 lwkt_deschedule_self(thread_t td
)
250 crit_enter_quick(td
);
256 * LWKTs operate on a per-cpu basis
258 * WARNING! Called from early boot, 'mycpu' may not work yet.
261 lwkt_gdinit(struct globaldata
*gd
)
265 for (i
= 0; i
< sizeof(gd
->gd_tdrunq
)/sizeof(gd
->gd_tdrunq
[0]); ++i
)
266 TAILQ_INIT(&gd
->gd_tdrunq
[i
]);
268 TAILQ_INIT(&gd
->gd_tdallq
);
272 * Create a new thread. The thread must be associated with a process context
273 * or LWKT start address before it can be scheduled. If the target cpu is
274 * -1 the thread will be created on the current cpu.
276 * If you intend to create a thread without a process context this function
277 * does everything except load the startup and switcher function.
280 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
282 globaldata_t gd
= mycpu
;
286 * If static thread storage is not supplied allocate a thread. Reuse
287 * a cached free thread if possible. gd_freetd is used to keep an exiting
288 * thread intact through the exit.
291 if ((td
= gd
->gd_freetd
) != NULL
)
292 gd
->gd_freetd
= NULL
;
294 td
= objcache_get(thread_cache
, M_WAITOK
);
295 KASSERT((td
->td_flags
&
296 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
297 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
298 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
302 * Try to reuse cached stack.
304 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
305 if (flags
& TDF_ALLOCATED_STACK
) {
306 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
311 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
312 flags
|= TDF_ALLOCATED_STACK
;
315 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
317 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
322 * Initialize a preexisting thread structure. This function is used by
323 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
325 * All threads start out in a critical section at a priority of
326 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
327 * appropriate. This function may send an IPI message when the
328 * requested cpu is not the current cpu and consequently gd_tdallq may
329 * not be initialized synchronously from the point of view of the originating
332 * NOTE! we have to be careful in regards to creating threads for other cpus
333 * if SMP has not yet been activated.
338 lwkt_init_thread_remote(void *arg
)
343 * Protected by critical section held by IPI dispatch
345 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
351 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
352 struct globaldata
*gd
)
354 globaldata_t mygd
= mycpu
;
356 bzero(td
, sizeof(struct thread
));
357 td
->td_kstack
= stack
;
358 td
->td_kstack_size
= stksize
;
359 td
->td_flags
= flags
;
361 td
->td_pri
= TDPRI_KERN_DAEMON
+ TDPRI_CRIT
;
363 if ((flags
& TDF_MPSAFE
) == 0)
366 if (lwkt_use_spin_port
)
367 lwkt_initport_spin(&td
->td_msgport
);
369 lwkt_initport_thread(&td
->td_msgport
, td
);
370 pmap_init_thread(td
);
373 * Normally initializing a thread for a remote cpu requires sending an
374 * IPI. However, the idlethread is setup before the other cpus are
375 * activated so we have to treat it as a special case. XXX manipulation
376 * of gd_tdallq requires the BGL.
378 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
380 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
383 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
387 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
393 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
398 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
403 lwkt_hold(thread_t td
)
409 lwkt_rele(thread_t td
)
411 KKASSERT(td
->td_refs
> 0);
416 lwkt_wait_free(thread_t td
)
419 tsleep(td
, 0, "tdreap", hz
);
423 lwkt_free_thread(thread_t td
)
425 KASSERT((td
->td_flags
& TDF_RUNNING
) == 0,
426 ("lwkt_free_thread: did not exit! %p", td
));
428 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
429 objcache_put(thread_cache
, td
);
430 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
431 /* client-allocated struct with internally allocated stack */
432 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
433 ("lwkt_free_thread: corrupted stack"));
434 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
435 td
->td_kstack
= NULL
;
436 td
->td_kstack_size
= 0;
442 * Switch to the next runnable lwkt. If no LWKTs are runnable then
443 * switch to the idlethread. Switching must occur within a critical
444 * section to avoid races with the scheduling queue.
446 * We always have full control over our cpu's run queue. Other cpus
447 * that wish to manipulate our queue must use the cpu_*msg() calls to
448 * talk to our cpu, so a critical section is all that is needed and
449 * the result is very, very fast thread switching.
451 * The LWKT scheduler uses a fixed priority model and round-robins at
452 * each priority level. User process scheduling is a totally
453 * different beast and LWKT priorities should not be confused with
454 * user process priorities.
