2 * Copyright (c) 2003-2011 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>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
58 #include <sys/mplock2.h>
60 #include <sys/dsched.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_extern.h>
71 #include <machine/stdarg.h>
72 #include <machine/smp.h>
74 #if !defined(KTR_CTXSW)
75 #define KTR_CTXSW KTR_ALL
77 KTR_INFO_MASTER(ctxsw);
78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
79 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
80 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
81 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
83 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
86 static int panic_on_cscount = 0;
88 static __int64_t switch_count = 0;
89 static __int64_t preempt_hit = 0;
90 static __int64_t preempt_miss = 0;
91 static __int64_t preempt_weird = 0;
92 static int lwkt_use_spin_port;
93 static struct objcache *thread_cache;
94 int cpu_mwait_spin = 0;
96 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
97 static void lwkt_setcpu_remote(void *arg);
99 extern void cpu_heavy_restore(void);
100 extern void cpu_lwkt_restore(void);
101 extern void cpu_kthread_restore(void);
102 extern void cpu_idle_restore(void);
105 * We can make all thread ports use the spin backend instead of the thread
106 * backend. This should only be set to debug the spin backend.
108 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
111 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
112 "Panic if attempting to switch lwkt's while mastering cpusync");
114 SYSCTL_INT(_hw, OID_AUTO, cpu_mwait_spin, CTLFLAG_RW, &cpu_mwait_spin, 0,
115 "monitor/mwait target state");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
117 "Number of switched threads");
118 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
119 "Successful preemption events");
120 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
121 "Failed preemption events");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
123 "Number of preempted threads.");
124 static int fairq_enable = 0;
125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
126 &fairq_enable, 0, "Turn on fairq priority accumulators");
127 static int fairq_bypass = -1;
128 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
129 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
130 extern int lwkt_sched_debug;
131 int lwkt_sched_debug = 0;
132 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
133 &lwkt_sched_debug, 0, "Scheduler debug");
134 static int lwkt_spin_loops = 10;
135 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
136 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
137 static int lwkt_spin_reseq = 0;
138 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
139 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
140 static int lwkt_spin_monitor = 0;
141 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
142 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
143 static int lwkt_spin_fatal = 0; /* disabled */
144 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
145 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
146 static int preempt_enable = 1;
147 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
148 &preempt_enable, 0, "Enable preemption");
149 static int lwkt_cache_threads = 0;
150 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
151 &lwkt_cache_threads, 0, "thread+kstack cache");
153 static __cachealign int lwkt_cseq_rindex;
154 static __cachealign int lwkt_cseq_windex;
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
167 _lwkt_dequeue(thread_t td)
169 if (td->td_flags & TDF_RUNQ) {
170 struct globaldata *gd = td->td_gd;
172 td->td_flags &= ~TDF_RUNQ;
173 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
174 --gd->gd_tdrunqcount;
175 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
176 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
183 * There are a limited number of lwkt threads runnable since user
184 * processes only schedule one at a time per cpu. However, there can
185 * be many user processes in kernel mode exiting from a tsleep() which
188 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
189 * will ignore user priority. This is to ensure that user threads in
190 * kernel mode get cpu at some point regardless of what the user
195 _lwkt_enqueue(thread_t td)
199 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
200 struct globaldata *gd = td->td_gd;
202 td->td_flags |= TDF_RUNQ;
203 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
205 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
209 * NOTE: td_upri - higher numbers more desireable, same sense
210 * as td_pri (typically reversed from lwp_upri).
212 * In the equal priority case we want the best selection
213 * at the beginning so the less desireable selections know
214 * that they have to setrunqueue/go-to-another-cpu, even
215 * though it means switching back to the 'best' selection.
216 * This also avoids degenerate situations when many threads
217 * are runnable or waking up at the same time.
219 * If upri matches exactly place at end/round-robin.
222 (xtd->td_pri >= td->td_pri ||
223 (xtd->td_pri == td->td_pri &&
224 xtd->td_upri >= td->td_upri))) {
225 xtd = TAILQ_NEXT(xtd, td_threadq);
228 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
230 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
232 ++gd->gd_tdrunqcount;
235 * Request a LWKT reschedule if we are now at the head of the queue.
237 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
243 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
245 struct thread *td = (struct thread *)obj;
247 td->td_kstack = NULL;
248 td->td_kstack_size = 0;
249 td->td_flags = TDF_ALLOCATED_THREAD;
255 _lwkt_thread_dtor(void *obj, void *privdata)
257 struct thread *td = (struct thread *)obj;
259 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
260 ("_lwkt_thread_dtor: not allocated from objcache"));
261 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
262 td->td_kstack_size > 0,
263 ("_lwkt_thread_dtor: corrupted stack"));
264 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
265 td->td_kstack = NULL;
270 * Initialize the lwkt s/system.
