2 * Copyright (c) 2003-2010 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/kinfo.h>
48 #include <sys/queue.h>
49 #include <sys/sysctl.h>
50 #include <sys/kthread.h>
51 #include <machine/cpu.h>
54 #include <sys/spinlock.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mplock2.h>
61 #include <sys/dsched.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>
72 #include <machine/stdarg.h>
73 #include <machine/smp.h>
75 #if !defined(KTR_CTXSW)
76 #define KTR_CTXSW KTR_ALL
78 KTR_INFO_MASTER(ctxsw
);
79 KTR_INFO(KTR_CTXSW
, ctxsw
, sw
, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread
*));
81 KTR_INFO(KTR_CTXSW
, ctxsw
, pre
, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread
*));
83 KTR_INFO(KTR_CTXSW
, ctxsw
, newtd
, 2, "#threads[%p].name = %s",
84 sizeof (struct thread
*) + sizeof(char *));
85 KTR_INFO(KTR_CTXSW
, ctxsw
, deadtd
, 3, "#threads[%p].name = <dead>", sizeof (struct thread
*));
87 static MALLOC_DEFINE(M_THREAD
, "thread", "lwkt threads");
90 static int panic_on_cscount
= 0;
92 static __int64_t switch_count
= 0;
93 static __int64_t preempt_hit
= 0;
94 static __int64_t preempt_miss
= 0;
95 static __int64_t preempt_weird
= 0;
96 static __int64_t token_contention_count __debugvar
= 0;
97 static int lwkt_use_spin_port
;
98 static struct objcache
*thread_cache
;
101 static void lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
);
103 static void lwkt_fairq_accumulate(globaldata_t gd
, thread_t td
);
105 extern void cpu_heavy_restore(void);
106 extern void cpu_lwkt_restore(void);
107 extern void cpu_kthread_restore(void);
108 extern void cpu_idle_restore(void);
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port
);
117 SYSCTL_INT(_lwkt
, OID_AUTO
, panic_on_cscount
, CTLFLAG_RW
, &panic_on_cscount
, 0, "");
119 SYSCTL_QUAD(_lwkt
, OID_AUTO
, switch_count
, CTLFLAG_RW
, &switch_count
, 0, "");
120 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_hit
, CTLFLAG_RW
, &preempt_hit
, 0,
121 "Successful preemption events");
122 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_miss
, CTLFLAG_RW
, &preempt_miss
, 0,
123 "Failed preemption events");
124 SYSCTL_QUAD(_lwkt
, OID_AUTO
, preempt_weird
, CTLFLAG_RW
, &preempt_weird
, 0, "");
126 SYSCTL_QUAD(_lwkt
, OID_AUTO
, token_contention_count
, CTLFLAG_RW
,
127 &token_contention_count
, 0, "spinning due to token contention");
129 static int fairq_enable
= 1;
130 SYSCTL_INT(_lwkt
, OID_AUTO
, fairq_enable
, CTLFLAG_RW
, &fairq_enable
, 0, "");
131 static int user_pri_sched
= 0;
132 SYSCTL_INT(_lwkt
, OID_AUTO
, user_pri_sched
, CTLFLAG_RW
, &user_pri_sched
, 0, "");
133 static int preempt_enable
= 1;
134 SYSCTL_INT(_lwkt
, OID_AUTO
, preempt_enable
, CTLFLAG_RW
, &preempt_enable
, 0, "");
138 * These helper procedures handle the runq, they can only be called from
139 * within a critical section.
141 * WARNING! Prior to SMP being brought up it is possible to enqueue and
142 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
143 * instead of 'mycpu' when referencing the globaldata structure. Once
144 * SMP live enqueuing and dequeueing only occurs on the current cpu.
148 _lwkt_dequeue(thread_t td
)
150 if (td
->td_flags
& TDF_RUNQ
) {
151 struct globaldata
*gd
= td
->td_gd
;
153 td
->td_flags
&= ~TDF_RUNQ
;
154 TAILQ_REMOVE(&gd
->gd_tdrunq
, td
, td_threadq
);
155 gd
->gd_fairq_total_pri
-= td
->td_pri
;
156 if (TAILQ_FIRST(&gd
->gd_tdrunq
) == NULL
)
157 atomic_clear_int_nonlocked(&gd
->gd_reqflags
, RQF_RUNNING
);
164 * NOTE: There are a limited number of lwkt threads runnable since user
165 * processes only schedule one at a time per cpu.
169 _lwkt_enqueue(thread_t td
)
173 if ((td
->td_flags
& (TDF_RUNQ
|TDF_MIGRATING
|TDF_BLOCKQ
)) == 0) {
174 struct globaldata
*gd
= td
->td_gd
;
176 td
->td_flags
|= TDF_RUNQ
;
177 xtd
= TAILQ_FIRST(&gd
->gd_tdrunq
);
179 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
, td
, td_threadq
);
180 atomic_set_int_nonlocked(&gd
->gd_reqflags
, RQF_RUNNING
);
182 while (xtd
&& xtd
->td_pri
> td
->td_pri
)
183 xtd
= TAILQ_NEXT(xtd
, td_threadq
);
185 TAILQ_INSERT_BEFORE(xtd
, td
, td_threadq
);
187 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
, td
, td_threadq
);
189 gd
->gd_fairq_total_pri
+= td
->td_pri
;
194 _lwkt_thread_ctor(void *obj
, void *privdata
, int ocflags
)
196 struct thread
*td
= (struct thread
*)obj
;
198 td
->td_kstack
= NULL
;
199 td
->td_kstack_size
= 0;
200 td
->td_flags
= TDF_ALLOCATED_THREAD
;
205 _lwkt_thread_dtor(void *obj
, void *privdata
)
207 struct thread
*td
= (struct thread
*)obj
;
209 KASSERT(td
->td_flags
& TDF_ALLOCATED_THREAD
,
210 ("_lwkt_thread_dtor: not allocated from objcache"));
211 KASSERT((td
->td_flags
& TDF_ALLOCATED_STACK
) && td
->td_kstack
&&
212 td
->td_kstack_size
> 0,
213 ("_lwkt_thread_dtor: corrupted stack"));
214 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
218 * Initialize the lwkt s/system.