456 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
457 * cleans it up. Note that the td_switch() function cannot do anything that
458 * requires the MP lock since the MP lock will have already been setup for
459 * the target thread (not the current thread). It's nice to have a scheduler
460 * that does not need the MP lock to work because it allows us to do some
461 * really cool high-performance MP lock optimizations.
463 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
464 * is not called by the current thread in the preemption case, only when
465 * the preempting thread blocks (in order to return to the original thread).
470 globaldata_t gd
= mycpu
;
471 thread_t td
= gd
->gd_curthread
;
478 * Switching from within a 'fast' (non thread switched) interrupt or IPI
479 * is illegal. However, we may have to do it anyway if we hit a fatal
480 * kernel trap or we have paniced.
482 * If this case occurs save and restore the interrupt nesting level.
484 if (gd
->gd_intr_nesting_level
) {
488 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
489 panic("lwkt_switch: cannot switch from within "
490 "a fast interrupt, yet, td %p\n", td
);
492 savegdnest
= gd
->gd_intr_nesting_level
;
493 savegdtrap
= gd
->gd_trap_nesting_level
;
494 gd
->gd_intr_nesting_level
= 0;
495 gd
->gd_trap_nesting_level
= 0;
496 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
497 td
->td_flags
|= TDF_PANICWARN
;
498 kprintf("Warning: thread switch from interrupt or IPI, "
499 "thread %p (%s)\n", td
, td
->td_comm
);
503 gd
->gd_intr_nesting_level
= savegdnest
;
504 gd
->gd_trap_nesting_level
= savegdtrap
;
510 * Passive release (used to transition from user to kernel mode
511 * when we block or switch rather then when we enter the kernel).
512 * This function is NOT called if we are switching into a preemption
513 * or returning from a preemption. Typically this causes us to lose
514 * our current process designation (if we have one) and become a true
515 * LWKT thread, and may also hand the current process designation to
516 * another process and schedule thread.
523 lwkt_relalltokens(td
);
526 * We had better not be holding any spin locks, but don't get into an
527 * endless panic loop.
529 KASSERT(gd
->gd_spinlock_rd
== NULL
|| panicstr
!= NULL
,
530 ("lwkt_switch: still holding a shared spinlock %p!",
531 gd
->gd_spinlock_rd
));
532 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
533 ("lwkt_switch: still holding %d exclusive spinlocks!",
534 gd
->gd_spinlocks_wr
));
539 * td_mpcount cannot be used to determine if we currently hold the
540 * MP lock because get_mplock() will increment it prior to attempting
541 * to get the lock, and switch out if it can't. Our ownership of
542 * the actual lock will remain stable while we are in a critical section
543 * (but, of course, another cpu may own or release the lock so the
544 * actual value of mp_lock is not stable).
546 mpheld
= MP_LOCK_HELD();
548 if (td
->td_cscount
) {
549 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
551 if (panic_on_cscount
)
552 panic("switching while mastering cpusync");
556 if ((ntd
= td
->td_preempted
) != NULL
) {
558 * We had preempted another thread on this cpu, resume the preempted
559 * thread. This occurs transparently, whether the preempted thread
560 * was scheduled or not (it may have been preempted after descheduling
563 * We have to setup the MP lock for the original thread after backing
564 * out the adjustment that was made to curthread when the original
567 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
569 if (ntd
->td_mpcount
&& mpheld
== 0) {
570 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
571 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
);
573 if (ntd
->td_mpcount
) {
574 td
->td_mpcount
-= ntd
->td_mpcount
;
575 KKASSERT(td
->td_mpcount
>= 0);
578 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
581 * The interrupt may have woken a thread up, we need to properly
582 * set the reschedule flag if the originally interrupted thread is
583 * at a lower priority.
585 if (gd
->gd_runqmask
> (2 << (ntd
->td_pri
& TDPRI_MASK
)) - 1)
587 /* YYY release mp lock on switchback if original doesn't need it */
590 * Priority queue / round-robin at each priority. Note that user
591 * processes run at a fixed, low priority and the user process
592 * scheduler deals with interactions between user processes
593 * by scheduling and descheduling them from the LWKT queue as
596 * We have to adjust the MP lock for the target thread. If we
597 * need the MP lock and cannot obtain it we try to locate a
598 * thread that does not need the MP lock. If we cannot, we spin
601 * A similar issue exists for the tokens held by the target thread.
