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;
95 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
96 static void lwkt_setcpu_remote(void *arg);
98 extern void cpu_heavy_restore(void);
99 extern void cpu_lwkt_restore(void);
100 extern void cpu_kthread_restore(void);
101 extern void cpu_idle_restore(void);
104 * We can make all thread ports use the spin backend instead of the thread
105 * backend. This should only be set to debug the spin backend.
107 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
110 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
111 "Panic if attempting to switch lwkt's while mastering cpusync");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
114 "Number of switched threads");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
116 "Successful preemption events");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
118 "Failed preemption events");
119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
120 "Number of preempted threads.");
121 static int fairq_enable = 0;
122 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
123 &fairq_enable, 0, "Turn on fairq priority accumulators");
124 static int fairq_bypass = -1;
125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
127 extern int lwkt_sched_debug;
128 int lwkt_sched_debug = 0;
129 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
130 &lwkt_sched_debug, 0, "Scheduler debug");
131 static int lwkt_spin_loops = 10;
132 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
134 static int lwkt_spin_reseq = 0;
135 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
136 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
137 static int lwkt_spin_monitor = 0;
138 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
139 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
140 static int lwkt_spin_fatal = 0; /* disabled */
141 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
142 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
143 static int preempt_enable = 1;
144 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
145 &preempt_enable, 0, "Enable preemption");
146 static int lwkt_cache_threads = 0;
147 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
148 &lwkt_cache_threads, 0, "thread+kstack cache");
150 static __cachealign int lwkt_cseq_rindex;
151 static __cachealign int lwkt_cseq_windex;
154 * These helper procedures handle the runq, they can only be called from
155 * within a critical section.
157 * WARNING! Prior to SMP being brought up it is possible to enqueue and
158 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
159 * instead of 'mycpu' when referencing the globaldata structure. Once
160 * SMP live enqueuing and dequeueing only occurs on the current cpu.
164 _lwkt_dequeue(thread_t td)
166 if (td->td_flags & TDF_RUNQ) {
167 struct globaldata *gd = td->td_gd;
169 td->td_flags &= ~TDF_RUNQ;
170 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
171 --gd->gd_tdrunqcount;
172 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
173 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
180 * There are a limited number of lwkt threads runnable since user
181 * processes only schedule one at a time per cpu. However, there can
182 * be many user processes in kernel mode exiting from a tsleep() which
185 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
186 * will ignore user priority. This is to ensure that user threads in
187 * kernel mode get cpu at some point regardless of what the user
192 _lwkt_enqueue(thread_t td)
196 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
197 struct globaldata *gd = td->td_gd;
199 td->td_flags |= TDF_RUNQ;
200 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
202 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
203 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
206 * NOTE: td_upri - higher numbers more desireable, same sense
207 * as td_pri (typically reversed from lwp_upri).
209 * In the equal priority case we want the best selection
210 * at the beginning so the less desireable selections know
211 * that they have to setrunqueue/go-to-another-cpu, even
212 * though it means switching back to the 'best' selection.
213 * This also avoids degenerate situations when many threads
214 * are runnable or waking up at the same time.
216 * If upri matches exactly place at end/round-robin.
219 (xtd->td_pri >= td->td_pri ||
220 (xtd->td_pri == td->td_pri &&
221 xtd->td_upri >= td->td_upri))) {
222 xtd = TAILQ_NEXT(xtd, td_threadq);
225 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
227 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
229 ++gd->gd_tdrunqcount;
232 * Request a LWKT reschedule if we are now at the head of the queue.
234 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
240 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
242 struct thread *td = (struct thread *)obj;
244 td->td_kstack = NULL;
245 td->td_kstack_size = 0;
246 td->td_flags = TDF_ALLOCATED_THREAD;
252 _lwkt_thread_dtor(void *obj, void *privdata)
254 struct thread *td = (struct thread *)obj;
256 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
257 ("_lwkt_thread_dtor: not allocated from objcache"));
258 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
259 td->td_kstack_size > 0,
260 ("_lwkt_thread_dtor: corrupted stack"));
261 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
262 td->td_kstack = NULL;
267 * Initialize the lwkt s/system.