272 * Nominally cache up to 32 thread + kstack structures. Cache more on
273 * systems with a lot of cpu cores.
278 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
279 if (lwkt_cache_threads == 0) {
280 lwkt_cache_threads = ncpus * 4;
281 if (lwkt_cache_threads < 32)
282 lwkt_cache_threads = 32;
284 thread_cache = objcache_create_mbacked(
285 M_THREAD, sizeof(struct thread),
286 0, lwkt_cache_threads,
287 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
291 * Schedule a thread to run. As the current thread we can always safely
292 * schedule ourselves, and a shortcut procedure is provided for that
295 * (non-blocking, self contained on a per cpu basis)
298 lwkt_schedule_self(thread_t td)
300 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
301 crit_enter_quick(td);
302 KASSERT(td != &td->td_gd->gd_idlethread,
303 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
304 KKASSERT(td->td_lwp == NULL ||
305 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
311 * Deschedule a thread.
313 * (non-blocking, self contained on a per cpu basis)
316 lwkt_deschedule_self(thread_t td)
318 crit_enter_quick(td);
324 * LWKTs operate on a per-cpu basis
326 * WARNING! Called from early boot, 'mycpu' may not work yet.
329 lwkt_gdinit(struct globaldata *gd)
331 TAILQ_INIT(&gd->gd_tdrunq);
332 TAILQ_INIT(&gd->gd_tdallq);
336 * Create a new thread. The thread must be associated with a process context
337 * or LWKT start address before it can be scheduled. If the target cpu is
338 * -1 the thread will be created on the current cpu.
340 * If you intend to create a thread without a process context this function
341 * does everything except load the startup and switcher function.
344 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
346 static int cpu_rotator;
347 globaldata_t gd = mycpu;
351 * If static thread storage is not supplied allocate a thread. Reuse
352 * a cached free thread if possible. gd_freetd is used to keep an exiting
353 * thread intact through the exit.
357 if ((td = gd->gd_freetd) != NULL) {
358 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
360 gd->gd_freetd = NULL;
362 td = objcache_get(thread_cache, M_WAITOK);
363 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
367 KASSERT((td->td_flags &
368 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
369 TDF_ALLOCATED_THREAD,
370 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
371 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
375 * Try to reuse cached stack.
377 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
378 if (flags & TDF_ALLOCATED_STACK) {
379 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
384 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
385 flags |= TDF_ALLOCATED_STACK;
392 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
397 * Initialize a preexisting thread structure. This function is used by
398 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
400 * All threads start out in a critical section at a priority of
401 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
402 * appropriate. This function may send an IPI message when the
403 * requested cpu is not the current cpu and consequently gd_tdallq may
404 * not be initialized synchronously from the point of view of the originating
407 * NOTE! we have to be careful in regards to creating threads for other cpus
408 * if SMP has not yet been activated.
411 lwkt_init_thread_remote(void *arg)
416 * Protected by critical section held by IPI dispatch
418 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
422 * lwkt core thread structural initialization.
424 * NOTE: All threads are initialized as mpsafe threads.
427 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
428 struct globaldata *gd)
430 globaldata_t mygd = mycpu;
432 bzero(td, sizeof(struct thread));
433 td->td_kstack = stack;
434 td->td_kstack_size = stksize;
435 td->td_flags = flags;
437 td->td_type = TD_TYPE_GENERIC;
439 td->td_pri = TDPRI_KERN_DAEMON;
440 td->td_critcount = 1;
441 td->td_toks_have = NULL;
442 td->td_toks_stop = &td->td_toks_base;
443 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
444 lwkt_initport_spin(&td->td_msgport, td);
446 lwkt_initport_thread(&td->td_msgport, td);
447 pmap_init_thread(td);
449 * Normally initializing a thread for a remote cpu requires sending an
450 * IPI. However, the idlethread is setup before the other cpus are
451 * activated so we have to treat it as a special case. XXX manipulation
452 * of gd_tdallq requires the BGL.
454 if (gd == mygd || td == &gd->gd_idlethread) {
456 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
459 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
461 dsched_new_thread(td);
465 lwkt_set_comm(thread_t td, const char *ctl, ...)
470 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
472 KTR_LOG(ctxsw_newtd, td, td->td_comm);
476 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
477 * this does not prevent the thread from migrating to another cpu so the
478 * gd_tdallq state is not protected by this.