602 * If we cannot obtain ownership of the tokens we cannot immediately
603 * schedule the thread.
607 * If an LWKT reschedule was requested, well that is what we are
608 * doing now so clear it.
610 clear_lwkt_resched();
612 if (gd
->gd_runqmask
) {
613 int nq
= bsrl(gd
->gd_runqmask
);
614 if ((ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
[nq
])) == NULL
) {
615 gd
->gd_runqmask
&= ~(1 << nq
);
620 * THREAD SELECTION FOR AN SMP MACHINE BUILD
622 * If the target needs the MP lock and we couldn't get it,
623 * or if the target is holding tokens and we could not
624 * gain ownership of the tokens, continue looking for a
625 * thread to schedule and spin instead of HLT if we can't.
627 * NOTE: the mpheld variable invalid after this conditional, it
628 * can change due to both cpu_try_mplock() returning success
629 * AND interactions in lwkt_getalltokens() due to the fact that
630 * we are trying to check the mpcount of a thread other then
631 * the current thread. Because of this, if the current thread
632 * is not holding td_mpcount, an IPI indirectly run via
633 * lwkt_getalltokens() can obtain and release the MP lock and
634 * cause the core MP lock to be released.
636 if ((ntd
->td_mpcount
&& mpheld
== 0 && !cpu_try_mplock()) ||
637 (ntd
->td_toks
&& lwkt_getalltokens(ntd
) == 0)
639 u_int32_t rqmask
= gd
->gd_runqmask
;
641 mpheld
= MP_LOCK_HELD();
644 TAILQ_FOREACH(ntd
, &gd
->gd_tdrunq
[nq
], td_threadq
) {
645 if (ntd
->td_mpcount
&& !mpheld
&& !cpu_try_mplock()) {
646 /* spinning due to MP lock being held */
648 ++mplock_contention_count
;
650 /* mplock still not held, 'mpheld' still valid */
655 * mpheld state invalid after getalltokens call returns
656 * failure, but the variable is only needed for
659 if (ntd
->td_toks
&& !lwkt_getalltokens(ntd
)) {
660 /* spinning due to token contention */
662 ++token_contention_count
;
664 mpheld
= MP_LOCK_HELD();
671 rqmask
&= ~(1 << nq
);
675 * We have two choices. We can either refuse to run a
676 * user thread when a kernel thread needs the MP lock
677 * but could not get it, or we can allow it to run but
678 * then expect an IPI (hopefully) later on to force a
679 * reschedule when the MP lock might become available.
681 if (nq
< TDPRI_KERN_LPSCHED
) {
682 if (chain_mplock
== 0)
684 atomic_set_int(&mp_lock_contention_mask
,
686 /* continue loop, allow user threads to be scheduled */
690 cpu_mplock_contested();
691 ntd
= &gd
->gd_idlethread
;
692 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
693 goto using_idle_thread
;
695 ++gd
->gd_cnt
.v_swtch
;
696 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
697 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
700 ++gd
->gd_cnt
.v_swtch
;
701 TAILQ_REMOVE(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
702 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
[nq
], ntd
, td_threadq
);
706 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
707 * worry about tokens or the BGL. However, we still have
708 * to call lwkt_getalltokens() in order to properly detect
709 * stale tokens. This call cannot fail for a UP build!
711 lwkt_getalltokens(ntd
);
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 * We have nothing to run but only let the idle loop halt
719 * the cpu if there are no pending interrupts.
721 ntd
= &gd
->gd_idlethread
;
722 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
723 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
727 * The idle thread should not be holding the MP lock unless we
728 * are trapping in the kernel or in a panic. Since we select the
729 * idle thread unconditionally when no other thread is available,
730 * if the MP lock is desired during a panic or kernel trap, we
731 * have to loop in the scheduler until we get it.
733 if (ntd
->td_mpcount
) {
734 mpheld
= MP_LOCK_HELD();
735 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
) {
736 panic("Idle thread %p was holding the BGL!", ntd
);
737 } else if (mpheld
== 0) {
738 cpu_mplock_contested();
745 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
,
746 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
749 * Do the actual switch. If the new target does not need the MP lock
750 * and we are holding it, release the MP lock. If the new target requires
751 * the MP lock we have already acquired it for the target.