223 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
224 thread_cache
= objcache_create_mbacked(M_THREAD
, sizeof(struct thread
),
225 NULL
, CACHE_NTHREADS
/2,
226 _lwkt_thread_ctor
, _lwkt_thread_dtor
, NULL
);
230 * Schedule a thread to run. As the current thread we can always safely
231 * schedule ourselves, and a shortcut procedure is provided for that
234 * (non-blocking, self contained on a per cpu basis)
237 lwkt_schedule_self(thread_t td
)
239 crit_enter_quick(td
);
240 KASSERT(td
!= &td
->td_gd
->gd_idlethread
,
241 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
242 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
248 * Deschedule a thread.
250 * (non-blocking, self contained on a per cpu basis)
253 lwkt_deschedule_self(thread_t td
)
255 crit_enter_quick(td
);
261 * LWKTs operate on a per-cpu basis
263 * WARNING! Called from early boot, 'mycpu' may not work yet.
266 lwkt_gdinit(struct globaldata
*gd
)
268 TAILQ_INIT(&gd
->gd_tdrunq
);
269 TAILQ_INIT(&gd
->gd_tdallq
);
273 * Create a new thread. The thread must be associated with a process context
274 * or LWKT start address before it can be scheduled. If the target cpu is
275 * -1 the thread will be created on the current cpu.
277 * If you intend to create a thread without a process context this function
278 * does everything except load the startup and switcher function.
281 lwkt_alloc_thread(struct thread
*td
, int stksize
, int cpu
, int flags
)
283 globaldata_t gd
= mycpu
;
287 * If static thread storage is not supplied allocate a thread. Reuse
288 * a cached free thread if possible. gd_freetd is used to keep an exiting
289 * thread intact through the exit.
293 if ((td
= gd
->gd_freetd
) != NULL
) {
294 KKASSERT((td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
|
296 gd
->gd_freetd
= NULL
;
298 td
= objcache_get(thread_cache
, M_WAITOK
);
299 KKASSERT((td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
|
303 KASSERT((td
->td_flags
&
304 (TDF_ALLOCATED_THREAD
|TDF_RUNNING
)) == TDF_ALLOCATED_THREAD
,
305 ("lwkt_alloc_thread: corrupted td flags 0x%X", td
->td_flags
));
306 flags
|= td
->td_flags
& (TDF_ALLOCATED_THREAD
|TDF_ALLOCATED_STACK
);
310 * Try to reuse cached stack.
312 if ((stack
= td
->td_kstack
) != NULL
&& td
->td_kstack_size
!= stksize
) {
313 if (flags
& TDF_ALLOCATED_STACK
) {
314 kmem_free(&kernel_map
, (vm_offset_t
)stack
, td
->td_kstack_size
);
319 stack
= (void *)kmem_alloc(&kernel_map
, stksize
);
320 flags
|= TDF_ALLOCATED_STACK
;
323 lwkt_init_thread(td
, stack
, stksize
, flags
, gd
);
325 lwkt_init_thread(td
, stack
, stksize
, flags
, globaldata_find(cpu
));
330 * Initialize a preexisting thread structure. This function is used by
331 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
333 * All threads start out in a critical section at a priority of
334 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
335 * appropriate. This function may send an IPI message when the
336 * requested cpu is not the current cpu and consequently gd_tdallq may
337 * not be initialized synchronously from the point of view of the originating
340 * NOTE! we have to be careful in regards to creating threads for other cpus
341 * if SMP has not yet been activated.
346 lwkt_init_thread_remote(void *arg
)
351 * Protected by critical section held by IPI dispatch
353 TAILQ_INSERT_TAIL(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
359 * lwkt core thread structural initialization.
361 * NOTE: All threads are initialized as mpsafe threads.
364 lwkt_init_thread(thread_t td
, void *stack
, int stksize
, int flags
,
365 struct globaldata
*gd
)
367 globaldata_t mygd
= mycpu
;
369 bzero(td
, sizeof(struct thread
));
370 td
->td_kstack
= stack
;
371 td
->td_kstack_size
= stksize
;
372 td
->td_flags
= flags
;
374 td
->td_pri
= TDPRI_KERN_DAEMON
;
375 td
->td_critcount
= 1;
376 td
->td_toks_stop
= &td
->td_toks_base
;
377 if (lwkt_use_spin_port
)
378 lwkt_initport_spin(&td
->td_msgport
);
380 lwkt_initport_thread(&td
->td_msgport
, td
);
381 pmap_init_thread(td
);
384 * Normally initializing a thread for a remote cpu requires sending an
385 * IPI. However, the idlethread is setup before the other cpus are
386 * activated so we have to treat it as a special case. XXX manipulation
387 * of gd_tdallq requires the BGL.
389 if (gd
== mygd
|| td
== &gd
->gd_idlethread
) {
391 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
394 lwkt_send_ipiq(gd
, lwkt_init_thread_remote
, td
);
398 TAILQ_INSERT_TAIL(&gd
->gd_tdallq
, td
, td_allq
);
402 dsched_new_thread(td
);
406 lwkt_set_comm(thread_t td
, const char *ctl
, ...)
411 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), ctl
, va
);
413 KTR_LOG(ctxsw_newtd
, td
, &td
->td_comm
[0]);
417 lwkt_hold(thread_t td
)
423 lwkt_rele(thread_t td
)
425 KKASSERT(td
->td_refs
> 0);
430 lwkt_wait_free(thread_t td
)
433 tsleep(td
, 0, "tdreap", hz
);
437 lwkt_free_thread(thread_t td
)
439 KKASSERT((td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
|TDF_RUNQ
)) == 0);
440 if (td
->td_flags
& TDF_ALLOCATED_THREAD
) {
441 objcache_put(thread_cache
, td
);
442 } else if (td
->td_flags
& TDF_ALLOCATED_STACK
) {
443 /* client-allocated struct with internally allocated stack */
444 KASSERT(td
->td_kstack
&& td
->td_kstack_size
> 0,
445 ("lwkt_free_thread: corrupted stack"));
446 kmem_free(&kernel_map
, (vm_offset_t
)td
->td_kstack
, td
->td_kstack_size
);
447 td
->td_kstack
= NULL
;
448 td
->td_kstack_size
= 0;
450 KTR_LOG(ctxsw_deadtd
, td
);
455 * Switch to the next runnable lwkt. If no LWKTs are runnable then
456 * switch to the idlethread. Switching must occur within a critical
457 * section to avoid races with the scheduling queue.