269 * Nominally cache up to 32 thread + kstack structures. Cache more on
270 * systems with a lot of cpu cores.
275 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
276 if (lwkt_cache_threads == 0) {
277 lwkt_cache_threads = ncpus * 4;
278 if (lwkt_cache_threads < 32)
279 lwkt_cache_threads = 32;
281 thread_cache = objcache_create_mbacked(
282 M_THREAD, sizeof(struct thread),
283 0, lwkt_cache_threads,
284 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
288 * Schedule a thread to run. As the current thread we can always safely
289 * schedule ourselves, and a shortcut procedure is provided for that
292 * (non-blocking, self contained on a per cpu basis)
295 lwkt_schedule_self(thread_t td)
297 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
298 crit_enter_quick(td);
299 KASSERT(td != &td->td_gd->gd_idlethread,
300 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
301 KKASSERT(td->td_lwp == NULL ||
302 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
308 * Deschedule a thread.
310 * (non-blocking, self contained on a per cpu basis)
313 lwkt_deschedule_self(thread_t td)
315 crit_enter_quick(td);
321 * LWKTs operate on a per-cpu basis
323 * WARNING! Called from early boot, 'mycpu' may not work yet.
326 lwkt_gdinit(struct globaldata *gd)
328 TAILQ_INIT(&gd->gd_tdrunq);
329 TAILQ_INIT(&gd->gd_tdallq);
333 * Create a new thread. The thread must be associated with a process context
334 * or LWKT start address before it can be scheduled. If the target cpu is
335 * -1 the thread will be created on the current cpu.
337 * If you intend to create a thread without a process context this function
338 * does everything except load the startup and switcher function.
341 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
343 static int cpu_rotator;
344 globaldata_t gd = mycpu;
348 * If static thread storage is not supplied allocate a thread. Reuse
349 * a cached free thread if possible. gd_freetd is used to keep an exiting
350 * thread intact through the exit.
354 if ((td = gd->gd_freetd) != NULL) {
355 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
357 gd->gd_freetd = NULL;
359 td = objcache_get(thread_cache, M_WAITOK);
360 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
364 KASSERT((td->td_flags &
365 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
366 TDF_ALLOCATED_THREAD,
367 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
368 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
372 * Try to reuse cached stack.
374 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
375 if (flags & TDF_ALLOCATED_STACK) {
376 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
381 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
382 flags |= TDF_ALLOCATED_STACK;
389 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
394 * Initialize a preexisting thread structure. This function is used by
395 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
397 * All threads start out in a critical section at a priority of
398 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
399 * appropriate. This function may send an IPI message when the
400 * requested cpu is not the current cpu and consequently gd_tdallq may
401 * not be initialized synchronously from the point of view of the originating
404 * NOTE! we have to be careful in regards to creating threads for other cpus
405 * if SMP has not yet been activated.
408 lwkt_init_thread_remote(void *arg)
413 * Protected by critical section held by IPI dispatch
415 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
419 * lwkt core thread structural initialization.
421 * NOTE: All threads are initialized as mpsafe threads.
424 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
425 struct globaldata *gd)
427 globaldata_t mygd = mycpu;
429 bzero(td, sizeof(struct thread));
430 td->td_kstack = stack;
431 td->td_kstack_size = stksize;
432 td->td_flags = flags;
434 td->td_type = TD_TYPE_GENERIC;
436 td->td_pri = TDPRI_KERN_DAEMON;
437 td->td_critcount = 1;
438 td->td_toks_have = NULL;
439 td->td_toks_stop = &td->td_toks_base;
440 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
441 lwkt_initport_spin(&td->td_msgport, td);
443 lwkt_initport_thread(&td->td_msgport, td);
444 pmap_init_thread(td);
446 * Normally initializing a thread for a remote cpu requires sending an
447 * IPI. However, the idlethread is setup before the other cpus are
448 * activated so we have to treat it as a special case. XXX manipulation
449 * of gd_tdallq requires the BGL.