754 if (ntd
->td_mpcount
== 0 ) {
758 ASSERT_MP_LOCK_HELD(ntd
);
763 KKASSERT(jg_tos_ok(ntd
));
766 /* NOTE: current cpu may have changed after switch */
771 * Request that the target thread preempt the current thread. Preemption
772 * only works under a specific set of conditions:
774 * - We are not preempting ourselves
775 * - The target thread is owned by the current cpu
776 * - We are not currently being preempted
777 * - The target is not currently being preempted
778 * - We are not holding any spin locks
779 * - The target thread is not holding any tokens
780 * - We are able to satisfy the target's MP lock requirements (if any).
782 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
783 * this is called via lwkt_schedule() through the td_preemptable callback.
784 * critpri is the managed critical priority that we should ignore in order
785 * to determine whether preemption is possible (aka usually just the crit
786 * priority of lwkt_schedule() itself).
788 * XXX at the moment we run the target thread in a critical section during
789 * the preemption in order to prevent the target from taking interrupts
790 * that *WE* can't. Preemption is strictly limited to interrupt threads
791 * and interrupt-like threads, outside of a critical section, and the
792 * preempted source thread will be resumed the instant the target blocks
793 * whether or not the source is scheduled (i.e. preemption is supposed to
794 * be as transparent as possible).
796 * The target thread inherits our MP count (added to its own) for the
797 * duration of the preemption in order to preserve the atomicy of the
798 * MP lock during the preemption. Therefore, any preempting targets must be
799 * careful in regards to MP assertions. Note that the MP count may be
800 * out of sync with the physical mp_lock, but we do not have to preserve
801 * the original ownership of the lock if it was out of synch (that is, we
802 * can leave it synchronized on return).
805 lwkt_preempt(thread_t ntd
, int critpri
)
807 struct globaldata
*gd
= mycpu
;
815 * The caller has put us in a critical section. We can only preempt
816 * if the caller of the caller was not in a critical section (basically
817 * a local interrupt), as determined by the 'critpri' parameter. We
818 * also can't preempt if the caller is holding any spinlocks (even if
819 * he isn't in a critical section). This also handles the tokens test.
821 * YYY The target thread must be in a critical section (else it must
822 * inherit our critical section? I dunno yet).
824 * Set need_lwkt_resched() unconditionally for now YYY.
826 KASSERT(ntd
->td_pri
>= TDPRI_CRIT
, ("BADCRIT0 %d", ntd
->td_pri
));
828 td
= gd
->gd_curthread
;
829 if ((ntd
->td_pri
& TDPRI_MASK
) <= (td
->td_pri
& TDPRI_MASK
)) {
833 if ((td
->td_pri
& ~TDPRI_MASK
) > critpri
) {
839 if (ntd
->td_gd
!= gd
) {
846 * Take the easy way out and do not preempt if we are holding
847 * any spinlocks. We could test whether the thread(s) being
848 * preempted interlock against the target thread's tokens and whether
849 * we can get all the target thread's tokens, but this situation
850 * should not occur very often so its easier to simply not preempt.
851 * Also, plain spinlocks are impossible to figure out at this point so
852 * just don't preempt.
854 * Do not try to preempt if the target thread is holding any tokens.
855 * We could try to acquire the tokens but this case is so rare there
856 * is no need to support it.
858 if (gd
->gd_spinlock_rd
|| gd
->gd_spinlocks_wr
) {
868 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
873 if (ntd
->td_preempted
) {
880 * note: an interrupt might have occured just as we were transitioning
881 * to or from the MP lock. In this case td_mpcount will be pre-disposed
882 * (non-zero) but not actually synchronized with the actual state of the
883 * lock. We can use it to imply an MP lock requirement for the
884 * preemption but we cannot use it to test whether we hold the MP lock
887 savecnt
= td
->td_mpcount
;
888 mpheld
= MP_LOCK_HELD();
889 ntd
->td_mpcount
+= td
->td_mpcount
;
890 if (mpheld
== 0 && ntd
->td_mpcount
&& !cpu_try_mplock()) {
891 ntd
->td_mpcount
-= td
->td_mpcount
;
899 * Since we are able to preempt the current thread, there is no need to
900 * call need_lwkt_resched().