459 * We always have full control over our cpu's run queue. Other cpus
460 * that wish to manipulate our queue must use the cpu_*msg() calls to
461 * talk to our cpu, so a critical section is all that is needed and
462 * the result is very, very fast thread switching.
464 * The LWKT scheduler uses a fixed priority model and round-robins at
465 * each priority level. User process scheduling is a totally
466 * different beast and LWKT priorities should not be confused with
467 * user process priorities.
469 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount.
470 * lwkt_switch() cleans it up.
472 * Note that the td_switch() function cannot do anything that requires
473 * the MP lock since the MP lock will have already been setup for
474 * the target thread (not the current thread). It's nice to have a scheduler
475 * that does not need the MP lock to work because it allows us to do some
476 * really cool high-performance MP lock optimizations.
478 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
479 * is not called by the current thread in the preemption case, only when
480 * the preempting thread blocks (in order to return to the original thread).
485 globaldata_t gd
= mycpu
;
486 thread_t td
= gd
->gd_curthread
;
495 const char *lmsg
; /* diagnostic - 'systat -pv 1' */
499 * Switching from within a 'fast' (non thread switched) interrupt or IPI
500 * is illegal. However, we may have to do it anyway if we hit a fatal
501 * kernel trap or we have paniced.
503 * If this case occurs save and restore the interrupt nesting level.
505 if (gd
->gd_intr_nesting_level
) {
509 if (gd
->gd_trap_nesting_level
== 0 && panic_cpu_gd
!= mycpu
) {
510 panic("lwkt_switch: Attempt to switch from a "
511 "a fast interrupt, ipi, or hard code section, "
515 savegdnest
= gd
->gd_intr_nesting_level
;
516 savegdtrap
= gd
->gd_trap_nesting_level
;
517 gd
->gd_intr_nesting_level
= 0;
518 gd
->gd_trap_nesting_level
= 0;
519 if ((td
->td_flags
& TDF_PANICWARN
) == 0) {
520 td
->td_flags
|= TDF_PANICWARN
;
521 kprintf("Warning: thread switch from interrupt, IPI, "
522 "or hard code section.\n"
523 "thread %p (%s)\n", td
, td
->td_comm
);
527 gd
->gd_intr_nesting_level
= savegdnest
;
528 gd
->gd_trap_nesting_level
= savegdtrap
;
534 * Passive release (used to transition from user to kernel mode
535 * when we block or switch rather then when we enter the kernel).
536 * This function is NOT called if we are switching into a preemption
537 * or returning from a preemption. Typically this causes us to lose
538 * our current process designation (if we have one) and become a true
539 * LWKT thread, and may also hand the current process designation to
540 * another process and schedule thread.
546 if (TD_TOKS_HELD(td
))
547 lwkt_relalltokens(td
);
550 * We had better not be holding any spin locks, but don't get into an
551 * endless panic loop.
553 KASSERT(gd
->gd_spinlocks_wr
== 0 || panicstr
!= NULL
,
554 ("lwkt_switch: still holding %d exclusive spinlocks!",
555 gd
->gd_spinlocks_wr
));
560 * td_mpcount + td_xpcount cannot be used to determine if we currently
561 * hold the MP lock because get_mplock() will increment it prior to
562 * attempting to get the lock, and switch out if it can't. Our
563 * ownership of the actual lock will remain stable while we are
564 * in a critical section, and once we actually acquire the underlying
565 * lock as long as the count is greater than 0.
567 mpheld
= MP_LOCK_HELD(gd
);
569 if (td
->td_cscount
) {
570 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
572 if (panic_on_cscount
)
573 panic("switching while mastering cpusync");
579 * If we had preempted another thread on this cpu, resume the preempted
580 * thread. This occurs transparently, whether the preempted thread
581 * was scheduled or not (it may have been preempted after descheduling
584 * We have to setup the MP lock for the original thread after backing
585 * out the adjustment that was made to curthread when the original
588 if ((ntd
= td
->td_preempted
) != NULL
) {
589 KKASSERT(ntd
->td_flags
& TDF_PREEMPT_LOCK
);
591 if (ntd
->td_mpcount
+ ntd
->td_xpcount
&& mpheld
== 0) {
592 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
593 td
, ntd
, td
->td_mpcount
, ntd
->td_mpcount
+ ntd
->td_xpcount
);
597 ntd
->td_flags
|= TDF_PREEMPT_DONE
;
600 * The interrupt may have woken a thread up, we need to properly
601 * set the reschedule flag if the originally interrupted thread is
602 * at a lower priority.
604 if (TAILQ_FIRST(&gd
->gd_tdrunq
) &&
605 TAILQ_FIRST(&gd
->gd_tdrunq
)->td_pri
> ntd
->td_pri
) {
608 /* YYY release mp lock on switchback if original doesn't need it */
609 goto havethread_preempted
;
613 * Implement round-robin fairq with priority insertion. The priority
614 * insertion is handled by _lwkt_enqueue()
616 * We have to adjust the MP lock for the target thread. If we
617 * need the MP lock and cannot obtain it we try to locate a
618 * thread that does not need the MP lock. If we cannot, we spin
621 * A similar issue exists for the tokens held by the target thread.
622 * If we cannot obtain ownership of the tokens we cannot immediately
623 * schedule the thread.
626 clear_lwkt_resched();
628 ntd
= TAILQ_FIRST(&gd
->gd_tdrunq
);
631 * Hotpath if we can get all necessary resources.