451 if (gd == mygd || td == &gd->gd_idlethread) {
453 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
456 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
458 dsched_new_thread(td);
462 lwkt_set_comm(thread_t td, const char *ctl, ...)
467 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
469 KTR_LOG(ctxsw_newtd, td, td->td_comm);
473 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
474 * this does not prevent the thread from migrating to another cpu so the
475 * gd_tdallq state is not protected by this.
478 lwkt_hold(thread_t td)
480 atomic_add_int(&td->td_refs, 1);
484 lwkt_rele(thread_t td)
486 KKASSERT(td->td_refs > 0);
487 atomic_add_int(&td->td_refs, -1);
491 lwkt_free_thread(thread_t td)
493 KKASSERT(td->td_refs == 0);
494 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
495 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
496 if (td->td_flags & TDF_ALLOCATED_THREAD) {
497 objcache_put(thread_cache, td);
498 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
499 /* client-allocated struct with internally allocated stack */
500 KASSERT(td->td_kstack && td->td_kstack_size > 0,
501 ("lwkt_free_thread: corrupted stack"));
502 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
503 td->td_kstack = NULL;
504 td->td_kstack_size = 0;
506 KTR_LOG(ctxsw_deadtd, td);
511 * Switch to the next runnable lwkt. If no LWKTs are runnable then
512 * switch to the idlethread. Switching must occur within a critical
513 * section to avoid races with the scheduling queue.
515 * We always have full control over our cpu's run queue. Other cpus
516 * that wish to manipulate our queue must use the cpu_*msg() calls to
517 * talk to our cpu, so a critical section is all that is needed and
518 * the result is very, very fast thread switching.
520 * The LWKT scheduler uses a fixed priority model and round-robins at
521 * each priority level. User process scheduling is a totally
522 * different beast and LWKT priorities should not be confused with
523 * user process priorities.
525 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
526 * is not called by the current thread in the preemption case, only when
527 * the preempting thread blocks (in order to return to the original thread).
529 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
530 * migration and tsleep deschedule the current lwkt thread and call
531 * lwkt_switch(). In particular, the target cpu of the migration fully
532 * expects the thread to become non-runnable and can deadlock against
533 * cpusync operations if we run any IPIs prior to switching the thread out.
535 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
536 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
541 globaldata_t gd = mycpu;
542 thread_t td = gd->gd_curthread;
546 KKASSERT(gd->gd_processing_ipiq == 0);
547 KKASSERT(td->td_flags & TDF_RUNNING);
550 * Switching from within a 'fast' (non thread switched) interrupt or IPI
551 * is illegal. However, we may have to do it anyway if we hit a fatal
552 * kernel trap or we have paniced.
554 * If this case occurs save and restore the interrupt nesting level.
556 if (gd->gd_intr_nesting_level) {
560 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
561 panic("lwkt_switch: Attempt to switch from a "
562 "fast interrupt, ipi, or hard code section, "
566 savegdnest = gd->gd_intr_nesting_level;
567 savegdtrap = gd->gd_trap_nesting_level;
568 gd->gd_intr_nesting_level = 0;
569 gd->gd_trap_nesting_level = 0;
570 if ((td->td_flags & TDF_PANICWARN) == 0) {
571 td->td_flags |= TDF_PANICWARN;
572 kprintf("Warning: thread switch from interrupt, IPI, "
573 "or hard code section.\n"
574 "thread %p (%s)\n", td, td->td_comm);
578 gd->gd_intr_nesting_level = savegdnest;
579 gd->gd_trap_nesting_level = savegdtrap;
585 * Release our current user process designation if we are blocking
586 * or if a user reschedule was requested.
588 * NOTE: This function is NOT called if we are switching into or
589 * returning from a preemption.