903 ntd
->td_preempted
= td
;
904 td
->td_flags
|= TDF_PREEMPT_LOCK
;
907 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
909 KKASSERT(savecnt
== td
->td_mpcount
);
910 mpheld
= MP_LOCK_HELD();
911 if (mpheld
&& td
->td_mpcount
== 0)
913 else if (mpheld
== 0 && td
->td_mpcount
)
914 panic("lwkt_preempt(): MP lock was not held through");
916 ntd
->td_preempted
= NULL
;
917 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
921 * Yield our thread while higher priority threads are pending. This is
922 * typically called when we leave a critical section but it can be safely
923 * called while we are in a critical section.
925 * This function will not generally yield to equal priority threads but it
926 * can occur as a side effect. Note that lwkt_switch() is called from
927 * inside the critical section to prevent its own crit_exit() from reentering
928 * lwkt_yield_quick().
930 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
931 * came along but was blocked and made pending.
933 * (self contained on a per cpu basis)
936 lwkt_yield_quick(void)
938 globaldata_t gd
= mycpu
;
939 thread_t td
= gd
->gd_curthread
;
942 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
943 * it with a non-zero cpl then we might not wind up calling splz after
944 * a task switch when the critical section is exited even though the
945 * new task could accept the interrupt.
947 * XXX from crit_exit() only called after last crit section is released.
948 * If called directly will run splz() even if in a critical section.
950 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
951 * except for this special case, we MUST call splz() here to handle any
952 * pending ints, particularly after we switch, or we might accidently
953 * halt the cpu with interrupts pending.
955 if (gd
->gd_reqflags
&& td
->td_nest_count
< 2)
959 * YYY enabling will cause wakeup() to task-switch, which really
960 * confused the old 4.x code. This is a good way to simulate
961 * preemption and MP without actually doing preemption or MP, because a
962 * lot of code assumes that wakeup() does not block.
964 if (untimely_switch
&& td
->td_nest_count
== 0 &&
965 gd
->gd_intr_nesting_level
== 0
967 crit_enter_quick(td
);
969 * YYY temporary hacks until we disassociate the userland scheduler
970 * from the LWKT scheduler.
972 if (td
->td_flags
& TDF_RUNQ
) {
973 lwkt_switch(); /* will not reenter yield function */
975 lwkt_schedule_self(td
); /* make sure we are scheduled */
976 lwkt_switch(); /* will not reenter yield function */
977 lwkt_deschedule_self(td
); /* make sure we are descheduled */
979 crit_exit_noyield(td
);
984 * This implements a normal yield which, unlike _quick, will yield to equal
985 * priority threads as well. Note that gd_reqflags tests will be handled by
986 * the crit_exit() call in lwkt_switch().
988 * (self contained on a per cpu basis)
993 lwkt_schedule_self(curthread
);
998 * Return 0 if no runnable threads are pending at the same or higher
999 * priority as the passed thread.
1001 * Return 1 if runnable threads are pending at the same priority.
1003 * Return 2 if runnable threads are pending at a higher priority.
1006 lwkt_check_resched(thread_t td
)
1008 int pri
= td
->td_pri
& TDPRI_MASK
;
1010 if (td
->td_gd
->gd_runqmask
> (2 << pri
) - 1)
1012 if (TAILQ_NEXT(td
, td_threadq
))
1018 * Generic schedule. Possibly schedule threads belonging to other cpus and
1019 * deal with threads that might be blocked on a wait queue.
1021 * We have a little helper inline function which does additional work after
1022 * the thread has been enqueued, including dealing with preemption and
1023 * setting need_lwkt_resched() (which prevents the kernel from returning
1024 * to userland until it has processed higher priority threads).
1026 * It is possible for this routine to be called after a failed _enqueue
1027 * (due to the target thread migrating, sleeping, or otherwise blocked).
1028 * We have to check that the thread is actually on the run queue!
1030 * reschedok is an optimized constant propagated from lwkt_schedule() or
1031 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1032 * reschedule to be requested if the target thread has a higher priority.
1033 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1034 * be 0, prevented undesired reschedules.
1038 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int cpri
, int reschedok
)
1042 if (ntd
->td_flags
& TDF_RUNQ
) {
1043 if (ntd
->td_preemptable
&& reschedok
) {
1044 ntd
->td_preemptable(ntd
, cpri
); /* YYY +token */
1045 } else if (reschedok
) {
1047 if ((ntd
->td_pri
& TDPRI_MASK
) > (otd
->td_pri
& TDPRI_MASK
))
1048 need_lwkt_resched();
1055 _lwkt_schedule(thread_t td
, int reschedok
)
1057 globaldata_t mygd
= mycpu
;
1059 KASSERT(td
!= &td
->td_gd
->gd_idlethread
, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1060 crit_enter_gd(mygd
);
1061 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1062 if (td
== mygd
->gd_curthread
) {
1066 * If we own the thread, there is no race (since we are in a
1067 * critical section). If we do not own the thread there might
1068 * be a race but the target cpu will deal with it.