633 * If nothing is runnable switch to the idle thread
636 ntd
= &gd
->gd_idlethread
;
637 if (gd
->gd_reqflags
& RQF_IDLECHECK_MASK
)
638 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
640 KKASSERT(ntd
->td_xpcount
== 0);
641 if (ntd
->td_mpcount
) {
642 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
)
643 panic("Idle thread %p was holding the BGL!", ntd
);
645 set_cpu_contention_mask(gd
);
646 handle_cpu_contention_mask();
648 mpheld
= MP_LOCK_HELD(gd
);
653 clr_cpu_contention_mask(gd
);
655 cpu_time
.cp_msg
[0] = 0;
656 cpu_time
.cp_stallpc
= 0;
663 * NOTE: For UP there is no mplock and lwkt_getalltokens()
666 if (ntd
->td_fairq_accum
>= 0 &&
668 (ntd
->td_mpcount
+ ntd
->td_xpcount
== 0 ||
669 mpheld
|| cpu_try_mplock()) &&
671 (!TD_TOKS_HELD(ntd
) || lwkt_getalltokens(ntd
, &lmsg
, &laddr
))
674 clr_cpu_contention_mask(gd
);
683 if (ntd
->td_fairq_accum
>= 0)
684 set_cpu_contention_mask(gd
);
685 /* Reload mpheld (it become stale after mplock/token ops) */
686 mpheld
= MP_LOCK_HELD(gd
);
687 if (ntd
->td_mpcount
+ ntd
->td_xpcount
&& mpheld
== 0) {
689 laddr
= ntd
->td_mplock_stallpc
;
694 * Coldpath - unable to schedule ntd, continue looking for threads
695 * to schedule. This is only allowed of the (presumably) kernel
696 * thread exhausted its fair share. A kernel thread stuck on
697 * resources does not currently allow a user thread to get in
701 nquserok
= ((ntd
->td_pri
< TDPRI_KERN_LPSCHED
) ||
702 (ntd
->td_fairq_accum
< 0));
710 * If the fair-share scheduler ran out ntd gets moved to the
711 * end and its accumulator will be bumped, if it didn't we
712 * maintain the same queue position.
714 * nlast keeps track of the last element prior to any moves.
716 if (ntd
->td_fairq_accum
< 0) {
717 lwkt_fairq_accumulate(gd
, ntd
);
723 xtd
= TAILQ_NEXT(ntd
, td_threadq
);
724 TAILQ_REMOVE(&gd
->gd_tdrunq
, ntd
, td_threadq
);
725 TAILQ_INSERT_TAIL(&gd
->gd_tdrunq
, ntd
, td_threadq
);
728 * Set terminal element (nlast)
737 ntd
= TAILQ_NEXT(ntd
, td_threadq
);
741 * If we exhausted the run list switch to the idle thread.
742 * Since one or more threads had resource acquisition issues
743 * we do not allow the idle thread to halt.
745 * NOTE: nlast can be NULL.
749 ntd
= &gd
->gd_idlethread
;
750 ntd
->td_flags
|= TDF_IDLE_NOHLT
;
752 KKASSERT(ntd
->td_xpcount
== 0);
753 if (ntd
->td_mpcount
) {
754 mpheld
= MP_LOCK_HELD(gd
);
755 if (gd
->gd_trap_nesting_level
== 0 && panicstr
== NULL
)
756 panic("Idle thread %p was holding the BGL!", ntd
);
758 set_cpu_contention_mask(gd
);
759 handle_cpu_contention_mask();
761 mpheld
= MP_LOCK_HELD(gd
);
763 break; /* try again from the top, almost */
769 * If fairq accumulations occured we do not schedule the
770 * idle thread. This will cause us to try again from
774 break; /* try again from the top, almost */
776 strlcpy(cpu_time
.cp_msg
, lmsg
, sizeof(cpu_time
.cp_msg
));
777 cpu_time
.cp_stallpc
= (uintptr_t)laddr
;
782 * Try to switch to this thread.
784 * NOTE: For UP there is no mplock and lwkt_getalltokens()
787 if ((ntd
->td_pri
>= TDPRI_KERN_LPSCHED
|| nquserok
||
788 user_pri_sched
) && ntd
->td_fairq_accum
>= 0 &&
790 (ntd
->td_mpcount
+ ntd
->td_xpcount
== 0 ||
791 mpheld
|| cpu_try_mplock()) &&
793 (!TD_TOKS_HELD(ntd
) || lwkt_getalltokens(ntd
, &lmsg
, &laddr
))
796 clr_cpu_contention_mask(gd
);
801 if (ntd
->td_fairq_accum
>= 0)
802 set_cpu_contention_mask(gd
);
804 * Reload mpheld (it become stale after mplock/token ops).
806 mpheld
= MP_LOCK_HELD(gd
);
807 if (ntd
->td_mpcount
+ ntd
->td_xpcount
&& mpheld
== 0) {
809 laddr
= ntd
->td_mplock_stallpc
;
811 if (ntd
->td_pri
>= TDPRI_KERN_LPSCHED
&& ntd
->td_fairq_accum
>= 0)
817 * All threads exhausted but we can loop due to a negative
820 * While we are looping in the scheduler be sure to service
821 * any interrupts which were made pending due to our critical
822 * section, otherwise we could livelock (e.g.) IPIs.
824 * NOTE: splz can enter and exit the mplock so mpheld is
825 * stale after this call.
831 * Our mplock can be cached and cause other cpus to livelock
832 * if we loop due to e.g. not being able to acquire tokens.
834 if (MP_LOCK_HELD(gd
))
835 cpu_rel_mplock(gd
->gd_cpuid
);
841 * Do the actual switch. WARNING: mpheld is stale here.
843 * We must always decrement td_fairq_accum on non-idle threads just
844 * in case a thread never gets a tick due to being in a continuous
845 * critical section. The page-zeroing code does that.
847 * If the thread we came up with is a higher or equal priority verses
848 * the thread at the head of the queue we move our thread to the
849 * front. This way we can always check the front of the queue.