481 lwkt_hold(thread_t td)
483 atomic_add_int(&td->td_refs, 1);
487 lwkt_rele(thread_t td)
489 KKASSERT(td->td_refs > 0);
490 atomic_add_int(&td->td_refs, -1);
494 lwkt_free_thread(thread_t td)
496 KKASSERT(td->td_refs == 0);
497 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
498 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
499 if (td->td_flags & TDF_ALLOCATED_THREAD) {
500 objcache_put(thread_cache, td);
501 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
502 /* client-allocated struct with internally allocated stack */
503 KASSERT(td->td_kstack && td->td_kstack_size > 0,
504 ("lwkt_free_thread: corrupted stack"));
505 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
506 td->td_kstack = NULL;
507 td->td_kstack_size = 0;
509 KTR_LOG(ctxsw_deadtd, td);
514 * Switch to the next runnable lwkt. If no LWKTs are runnable then
515 * switch to the idlethread. Switching must occur within a critical
516 * section to avoid races with the scheduling queue.
518 * We always have full control over our cpu's run queue. Other cpus
519 * that wish to manipulate our queue must use the cpu_*msg() calls to
520 * talk to our cpu, so a critical section is all that is needed and
521 * the result is very, very fast thread switching.
523 * The LWKT scheduler uses a fixed priority model and round-robins at
524 * each priority level. User process scheduling is a totally
525 * different beast and LWKT priorities should not be confused with
526 * user process priorities.
528 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
529 * is not called by the current thread in the preemption case, only when
530 * the preempting thread blocks (in order to return to the original thread).
532 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
533 * migration and tsleep deschedule the current lwkt thread and call
534 * lwkt_switch(). In particular, the target cpu of the migration fully
535 * expects the thread to become non-runnable and can deadlock against
536 * cpusync operations if we run any IPIs prior to switching the thread out.
538 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
539 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
544 globaldata_t gd = mycpu;
545 thread_t td = gd->gd_curthread;
549 KKASSERT(gd->gd_processing_ipiq == 0);
550 KKASSERT(td->td_flags & TDF_RUNNING);
553 * Switching from within a 'fast' (non thread switched) interrupt or IPI
554 * is illegal. However, we may have to do it anyway if we hit a fatal
555 * kernel trap or we have paniced.
557 * If this case occurs save and restore the interrupt nesting level.
559 if (gd->gd_intr_nesting_level) {
563 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
564 panic("lwkt_switch: Attempt to switch from a "
565 "fast interrupt, ipi, or hard code section, "
569 savegdnest = gd->gd_intr_nesting_level;
570 savegdtrap = gd->gd_trap_nesting_level;
571 gd->gd_intr_nesting_level = 0;
572 gd->gd_trap_nesting_level = 0;
573 if ((td->td_flags & TDF_PANICWARN) == 0) {
574 td->td_flags |= TDF_PANICWARN;
575 kprintf("Warning: thread switch from interrupt, IPI, "
576 "or hard code section.\n"
577 "thread %p (%s)\n", td, td->td_comm);
581 gd->gd_intr_nesting_level = savegdnest;
582 gd->gd_trap_nesting_level = savegdtrap;
588 * Release our current user process designation if we are blocking
589 * or if a user reschedule was requested.
591 * NOTE: This function is NOT called if we are switching into or
592 * returning from a preemption.
594 * NOTE: Releasing our current user process designation may cause
595 * it to be assigned to another thread, which in turn will
596 * cause us to block in the usched acquire code when we attempt
597 * to return to userland.
599 * NOTE: On SMP systems this can be very nasty when heavy token
600 * contention is present so we want to be careful not to
601 * release the designation gratuitously.
603 if (td->td_release &&
604 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
612 if (TD_TOKS_HELD(td))
613 lwkt_relalltokens(td);
616 * We had better not be holding any spin locks, but don't get into an
617 * endless panic loop.
619 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
620 ("lwkt_switch: still holding %d exclusive spinlocks!",
625 if (td->td_cscount) {
626 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
628 if (panic_on_cscount)
629 panic("switching while mastering cpusync");
634 * If we had preempted another thread on this cpu, resume the preempted
635 * thread. This occurs transparently, whether the preempted thread
636 * was scheduled or not (it may have been preempted after descheduling
639 * We have to setup the MP lock for the original thread after backing
640 * out the adjustment that was made to curthread when the original
643 if ((ntd = td->td_preempted) != NULL) {
644 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
645 ntd->td_flags |= TDF_PREEMPT_DONE;
648 * The interrupt may have woken a thread up, we need to properly
649 * set the reschedule flag if the originally interrupted thread is
650 * at a lower priority.
652 * The interrupt may not have descheduled.
654 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
656 goto havethread_preempted;
660 * If we cannot obtain ownership of the tokens we cannot immediately
661 * schedule the target thread.
663 * Reminder: Again, we cannot afford to run any IPIs in this path if
664 * the current thread has been descheduled.
667 clear_lwkt_resched();
670 * Hotpath - pull the head of the run queue and attempt to schedule
673 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
677 * Runq is empty, switch to idle to allow it to halt.