1071 if (td
->td_gd
== mygd
) {
1073 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1075 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_schedule
, td
);
1079 _lwkt_schedule_post(mygd
, td
, TDPRI_CRIT
, reschedok
);
1086 lwkt_schedule(thread_t td
)
1088 _lwkt_schedule(td
, 1);
1092 lwkt_schedule_noresched(thread_t td
)
1094 _lwkt_schedule(td
, 0);
1100 * Thread migration using a 'Pull' method. The thread may or may not be
1101 * the current thread. It MUST be descheduled and in a stable state.
1102 * lwkt_giveaway() must be called on the cpu owning the thread.
1104 * At any point after lwkt_giveaway() is called, the target cpu may
1105 * 'pull' the thread by calling lwkt_acquire().
1107 * MPSAFE - must be called under very specific conditions.
1110 lwkt_giveaway(thread_t td
)
1112 globaldata_t gd
= mycpu
;
1115 KKASSERT(td
->td_gd
== gd
);
1116 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1117 td
->td_flags
|= TDF_MIGRATING
;
1122 lwkt_acquire(thread_t td
)
1127 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1132 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1133 crit_enter_gd(mygd
);
1134 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1136 lwkt_process_ipiq();
1141 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1142 td
->td_flags
&= ~TDF_MIGRATING
;
1145 crit_enter_gd(mygd
);
1146 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1147 td
->td_flags
&= ~TDF_MIGRATING
;
1155 * Generic deschedule. Descheduling threads other then your own should be
1156 * done only in carefully controlled circumstances. Descheduling is
1159 * This function may block if the cpu has run out of messages.
1162 lwkt_deschedule(thread_t td
)
1166 if (td
== curthread
) {
1169 if (td
->td_gd
== mycpu
) {
1172 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1182 * Set the target thread's priority. This routine does not automatically
1183 * switch to a higher priority thread, LWKT threads are not designed for
1184 * continuous priority changes. Yield if you want to switch.
1186 * We have to retain the critical section count which uses the high bits
1187 * of the td_pri field. The specified priority may also indicate zero or
1188 * more critical sections by adding TDPRI_CRIT*N.
1190 * Note that we requeue the thread whether it winds up on a different runq
1191 * or not. uio_yield() depends on this and the routine is not normally
1192 * called with the same priority otherwise.
1195 lwkt_setpri(thread_t td
, int pri
)
1198 KKASSERT(td
->td_gd
== mycpu
);
1200 if (td
->td_flags
& TDF_RUNQ
) {
1202 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1205 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1211 lwkt_setpri_self(int pri
)
1213 thread_t td
= curthread
;
1215 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1217 if (td
->td_flags
& TDF_RUNQ
) {
1219 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1222 td
->td_pri
= (td
->td_pri
& ~TDPRI_MASK
) + pri
;
1228 * Migrate the current thread to the specified cpu.
1230 * This is accomplished by descheduling ourselves from the current cpu,
1231 * moving our thread to the tdallq of the target cpu, IPI messaging the
1232 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1233 * races while the thread is being migrated.
1236 static void lwkt_setcpu_remote(void *arg
);
1240 lwkt_setcpu_self(globaldata_t rgd
)
1243 thread_t td
= curthread
;
1245 if (td
->td_gd
!= rgd
) {
1246 crit_enter_quick(td
);
1247 td
->td_flags
|= TDF_MIGRATING
;
1248 lwkt_deschedule_self(td
);
1249 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1250 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1252 /* we are now on the target cpu */
1253 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1254 crit_exit_quick(td
);
1260 lwkt_migratecpu(int cpuid
)
1265 rgd
= globaldata_find(cpuid
);
1266 lwkt_setcpu_self(rgd
);
1271 * Remote IPI for cpu migration (called while in a critical section so we
1272 * do not have to enter another one). The thread has already been moved to
1273 * our cpu's allq, but we must wait for the thread to be completely switched
1274 * out on the originating cpu before we schedule it on ours or the stack
1275 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1276 * change to main memory.
1278 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1279 * against wakeups. It is best if this interface is used only when there
1280 * are no pending events that might try to schedule the thread.