852 ++gd
->gd_cnt
.v_swtch
;
853 --ntd
->td_fairq_accum
;
854 xtd
= TAILQ_FIRST(&gd
->gd_tdrunq
);
855 if (ntd
!= xtd
&& ntd
->td_pri
>= xtd
->td_pri
) {
856 TAILQ_REMOVE(&gd
->gd_tdrunq
, ntd
, td_threadq
);
857 TAILQ_INSERT_HEAD(&gd
->gd_tdrunq
, ntd
, td_threadq
);
859 havethread_preempted
:
862 * If the new target does not need the MP lock and we are holding it,
863 * release the MP lock. If the new target requires the MP lock we have
864 * already acquired it for the target.
866 * WARNING: mpheld is stale here.
869 KASSERT(ntd
->td_critcount
,
870 ("priority problem in lwkt_switch %d %d", td
->td_pri
, ntd
->td_pri
));
872 if (ntd
->td_mpcount
+ ntd
->td_xpcount
== 0 ) {
873 if (MP_LOCK_HELD(gd
))
874 cpu_rel_mplock(gd
->gd_cpuid
);
876 ASSERT_MP_LOCK_HELD(ntd
);
881 KTR_LOG(ctxsw_sw
, gd
->gd_cpuid
, ntd
);
884 /* NOTE: current cpu may have changed after switch */
889 * Request that the target thread preempt the current thread. Preemption
890 * only works under a specific set of conditions:
892 * - We are not preempting ourselves
893 * - The target thread is owned by the current cpu
894 * - We are not currently being preempted
895 * - The target is not currently being preempted
896 * - We are not holding any spin locks
897 * - The target thread is not holding any tokens
898 * - We are able to satisfy the target's MP lock requirements (if any).
900 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
901 * this is called via lwkt_schedule() through the td_preemptable callback.
902 * critcount is the managed critical priority that we should ignore in order
903 * to determine whether preemption is possible (aka usually just the crit
904 * priority of lwkt_schedule() itself).
906 * XXX at the moment we run the target thread in a critical section during
907 * the preemption in order to prevent the target from taking interrupts
908 * that *WE* can't. Preemption is strictly limited to interrupt threads
909 * and interrupt-like threads, outside of a critical section, and the
910 * preempted source thread will be resumed the instant the target blocks
911 * whether or not the source is scheduled (i.e. preemption is supposed to
912 * be as transparent as possible).
914 * The target thread inherits our MP count (added to its own) for the
915 * duration of the preemption in order to preserve the atomicy of the
916 * MP lock during the preemption. Therefore, any preempting targets must be
917 * careful in regards to MP assertions. Note that the MP count may be
918 * out of sync with the physical mp_lock, but we do not have to preserve
919 * the original ownership of the lock if it was out of synch (that is, we
920 * can leave it synchronized on return).
923 lwkt_preempt(thread_t ntd
, int critcount
)
925 struct globaldata
*gd
= mycpu
;
933 * The caller has put us in a critical section. We can only preempt
934 * if the caller of the caller was not in a critical section (basically
935 * a local interrupt), as determined by the 'critcount' parameter. We
936 * also can't preempt if the caller is holding any spinlocks (even if
937 * he isn't in a critical section). This also handles the tokens test.
939 * YYY The target thread must be in a critical section (else it must
940 * inherit our critical section? I dunno yet).
942 * Set need_lwkt_resched() unconditionally for now YYY.
944 KASSERT(ntd
->td_critcount
, ("BADCRIT0 %d", ntd
->td_pri
));
946 if (preempt_enable
== 0) {
951 td
= gd
->gd_curthread
;
952 if (ntd
->td_pri
<= td
->td_pri
) {
956 if (td
->td_critcount
> critcount
) {
962 if (ntd
->td_gd
!= gd
) {
969 * We don't have to check spinlocks here as they will also bump
972 * Do not try to preempt if the target thread is holding any tokens.
973 * We could try to acquire the tokens but this case is so rare there
974 * is no need to support it.
976 KKASSERT(gd
->gd_spinlocks_wr
== 0);
978 if (TD_TOKS_HELD(ntd
)) {
983 if (td
== ntd
|| ((td
->td_flags
| ntd
->td_flags
) & TDF_PREEMPT_LOCK
)) {
988 if (ntd
->td_preempted
) {
995 * NOTE: An interrupt might have occured just as we were transitioning
996 * to or from the MP lock. In this case td_mpcount will be pre-disposed
997 * (non-zero) but not actually synchronized with the mp_lock itself.
998 * We can use it to imply an MP lock requirement for the preemption but
999 * we cannot use it to test whether we hold the MP lock or not.
1001 savecnt
= td
->td_mpcount
;
1002 mpheld
= MP_LOCK_HELD(gd
);
1003 ntd
->td_xpcount
= td
->td_mpcount
+ td
->td_xpcount
;
1004 if (mpheld
== 0 && ntd
->td_mpcount
+ ntd
->td_xpcount
&& !cpu_try_mplock()) {
1005 ntd
->td_xpcount
= 0;
1007 need_lwkt_resched();
1013 * Since we are able to preempt the current thread, there is no need to
1014 * call need_lwkt_resched().
1017 ntd
->td_preempted
= td
;
1018 td
->td_flags
|= TDF_PREEMPT_LOCK
;
1019 KTR_LOG(ctxsw_pre
, gd
->gd_cpuid
, ntd
);
1022 KKASSERT(ntd
->td_preempted
&& (td
->td_flags
& TDF_PREEMPT_DONE
));
1024 KKASSERT(savecnt
== td
->td_mpcount
);
1025 mpheld
= MP_LOCK_HELD(gd
);
1026 if (mpheld
&& td
->td_mpcount
== 0)
1027 cpu_rel_mplock(gd
->gd_cpuid
);
1028 else if (mpheld
== 0 && td
->td_mpcount
+ td
->td_xpcount
)
1029 panic("lwkt_preempt(): MP lock was not held through");
1031 ntd
->td_preempted
= NULL
;
1032 td
->td_flags
&= ~(TDF_PREEMPT_LOCK
|TDF_PREEMPT_DONE
);
1036 * Conditionally call splz() if gd_reqflags indicates work is pending.