679 ntd = &gd->gd_idlethread;
680 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
681 ASSERT_NO_TOKENS_HELD(ntd);
682 cpu_time.cp_msg[0] = 0;
683 cpu_time.cp_stallpc = 0;
688 * Hotpath - schedule ntd.
690 * NOTE: For UP there is no mplock and lwkt_getalltokens()
693 if (TD_TOKS_NOT_HELD(ntd) ||
694 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
700 * Coldpath (SMP only since tokens always succeed on UP)
702 * We had some contention on the thread we wanted to schedule.
703 * What we do now is try to find a thread that we can schedule
706 * The coldpath scan does NOT rearrange threads in the run list.
707 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
708 * the next tick whenever the current head is not the current thread.
713 ++gd->gd_cnt.v_token_colls;
715 if (fairq_bypass > 0)
718 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
719 #ifndef NO_LWKT_SPLIT_USERPRI
721 * Never schedule threads returning to userland or the
722 * user thread scheduler helper thread when higher priority
723 * threads are present. The runq is sorted by priority
724 * so we can give up traversing it when we find the first
725 * low priority thread.
727 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
736 if (TD_TOKS_NOT_HELD(ntd) ||
737 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
743 ++gd->gd_cnt.v_token_colls;
748 * We exhausted the run list, meaning that all runnable threads
752 ntd = &gd->gd_idlethread;
753 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
754 ASSERT_NO_TOKENS_HELD(ntd);
755 /* contention case, do not clear contention mask */
758 * We are going to have to retry but if the current thread is not
759 * on the runq we instead switch through the idle thread to get away
760 * from the current thread. We have to flag for lwkt reschedule
761 * to prevent the idle thread from halting.
763 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
764 * instruct it to deal with the potential for deadlocks by
765 * ordering the tokens by address.
767 if ((td->td_flags & TDF_RUNQ) == 0) {
768 need_lwkt_resched(); /* prevent hlt */
771 #if defined(INVARIANTS) && defined(__x86_64__)
772 if ((read_rflags() & PSL_I) == 0) {
774 panic("lwkt_switch() called with interrupts disabled");
779 * Number iterations so far. After a certain point we switch to
780 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
782 if (spinning < 0x7FFFFFFF)
786 * lwkt_getalltokens() failed in sorted token mode, we can use
787 * monitor/mwait in this case.
789 if (spinning >= lwkt_spin_loops &&
790 (cpu_mi_feature & CPU_MI_MONITOR) &&
793 cpu_mmw_pause_int(&gd->gd_reqflags,
794 (gd->gd_reqflags | RQF_SPINNING) &
795 ~RQF_IDLECHECK_WK_MASK,
800 * We already checked that td is still scheduled so this should be
806 * This experimental resequencer is used as a fall-back to reduce
807 * hw cache line contention by placing each core's scheduler into a
808 * time-domain-multplexed slot.
810 * The resequencer is disabled by default. It's functionality has
811 * largely been superceeded by the token algorithm which limits races
812 * to a subset of cores.
814 * The resequencer algorithm tends to break down when more than
815 * 20 cores are contending. What appears to happen is that new
816 * tokens can be obtained out of address-sorted order by new cores
817 * while existing cores languish in long delays between retries and
818 * wind up being starved-out of the token acquisition.
820 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
821 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
824 while ((oseq = lwkt_cseq_rindex) != cseq) {
827 if (cpu_mi_feature & CPU_MI_MONITOR) {
828 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq, cpu_mwait_spin);
838 atomic_add_int(&lwkt_cseq_rindex, 1);
840 /* highest level for(;;) loop */
845 * Clear gd_idle_repeat when doing a normal switch to a non-idle
848 ntd->td_wmesg = NULL;
849 ++gd->gd_cnt.v_swtch;
850 gd->gd_idle_repeat = 0;
852 havethread_preempted:
854 * If the new target does not need the MP lock and we are holding it,
855 * release the MP lock. If the new target requires the MP lock we have
856 * already acquired it for the target.
860 KASSERT(ntd->td_critcount,
861 ("priority problem in lwkt_switch %d %d",
862 td->td_critcount, ntd->td_critcount));
866 * Execute the actual thread switch operation. This function
867 * returns to the current thread and returns the previous thread
868 * (which may be different from the thread we switched to).
870 * We are responsible for marking ntd as TDF_RUNNING.
872 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
874 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
875 ntd->td_flags |= TDF_RUNNING;
876 lwkt_switch_return(td->td_switch(ntd));
877 /* ntd invalid, td_switch() can return a different thread_t */
881 * catch-all. XXX is this strictly needed?
885 /* NOTE: current cpu may have changed after switch */
890 * Called by assembly in the td_switch (thread restore path) for thread
891 * bootstrap cases which do not 'return' to lwkt_switch().