1284 lwkt_setcpu_remote(void *arg
)
1287 globaldata_t gd
= mycpu
;
1289 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1291 lwkt_process_ipiq();
1297 td
->td_flags
&= ~TDF_MIGRATING
;
1298 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1304 lwkt_preempted_proc(void)
1306 thread_t td
= curthread
;
1307 while (td
->td_preempted
)
1308 td
= td
->td_preempted
;
1313 * Create a kernel process/thread/whatever. It shares it's address space
1314 * with proc0 - ie: kernel only.
1316 * NOTE! By default new threads are created with the MP lock held. A
1317 * thread which does not require the MP lock should release it by calling
1318 * rel_mplock() at the start of the new thread.
1321 lwkt_create(void (*func
)(void *), void *arg
,
1322 struct thread
**tdp
, thread_t
template, int tdflags
, int cpu
,
1323 const char *fmt
, ...)
1328 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1332 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1335 * Set up arg0 for 'ps' etc
1337 __va_start(ap
, fmt
);
1338 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1342 * Schedule the thread to run
1344 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1347 td
->td_flags
&= ~TDF_STOPREQ
;
1352 * Destroy an LWKT thread. Warning! This function is not called when
1353 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1354 * uses a different reaping mechanism.
1359 thread_t td
= curthread
;
1363 if (td
->td_flags
& TDF_VERBOSE
)
1364 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1368 * Get us into a critical section to interlock gd_freetd and loop
1369 * until we can get it freed.
1371 * We have to cache the current td in gd_freetd because objcache_put()ing
1372 * it would rip it out from under us while our thread is still active.
1375 crit_enter_quick(td
);
1376 while ((std
= gd
->gd_freetd
) != NULL
) {
1377 gd
->gd_freetd
= NULL
;
1378 objcache_put(thread_cache
, std
);
1380 lwkt_deschedule_self(td
);
1381 lwkt_remove_tdallq(td
);
1382 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1388 lwkt_remove_tdallq(thread_t td
)
1390 KKASSERT(td
->td_gd
== mycpu
);
1391 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1397 thread_t td
= curthread
;
1398 int lpri
= td
->td_pri
;
1401 panic("td_pri is/would-go negative! %p %d", td
, lpri
);
1407 * Called from debugger/panic on cpus which have been stopped. We must still
1408 * process the IPIQ while stopped, even if we were stopped while in a critical
1411 * If we are dumping also try to process any pending interrupts. This may
1412 * or may not work depending on the state of the cpu at the point it was
1416 lwkt_smp_stopped(void)
1418 globaldata_t gd
= mycpu
;
1422 lwkt_process_ipiq();
1425 lwkt_process_ipiq();
1431 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1432 * get_mplock() has already incremented td_mpcount. We must block and
1433 * not return until giant is held.
1435 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1436 * reschedule the thread until it can obtain the giant lock for it.
1439 lwkt_mp_lock_contested(void)
1447 * The rel_mplock() code will call this function after releasing the
1448 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1450 * We then chain an IPI to a single other cpu potentially needing the
1451 * lock. This is a bit heuristical and we can wind up with IPIs flying
1452 * all over the place.
1454 static void lwkt_mp_lock_uncontested_remote(void *arg __unused
);
1457 lwkt_mp_lock_uncontested(void)
1467 atomic_clear_int(&mp_lock_contention_mask
, gd
->gd_cpumask
);
1468 mask
= mp_lock_contention_mask
;
1469 tmpmask
= ~((1 << gd
->gd_cpuid
) - 1);
1473 cpuid
= bsfl(mask
& tmpmask
);
1476 atomic_clear_int(&mp_lock_contention_mask
, 1 << cpuid
);
1477 dgd
= globaldata_find(cpuid
);
1478 lwkt_send_ipiq(dgd
, lwkt_mp_lock_uncontested_remote
, NULL
);
1484 * The idea is for this IPI to interrupt a potentially lower priority
1485 * thread, such as a user thread, to allow the scheduler to reschedule
1486 * a higher priority kernel thread that needs the MP lock.
1488 * For now we set the LWKT reschedule flag which generates an AST in
1489 * doreti, though theoretically it is also possible to possibly preempt
1490 * here if the underlying thread was operating in user mode. Nah.
1493 lwkt_mp_lock_uncontested_remote(void *arg __unused
)
1495 need_lwkt_resched();