1037 * This will work inside a critical section but not inside a hard code
1040 * (self contained on a per cpu basis)
1045 globaldata_t gd
= mycpu
;
1046 thread_t td
= gd
->gd_curthread
;
1048 if ((gd
->gd_reqflags
& RQF_IDLECHECK_MASK
) &&
1049 gd
->gd_intr_nesting_level
== 0 &&
1050 td
->td_nest_count
< 2)
1057 * This version is integrated into crit_exit, reqflags has already
1058 * been tested but td_critcount has not.
1060 * We only want to execute the splz() on the 1->0 transition of
1061 * critcount and not in a hard code section or if too deeply nested.
1064 lwkt_maybe_splz(thread_t td
)
1066 globaldata_t gd
= td
->td_gd
;
1068 if (td
->td_critcount
== 0 &&
1069 gd
->gd_intr_nesting_level
== 0 &&
1070 td
->td_nest_count
< 2)
1077 * This function is used to negotiate a passive release of the current
1078 * process/lwp designation with the user scheduler, allowing the user
1079 * scheduler to schedule another user thread. The related kernel thread
1080 * (curthread) continues running in the released state.
1083 lwkt_passive_release(struct thread
*td
)
1085 struct lwp
*lp
= td
->td_lwp
;
1087 td
->td_release
= NULL
;
1088 lwkt_setpri_self(TDPRI_KERN_USER
);
1089 lp
->lwp_proc
->p_usched
->release_curproc(lp
);
1094 * This implements a normal yield. This routine is virtually a nop if
1095 * there is nothing to yield to but it will always run any pending interrupts
1096 * if called from a critical section.
1098 * This yield is designed for kernel threads without a user context.
1100 * (self contained on a per cpu basis)
1105 globaldata_t gd
= mycpu
;
1106 thread_t td
= gd
->gd_curthread
;
1109 if ((gd
->gd_reqflags
& RQF_IDLECHECK_MASK
) && td
->td_nest_count
< 2)
1111 if (td
->td_fairq_accum
< 0) {
1112 lwkt_schedule_self(curthread
);
1115 xtd
= TAILQ_FIRST(&gd
->gd_tdrunq
);
1116 if (xtd
&& xtd
->td_pri
> td
->td_pri
) {
1117 lwkt_schedule_self(curthread
);
1124 * This yield is designed for kernel threads with a user context.
1126 * The kernel acting on behalf of the user is potentially cpu-bound,
1127 * this function will efficiently allow other threads to run and also
1128 * switch to other processes by releasing.
1130 * The lwkt_user_yield() function is designed to have very low overhead
1131 * if no yield is determined to be needed.
1134 lwkt_user_yield(void)
1136 globaldata_t gd
= mycpu
;
1137 thread_t td
= gd
->gd_curthread
;
1140 * Always run any pending interrupts in case we are in a critical
1143 if ((gd
->gd_reqflags
& RQF_IDLECHECK_MASK
) && td
->td_nest_count
< 2)
1148 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1149 * kernel can prevent other cpus from servicing interrupt threads
1150 * which still require the MP lock (which is a lot of them). This
1151 * has a chaining effect since if the interrupt is blocked, so is
1152 * the event, so normal scheduling will not pick up on the problem.
1154 if (cpu_contention_mask
&& td
->td_mpcount
+ td
->td_xpcount
) {
1160 * Switch (which forces a release) if another kernel thread needs
1161 * the cpu, if userland wants us to resched, or if our kernel
1162 * quantum has run out.
1164 if (lwkt_resched_wanted() ||
1165 user_resched_wanted() ||
1166 td
->td_fairq_accum
< 0)
1173 * Reacquire the current process if we are released.
1175 * XXX not implemented atm. The kernel may be holding locks and such,
1176 * so we want the thread to continue to receive cpu.
1178 if (td
->td_release
== NULL
&& lp
) {
1179 lp
->lwp_proc
->p_usched
->acquire_curproc(lp
);
1180 td
->td_release
= lwkt_passive_release
;
1181 lwkt_setpri_self(TDPRI_USER_NORM
);
1187 * Generic schedule. Possibly schedule threads belonging to other cpus and
1188 * deal with threads that might be blocked on a wait queue.
1190 * We have a little helper inline function which does additional work after
1191 * the thread has been enqueued, including dealing with preemption and
1192 * setting need_lwkt_resched() (which prevents the kernel from returning
1193 * to userland until it has processed higher priority threads).
1195 * It is possible for this routine to be called after a failed _enqueue
1196 * (due to the target thread migrating, sleeping, or otherwise blocked).
1197 * We have to check that the thread is actually on the run queue!
1199 * reschedok is an optimized constant propagated from lwkt_schedule() or
1200 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1201 * reschedule to be requested if the target thread has a higher priority.
1202 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1203 * be 0, prevented undesired reschedules.
1207 _lwkt_schedule_post(globaldata_t gd
, thread_t ntd
, int ccount
, int reschedok
)
1211 if (ntd
->td_flags
& TDF_RUNQ
) {
1212 if (ntd
->td_preemptable
&& reschedok
) {
1213 ntd
->td_preemptable(ntd
, ccount
); /* YYY +token */
1214 } else if (reschedok
) {
1216 if (ntd
->td_pri
> otd
->td_pri
)
1217 need_lwkt_resched();
1221 * Give the thread a little fair share scheduler bump if it
1222 * has been asleep for a while. This is primarily to avoid
1223 * a degenerate case for interrupt threads where accumulator
1224 * crosses into negative territory unnecessarily.
1226 if (ntd
->td_fairq_lticks
!= ticks
) {
1227 ntd
->td_fairq_lticks
= ticks
;
1228 ntd
->td_fairq_accum
+= gd
->gd_fairq_total_pri
;
1229 if (ntd
->td_fairq_accum
> TDFAIRQ_MAX(gd
))
1230 ntd
->td_fairq_accum
= TDFAIRQ_MAX(gd
);
1237 _lwkt_schedule(thread_t td
, int reschedok
)
1239 globaldata_t mygd
= mycpu
;
1241 KASSERT(td
!= &td
->td_gd
->gd_idlethread
,
1242 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1243 crit_enter_gd(mygd
);
1244 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1245 if (td
== mygd
->gd_curthread
) {
1249 * If we own the thread, there is no race (since we are in a
1250 * critical section). If we do not own the thread there might
1251 * be a race but the target cpu will deal with it.