894 lwkt_switch_return(thread_t otd)
899 * Check if otd was migrating. Now that we are on ntd we can finish
900 * up the migration. This is a bit messy but it is the only place
901 * where td is known to be fully descheduled.
903 * We can only activate the migration if otd was migrating but not
904 * held on the cpu due to a preemption chain. We still have to
905 * clear TDF_RUNNING on the old thread either way.
907 * We are responsible for clearing the previously running thread's
910 if ((rgd = otd->td_migrate_gd) != NULL &&
911 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
912 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
913 (TDF_MIGRATING | TDF_RUNNING));
914 otd->td_migrate_gd = NULL;
915 otd->td_flags &= ~TDF_RUNNING;
916 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
918 otd->td_flags &= ~TDF_RUNNING;
922 * Final exit validations (see lwp_wait()). Note that otd becomes
923 * invalid the *instant* we set TDF_MP_EXITSIG.
925 while (otd->td_flags & TDF_EXITING) {
928 mpflags = otd->td_mpflags;
931 if (mpflags & TDF_MP_EXITWAIT) {
932 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
933 mpflags | TDF_MP_EXITSIG)) {
938 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
939 mpflags | TDF_MP_EXITSIG)) {
948 * Request that the target thread preempt the current thread. Preemption
949 * can only occur if our only critical section is the one that we were called
950 * with, the relative priority of the target thread is higher, and the target
951 * thread holds no tokens. This also only works if we are not holding any
952 * spinlocks (obviously).
954 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
955 * this is called via lwkt_schedule() through the td_preemptable callback.
956 * critcount is the managed critical priority that we should ignore in order
957 * to determine whether preemption is possible (aka usually just the crit
958 * priority of lwkt_schedule() itself).
960 * Preemption is typically limited to interrupt threads.
962 * Operation works in a fairly straight-forward manner. The normal
963 * scheduling code is bypassed and we switch directly to the target
964 * thread. When the target thread attempts to block or switch away
965 * code at the base of lwkt_switch() will switch directly back to our
966 * thread. Our thread is able to retain whatever tokens it holds and
967 * if the target needs one of them the target will switch back to us
968 * and reschedule itself normally.
971 lwkt_preempt(thread_t ntd, int critcount)
973 struct globaldata *gd = mycpu;
976 int save_gd_intr_nesting_level;
979 * The caller has put us in a critical section. We can only preempt
980 * if the caller of the caller was not in a critical section (basically
981 * a local interrupt), as determined by the 'critcount' parameter. We
982 * also can't preempt if the caller is holding any spinlocks (even if
983 * he isn't in a critical section). This also handles the tokens test.
985 * YYY The target thread must be in a critical section (else it must
986 * inherit our critical section? I dunno yet).
988 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
990 td = gd->gd_curthread;
991 if (preempt_enable == 0) {
995 if (ntd->td_pri <= td->td_pri) {
999 if (td->td_critcount > critcount) {
1003 if (td->td_cscount) {
1007 if (ntd->td_gd != gd) {
1012 * We don't have to check spinlocks here as they will also bump
1015 * Do not try to preempt if the target thread is holding any tokens.
1016 * We could try to acquire the tokens but this case is so rare there
1017 * is no need to support it.
1019 KKASSERT(gd->gd_spinlocks == 0);
1021 if (TD_TOKS_HELD(ntd)) {
1025 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1029 if (ntd->td_preempted) {
1033 KKASSERT(gd->gd_processing_ipiq == 0);
1036 * Since we are able to preempt the current thread, there is no need to
1037 * call need_lwkt_resched().
1039 * We must temporarily clear gd_intr_nesting_level around the switch
1040 * since switchouts from the target thread are allowed (they will just
1041 * return to our thread), and since the target thread has its own stack.
1043 * A preemption must switch back to the original thread, assert the
1047 ntd->td_preempted = td;
1048 td->td_flags |= TDF_PREEMPT_LOCK;
1049 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1050 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1051 gd->gd_intr_nesting_level = 0;
1053 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1054 ntd->td_flags |= TDF_RUNNING;
1055 xtd = td->td_switch(ntd);
1056 KKASSERT(xtd == ntd);
1057 lwkt_switch_return(xtd);
1058 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1060 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1061 ntd->td_preempted = NULL;
1062 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1066 * Conditionally call splz() if gd_reqflags indicates work is pending.
1067 * This will work inside a critical section but not inside a hard code
1070 * (self contained on a per cpu basis)
1075 globaldata_t gd = mycpu;
1076 thread_t td = gd->gd_curthread;
1078 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1079 gd->gd_intr_nesting_level == 0 &&
1080 td->td_nest_count < 2)
1087 * This version is integrated into crit_exit, reqflags has already
1088 * been tested but td_critcount has not.
1090 * We only want to execute the splz() on the 1->0 transition of
1091 * critcount and not in a hard code section or if too deeply nested.
1093 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
1096 lwkt_maybe_splz(thread_t td)
1098 globaldata_t gd = td->td_gd;
1100 if (td->td_critcount == 0 &&
1101 gd->gd_intr_nesting_level == 0 &&
1102 td->td_nest_count < 2)
1109 * Drivers which set up processing co-threads can call this function to
1110 * run the co-thread at a higher priority and to allow it to preempt
1114 lwkt_set_interrupt_support_thread(void)
1116 thread_t td = curthread;
1118 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1119 td->td_flags |= TDF_INTTHREAD;
1120 td->td_preemptable = lwkt_preempt;
1125 * This function is used to negotiate a passive release of the current
1126 * process/lwp designation with the user scheduler, allowing the user
1127 * scheduler to schedule another user thread. The related kernel thread
1128 * (curthread) continues running in the released state.
1131 lwkt_passive_release(struct thread *td)
1133 struct lwp *lp = td->td_lwp;
1135 #ifndef NO_LWKT_SPLIT_USERPRI
1136 td->td_release = NULL;
1137 lwkt_setpri_self(TDPRI_KERN_USER);
1140 lp->lwp_proc->p_usched->release_curproc(lp);
1145 * This implements a LWKT yield, allowing a kernel thread to yield to other
1146 * kernel threads at the same or higher priority. This function can be
1147 * called in a tight loop and will typically only yield once per tick.
1149 * Most kernel threads run at the same priority in order to allow equal
1152 * (self contained on a per cpu basis)
1157 globaldata_t gd = mycpu;
1158 thread_t td = gd->gd_curthread;
1160 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1162 if (lwkt_resched_wanted()) {
1163 lwkt_schedule_self(curthread);
1169 * The quick version processes pending interrupts and higher-priority
1170 * LWKT threads but will not round-robin same-priority LWKT threads.
1172 * When called while attempting to return to userland the only same-pri
1173 * threads are the ones which have already tried to become the current
1177 lwkt_yield_quick(void)
1179 globaldata_t gd = mycpu;
1180 thread_t td = gd->gd_curthread;
1182 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1184 if (lwkt_resched_wanted()) {
1186 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1187 clear_lwkt_resched();
1189 lwkt_schedule_self(curthread);
1197 * This yield is designed for kernel threads with a user context.
1199 * The kernel acting on behalf of the user is potentially cpu-bound,
1200 * this function will efficiently allow other threads to run and also
1201 * switch to other processes by releasing.
1203 * The lwkt_user_yield() function is designed to have very low overhead
1204 * if no yield is determined to be needed.
1207 lwkt_user_yield(void)
1209 globaldata_t gd = mycpu;
1210 thread_t td = gd->gd_curthread;
1213 * Always run any pending interrupts in case we are in a critical
1216 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1220 * Switch (which forces a release) if another kernel thread needs
1221 * the cpu, if userland wants us to resched, or if our kernel
1222 * quantum has run out.
1224 if (lwkt_resched_wanted() ||
1225 user_resched_wanted())
1232 * Reacquire the current process if we are released.
1234 * XXX not implemented atm. The kernel may be holding locks and such,
1235 * so we want the thread to continue to receive cpu.
1237 if (td->td_release == NULL && lp) {
1238 lp->lwp_proc->p_usched->acquire_curproc(lp);
1239 td->td_release = lwkt_passive_release;
1240 lwkt_setpri_self(TDPRI_USER_NORM);
1246 * Generic schedule. Possibly schedule threads belonging to other cpus and
1247 * deal with threads that might be blocked on a wait queue.
1249 * We have a little helper inline function which does additional work after
1250 * the thread has been enqueued, including dealing with preemption and
1251 * setting need_lwkt_resched() (which prevents the kernel from returning
1252 * to userland until it has processed higher priority threads).
1254 * It is possible for this routine to be called after a failed _enqueue
1255 * (due to the target thread migrating, sleeping, or otherwise blocked).
1256 * We have to check that the thread is actually on the run queue!
1260 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1262 if (ntd->td_flags & TDF_RUNQ) {
1263 if (ntd->td_preemptable) {
1264 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1271 _lwkt_schedule(thread_t td)
1273 globaldata_t mygd = mycpu;
1275 KASSERT(td != &td->td_gd->gd_idlethread,
1276 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1277 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1278 crit_enter_gd(mygd);
1279 KKASSERT(td->td_lwp == NULL ||
1280 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1282 if (td == mygd->gd_curthread) {
1286 * If we own the thread, there is no race (since we are in a
1287 * critical section). If we do not own the thread there might
1288 * be a race but the target cpu will deal with it.
1290 if (td->td_gd == mygd) {
1292 _lwkt_schedule_post(mygd, td, 1);
1294 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1301 lwkt_schedule(thread_t td)
1307 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1313 * When scheduled remotely if frame != NULL the IPIQ is being
1314 * run via doreti or an interrupt then preemption can be allowed.