1254 if (td
->td_gd
== mygd
) {
1256 _lwkt_schedule_post(mygd
, td
, 1, reschedok
);
1258 lwkt_send_ipiq3(td
->td_gd
, lwkt_schedule_remote
, td
, 0);
1262 _lwkt_schedule_post(mygd
, td
, 1, reschedok
);
1269 lwkt_schedule(thread_t td
)
1271 _lwkt_schedule(td
, 1);
1275 lwkt_schedule_noresched(thread_t td
)
1277 _lwkt_schedule(td
, 0);
1283 * When scheduled remotely if frame != NULL the IPIQ is being
1284 * run via doreti or an interrupt then preemption can be allowed.
1286 * To allow preemption we have to drop the critical section so only
1287 * one is present in _lwkt_schedule_post.
1290 lwkt_schedule_remote(void *arg
, int arg2
, struct intrframe
*frame
)
1292 thread_t td
= curthread
;
1295 if (frame
&& ntd
->td_preemptable
) {
1296 crit_exit_noyield(td
);
1297 _lwkt_schedule(ntd
, 1);
1298 crit_enter_quick(td
);
1300 _lwkt_schedule(ntd
, 1);
1305 * Thread migration using a 'Pull' method. The thread may or may not be
1306 * the current thread. It MUST be descheduled and in a stable state.
1307 * lwkt_giveaway() must be called on the cpu owning the thread.
1309 * At any point after lwkt_giveaway() is called, the target cpu may
1310 * 'pull' the thread by calling lwkt_acquire().
1312 * We have to make sure the thread is not sitting on a per-cpu tsleep
1313 * queue or it will blow up when it moves to another cpu.
1315 * MPSAFE - must be called under very specific conditions.
1318 lwkt_giveaway(thread_t td
)
1320 globaldata_t gd
= mycpu
;
1323 if (td
->td_flags
& TDF_TSLEEPQ
)
1325 KKASSERT(td
->td_gd
== gd
);
1326 TAILQ_REMOVE(&gd
->gd_tdallq
, td
, td_allq
);
1327 td
->td_flags
|= TDF_MIGRATING
;
1332 lwkt_acquire(thread_t td
)
1337 KKASSERT(td
->td_flags
& TDF_MIGRATING
);
1342 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1343 crit_enter_gd(mygd
);
1344 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1346 lwkt_process_ipiq();
1352 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1353 td
->td_flags
&= ~TDF_MIGRATING
;
1356 crit_enter_gd(mygd
);
1357 TAILQ_INSERT_TAIL(&mygd
->gd_tdallq
, td
, td_allq
);
1358 td
->td_flags
&= ~TDF_MIGRATING
;
1366 * Generic deschedule. Descheduling threads other then your own should be
1367 * done only in carefully controlled circumstances. Descheduling is
1370 * This function may block if the cpu has run out of messages.
1373 lwkt_deschedule(thread_t td
)
1377 if (td
== curthread
) {
1380 if (td
->td_gd
== mycpu
) {
1383 lwkt_send_ipiq(td
->td_gd
, (ipifunc1_t
)lwkt_deschedule
, td
);
1393 * Set the target thread's priority. This routine does not automatically
1394 * switch to a higher priority thread, LWKT threads are not designed for
1395 * continuous priority changes. Yield if you want to switch.
1398 lwkt_setpri(thread_t td
, int pri
)
1400 KKASSERT(td
->td_gd
== mycpu
);
1401 if (td
->td_pri
!= pri
) {
1404 if (td
->td_flags
& TDF_RUNQ
) {
1416 * Set the initial priority for a thread prior to it being scheduled for
1417 * the first time. The thread MUST NOT be scheduled before or during
1418 * this call. The thread may be assigned to a cpu other then the current
1421 * Typically used after a thread has been created with TDF_STOPPREQ,
1422 * and before the thread is initially scheduled.
1425 lwkt_setpri_initial(thread_t td
, int pri
)
1428 KKASSERT((td
->td_flags
& TDF_RUNQ
) == 0);
1433 lwkt_setpri_self(int pri
)
1435 thread_t td
= curthread
;
1437 KKASSERT(pri
>= 0 && pri
<= TDPRI_MAX
);
1439 if (td
->td_flags
& TDF_RUNQ
) {
1450 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1452 * Example: two competing threads, same priority N. decrement by (2*N)
1453 * increment by N*8, each thread will get 4 ticks.
1456 lwkt_fairq_schedulerclock(thread_t td
)
1460 if (td
!= &td
->td_gd
->gd_idlethread
) {
1461 td
->td_fairq_accum
-= td
->td_gd
->gd_fairq_total_pri
;
1462 if (td
->td_fairq_accum
< -TDFAIRQ_MAX(td
->td_gd
))
1463 td
->td_fairq_accum
= -TDFAIRQ_MAX(td
->td_gd
);
1464 if (td
->td_fairq_accum
< 0)
1465 need_lwkt_resched();
1466 td
->td_fairq_lticks
= ticks
;
1468 td
= td
->td_preempted
;
1474 lwkt_fairq_accumulate(globaldata_t gd
, thread_t td
)
1476 td
->td_fairq_accum
+= td
->td_pri
* TDFAIRQ_SCALE
;
1477 if (td
->td_fairq_accum
> TDFAIRQ_MAX(td
->td_gd
))
1478 td
->td_fairq_accum
= TDFAIRQ_MAX(td
->td_gd
);
1482 * Migrate the current thread to the specified cpu.
1484 * This is accomplished by descheduling ourselves from the current cpu,
1485 * moving our thread to the tdallq of the target cpu, IPI messaging the
1486 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1487 * races while the thread is being migrated.