1316 * To allow preemption we have to drop the critical section so only
1317 * one is present in _lwkt_schedule_post.
1320 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1322 thread_t td = curthread;
1325 if (frame && ntd->td_preemptable) {
1326 crit_exit_noyield(td);
1327 _lwkt_schedule(ntd);
1328 crit_enter_quick(td);
1330 _lwkt_schedule(ntd);
1335 * Thread migration using a 'Pull' method. The thread may or may not be
1336 * the current thread. It MUST be descheduled and in a stable state.
1337 * lwkt_giveaway() must be called on the cpu owning the thread.
1339 * At any point after lwkt_giveaway() is called, the target cpu may
1340 * 'pull' the thread by calling lwkt_acquire().
1342 * We have to make sure the thread is not sitting on a per-cpu tsleep
1343 * queue or it will blow up when it moves to another cpu.
1345 * MPSAFE - must be called under very specific conditions.
1348 lwkt_giveaway(thread_t td)
1350 globaldata_t gd = mycpu;
1353 if (td->td_flags & TDF_TSLEEPQ)
1355 KKASSERT(td->td_gd == gd);
1356 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1357 td->td_flags |= TDF_MIGRATING;
1362 lwkt_acquire(thread_t td)
1366 int retry = 10000000;
1368 KKASSERT(td->td_flags & TDF_MIGRATING);
1373 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1374 crit_enter_gd(mygd);
1375 DEBUG_PUSH_INFO("lwkt_acquire");
1376 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1377 lwkt_process_ipiq();
1380 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1388 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1389 td->td_flags &= ~TDF_MIGRATING;
1392 crit_enter_gd(mygd);
1393 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1394 td->td_flags &= ~TDF_MIGRATING;
1400 * Generic deschedule. Descheduling threads other then your own should be
1401 * done only in carefully controlled circumstances. Descheduling is
1404 * This function may block if the cpu has run out of messages.
1407 lwkt_deschedule(thread_t td)
1410 if (td == curthread) {
1413 if (td->td_gd == mycpu) {
1416 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1423 * Set the target thread's priority. This routine does not automatically
1424 * switch to a higher priority thread, LWKT threads are not designed for
1425 * continuous priority changes. Yield if you want to switch.
1428 lwkt_setpri(thread_t td, int pri)
1430 if (td->td_pri != pri) {
1433 if (td->td_flags & TDF_RUNQ) {
1434 KKASSERT(td->td_gd == mycpu);
1446 * Set the initial priority for a thread prior to it being scheduled for
1447 * the first time. The thread MUST NOT be scheduled before or during
1448 * this call. The thread may be assigned to a cpu other then the current
1451 * Typically used after a thread has been created with TDF_STOPPREQ,
1452 * and before the thread is initially scheduled.
1455 lwkt_setpri_initial(thread_t td, int pri)
1458 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1463 lwkt_setpri_self(int pri)
1465 thread_t td = curthread;
1467 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1469 if (td->td_flags & TDF_RUNQ) {
1480 * hz tick scheduler clock for LWKT threads
1483 lwkt_schedulerclock(thread_t td)
1485 globaldata_t gd = td->td_gd;
1488 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1490 * If the current thread is at the head of the runq shift it to the
1491 * end of any equal-priority threads and request a LWKT reschedule
1494 * Ignore upri in this situation. There will only be one user thread
1495 * in user mode, all others will be user threads running in kernel
1496 * mode and we have to make sure they get some cpu.
1498 xtd = TAILQ_NEXT(td, td_threadq);
1499 if (xtd && xtd->td_pri == td->td_pri) {
1500 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1501 while (xtd && xtd->td_pri == td->td_pri)
1502 xtd = TAILQ_NEXT(xtd, td_threadq);
1504 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1506 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1507 need_lwkt_resched();
1511 * If we scheduled a thread other than the one at the head of the
1512 * queue always request a reschedule every tick.
1514 need_lwkt_resched();
1519 * Migrate the current thread to the specified cpu.
1521 * This is accomplished by descheduling ourselves from the current cpu
1522 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1523 * 'old' thread wants to migrate after it has been completely switched out
1524 * and will complete the migration.
1526 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1528 * We must be sure to release our current process designation (if a user
1529 * process) before clearing out any tsleepq we are on because the release
1530 * code may re-add us.
1532 * We must be sure to remove ourselves from the current cpu's tsleepq
1533 * before potentially moving to another queue. The thread can be on
1534 * a tsleepq due to a left-over tsleep_interlock().
1538 lwkt_setcpu_self(globaldata_t rgd)
1540 thread_t td = curthread;
1542 if (td->td_gd != rgd) {
1543 crit_enter_quick(td);
1547 if (td->td_flags & TDF_TSLEEPQ)
1551 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1552 * trying to deschedule ourselves and switch away, then deschedule
1553 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1554 * call lwkt_switch() to complete the operation.
1556 td->td_flags |= TDF_MIGRATING;
1557 lwkt_deschedule_self(td);
1558 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1559 td->td_migrate_gd = rgd;
1563 * We are now on the target cpu
1565 KKASSERT(rgd == mycpu);
1566 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1567 crit_exit_quick(td);
1572 lwkt_migratecpu(int cpuid)
1576 rgd = globaldata_find(cpuid);
1577 lwkt_setcpu_self(rgd);
1581 * Remote IPI for cpu migration (called while in a critical section so we
1582 * do not have to enter another one).
1584 * The thread (td) has already been completely descheduled from the
1585 * originating cpu and we can simply assert the case. The thread is
1586 * assigned to the new cpu and enqueued.
1588 * The thread will re-add itself to tdallq when it resumes execution.
1591 lwkt_setcpu_remote(void *arg)
1594 globaldata_t gd = mycpu;
1596 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1599 td->td_flags &= ~TDF_MIGRATING;
1600 KKASSERT(td->td_migrate_gd == NULL);
1601 KKASSERT(td->td_lwp == NULL ||
1602 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1607 lwkt_preempted_proc(void)
1609 thread_t td = curthread;
1610 while (td->td_preempted)
1611 td = td->td_preempted;
1616 * Create a kernel process/thread/whatever. It shares it's address space
1617 * with proc0 - ie: kernel only.
1619 * If the cpu is not specified one will be selected. In the future
1620 * specifying a cpu of -1 will enable kernel thread migration between
1624 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1625 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1630 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1634 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1637 * Set up arg0 for 'ps' etc
1639 __va_start(ap, fmt);
1640 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1644 * Schedule the thread to run
1646 if (td->td_flags & TDF_NOSTART)
1647 td->td_flags &= ~TDF_NOSTART;
1654 * Destroy an LWKT thread. Warning! This function is not called when
1655 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1656 * uses a different reaping mechanism.
1661 thread_t td = curthread;
1666 * Do any cleanup that might block here
1668 if (td->td_flags & TDF_VERBOSE)
1669 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1671 dsched_exit_thread(td);
1674 * Get us into a critical section to interlock gd_freetd and loop
1675 * until we can get it freed.
1677 * We have to cache the current td in gd_freetd because objcache_put()ing
1678 * it would rip it out from under us while our thread is still active.
1680 * We are the current thread so of course our own TDF_RUNNING bit will
1681 * be set, so unlike the lwp reap code we don't wait for it to clear.
1684 crit_enter_quick(td);
1687 tsleep(td, 0, "tdreap", 1);
1690 if ((std = gd->gd_freetd) != NULL) {
1691 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1692 gd->gd_freetd = NULL;
1693 objcache_put(thread_cache, std);
1700 * Remove thread resources from kernel lists and deschedule us for
1701 * the last time. We cannot block after this point or we may end
1702 * up with a stale td on the tsleepq.
1704 * None of this may block, the critical section is the only thing
1705 * protecting tdallq and the only thing preventing new lwkt_hold()
1708 if (td->td_flags & TDF_TSLEEPQ)
1710 lwkt_deschedule_self(td);
1711 lwkt_remove_tdallq(td);
1712 KKASSERT(td->td_refs == 0);
1717 KKASSERT(gd->gd_freetd == NULL);
1718 if (td->td_flags & TDF_ALLOCATED_THREAD)
1724 lwkt_remove_tdallq(thread_t td)
1726 KKASSERT(td->td_gd == mycpu);
1727 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1731 * Code reduction and branch prediction improvements. Call/return
1732 * overhead on modern cpus often degenerates into 0 cycles due to
1733 * the cpu's branch prediction hardware and return pc cache. We
1734 * can take advantage of this by not inlining medium-complexity
1735 * functions and we can also reduce the branch prediction impact
1736 * by collapsing perfectly predictable branches into a single
1737 * procedure instead of duplicating it.
1739 * Is any of this noticeable? Probably not, so I'll take the
1740 * smaller code size.
1743 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1745 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1751 thread_t td = curthread;
1752 int lcrit = td->td_critcount;
1754 td->td_critcount = 0;
1755 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1760 * Called from debugger/panic on cpus which have been stopped. We must still
1761 * process the IPIQ while stopped, even if we were stopped while in a critical
1764 * If we are dumping also try to process any pending interrupts. This may
1765 * or may not work depending on the state of the cpu at the point it was
1769 lwkt_smp_stopped(void)
1771 globaldata_t gd = mycpu;
1775 lwkt_process_ipiq();
1778 lwkt_process_ipiq();