1489 * We must be sure to remove ourselves from the current cpu's tsleepq
1490 * before potentially moving to another queue. The thread can be on
1491 * a tsleepq due to a left-over tsleep_interlock().
1494 static void lwkt_setcpu_remote(void *arg
);
1498 lwkt_setcpu_self(globaldata_t rgd
)
1501 thread_t td
= curthread
;
1503 if (td
->td_gd
!= rgd
) {
1504 crit_enter_quick(td
);
1505 if (td
->td_flags
& TDF_TSLEEPQ
)
1507 td
->td_flags
|= TDF_MIGRATING
;
1508 lwkt_deschedule_self(td
);
1509 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1510 lwkt_send_ipiq(rgd
, (ipifunc1_t
)lwkt_setcpu_remote
, td
);
1512 /* we are now on the target cpu */
1513 TAILQ_INSERT_TAIL(&rgd
->gd_tdallq
, td
, td_allq
);
1514 crit_exit_quick(td
);
1520 lwkt_migratecpu(int cpuid
)
1525 rgd
= globaldata_find(cpuid
);
1526 lwkt_setcpu_self(rgd
);
1531 * Remote IPI for cpu migration (called while in a critical section so we
1532 * do not have to enter another one). The thread has already been moved to
1533 * our cpu's allq, but we must wait for the thread to be completely switched
1534 * out on the originating cpu before we schedule it on ours or the stack
1535 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1536 * change to main memory.
1538 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1539 * against wakeups. It is best if this interface is used only when there
1540 * are no pending events that might try to schedule the thread.
1544 lwkt_setcpu_remote(void *arg
)
1547 globaldata_t gd
= mycpu
;
1549 while (td
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) {
1551 lwkt_process_ipiq();
1558 td
->td_flags
&= ~TDF_MIGRATING
;
1559 KKASSERT(td
->td_lwp
== NULL
|| (td
->td_lwp
->lwp_flag
& LWP_ONRUNQ
) == 0);
1565 lwkt_preempted_proc(void)
1567 thread_t td
= curthread
;
1568 while (td
->td_preempted
)
1569 td
= td
->td_preempted
;
1574 * Create a kernel process/thread/whatever. It shares it's address space
1575 * with proc0 - ie: kernel only.
1577 * NOTE! By default new threads are created with the MP lock held. A
1578 * thread which does not require the MP lock should release it by calling
1579 * rel_mplock() at the start of the new thread.
1582 lwkt_create(void (*func
)(void *), void *arg
, struct thread
**tdp
,
1583 thread_t
template, int tdflags
, int cpu
, const char *fmt
, ...)
1588 td
= lwkt_alloc_thread(template, LWKT_THREAD_STACK
, cpu
,
1592 cpu_set_thread_handler(td
, lwkt_exit
, func
, arg
);
1595 * Set up arg0 for 'ps' etc
1597 __va_start(ap
, fmt
);
1598 kvsnprintf(td
->td_comm
, sizeof(td
->td_comm
), fmt
, ap
);
1602 * Schedule the thread to run
1604 if ((td
->td_flags
& TDF_STOPREQ
) == 0)
1607 td
->td_flags
&= ~TDF_STOPREQ
;
1612 * Destroy an LWKT thread. Warning! This function is not called when
1613 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1614 * uses a different reaping mechanism.
1619 thread_t td
= curthread
;
1624 * Do any cleanup that might block here
1626 if (td
->td_flags
& TDF_VERBOSE
)
1627 kprintf("kthread %p %s has exited\n", td
, td
->td_comm
);
1630 dsched_exit_thread(td
);
1633 * Get us into a critical section to interlock gd_freetd and loop
1634 * until we can get it freed.
1636 * We have to cache the current td in gd_freetd because objcache_put()ing
1637 * it would rip it out from under us while our thread is still active.
1640 crit_enter_quick(td
);
1641 while ((std
= gd
->gd_freetd
) != NULL
) {
1642 KKASSERT((std
->td_flags
& (TDF_RUNNING
|TDF_PREEMPT_LOCK
)) == 0);
1643 gd
->gd_freetd
= NULL
;
1644 objcache_put(thread_cache
, std
);
1648 * Remove thread resources from kernel lists and deschedule us for
1649 * the last time. We cannot block after this point or we may end
1650 * up with a stale td on the tsleepq.
1652 if (td
->td_flags
& TDF_TSLEEPQ
)
1654 lwkt_deschedule_self(td
);
1655 lwkt_remove_tdallq(td
);
1660 KKASSERT(gd
->gd_freetd
== NULL
);
1661 if (td
->td_flags
& TDF_ALLOCATED_THREAD
)
1667 lwkt_remove_tdallq(thread_t td
)
1669 KKASSERT(td
->td_gd
== mycpu
);
1670 TAILQ_REMOVE(&td
->td_gd
->gd_tdallq
, td
, td_allq
);
1674 * Code reduction and branch prediction improvements. Call/return
1675 * overhead on modern cpus often degenerates into 0 cycles due to
1676 * the cpu's branch prediction hardware and return pc cache. We
1677 * can take advantage of this by not inlining medium-complexity
1678 * functions and we can also reduce the branch prediction impact
1679 * by collapsing perfectly predictable branches into a single
1680 * procedure instead of duplicating it.
1682 * Is any of this noticeable? Probably not, so I'll take the
1683 * smaller code size.
1686 crit_exit_wrapper(__DEBUG_CRIT_ARG__
)
1688 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__
);
1694 thread_t td
= curthread
;
1695 int lcrit
= td
->td_critcount
;
1697 td
->td_critcount
= 0;
1698 panic("td_critcount is/would-go negative! %p %d", td
, lcrit
);
1705 * Called from debugger/panic on cpus which have been stopped. We must still
1706 * process the IPIQ while stopped, even if we were stopped while in a critical
1709 * If we are dumping also try to process any pending interrupts. This may
1710 * or may not work depending on the state of the cpu at the point it was
1714 lwkt_smp_stopped(void)
1716 globaldata_t gd
= mycpu
;
1720 lwkt_process_ipiq();
1723 lwkt_process_ipiq();