591 * NOTE: Releasing our current user process designation may cause
592 * it to be assigned to another thread, which in turn will
593 * cause us to block in the usched acquire code when we attempt
594 * to return to userland.
596 * NOTE: On SMP systems this can be very nasty when heavy token
597 * contention is present so we want to be careful not to
598 * release the designation gratuitously.
600 if (td->td_release &&
601 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
609 if (TD_TOKS_HELD(td))
610 lwkt_relalltokens(td);
613 * We had better not be holding any spin locks, but don't get into an
614 * endless panic loop.
616 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
617 ("lwkt_switch: still holding %d exclusive spinlocks!",
622 if (td->td_cscount) {
623 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
625 if (panic_on_cscount)
626 panic("switching while mastering cpusync");
631 * If we had preempted another thread on this cpu, resume the preempted
632 * thread. This occurs transparently, whether the preempted thread
633 * was scheduled or not (it may have been preempted after descheduling
636 * We have to setup the MP lock for the original thread after backing
637 * out the adjustment that was made to curthread when the original
640 if ((ntd = td->td_preempted) != NULL) {
641 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
642 ntd->td_flags |= TDF_PREEMPT_DONE;
645 * The interrupt may have woken a thread up, we need to properly
646 * set the reschedule flag if the originally interrupted thread is
647 * at a lower priority.
649 * The interrupt may not have descheduled.
651 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
653 goto havethread_preempted;
657 * If we cannot obtain ownership of the tokens we cannot immediately
658 * schedule the target thread.
660 * Reminder: Again, we cannot afford to run any IPIs in this path if
661 * the current thread has been descheduled.
664 clear_lwkt_resched();
667 * Hotpath - pull the head of the run queue and attempt to schedule
670 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
674 * Runq is empty, switch to idle to allow it to halt.
676 ntd = &gd->gd_idlethread;
677 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
678 ASSERT_NO_TOKENS_HELD(ntd);
679 cpu_time.cp_msg[0] = 0;
680 cpu_time.cp_stallpc = 0;
685 * Hotpath - schedule ntd.
687 * NOTE: For UP there is no mplock and lwkt_getalltokens()
690 if (TD_TOKS_NOT_HELD(ntd) ||
691 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
697 * Coldpath (SMP only since tokens always succeed on UP)
699 * We had some contention on the thread we wanted to schedule.
700 * What we do now is try to find a thread that we can schedule
703 * The coldpath scan does NOT rearrange threads in the run list.
704 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
705 * the next tick whenever the current head is not the current thread.
710 ++gd->gd_cnt.v_token_colls;
712 if (fairq_bypass > 0)
715 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
716 #ifndef NO_LWKT_SPLIT_USERPRI
718 * Never schedule threads returning to userland or the
719 * user thread scheduler helper thread when higher priority
720 * threads are present. The runq is sorted by priority
721 * so we can give up traversing it when we find the first
722 * low priority thread.
724 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
733 if (TD_TOKS_NOT_HELD(ntd) ||
734 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
740 ++gd->gd_cnt.v_token_colls;
745 * We exhausted the run list, meaning that all runnable threads
749 ntd = &gd->gd_idlethread;
750 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
751 ASSERT_NO_TOKENS_HELD(ntd);
752 /* contention case, do not clear contention mask */
755 * We are going to have to retry but if the current thread is not
756 * on the runq we instead switch through the idle thread to get away
757 * from the current thread. We have to flag for lwkt reschedule
758 * to prevent the idle thread from halting.
760 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
761 * instruct it to deal with the potential for deadlocks by
762 * ordering the tokens by address.
764 if ((td->td_flags & TDF_RUNQ) == 0) {
765 need_lwkt_resched(); /* prevent hlt */
768 #if defined(INVARIANTS) && defined(__x86_64__)
769 if ((read_rflags() & PSL_I) == 0) {
771 panic("lwkt_switch() called with interrupts disabled");
776 * Number iterations so far. After a certain point we switch to
777 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
779 if (spinning < 0x7FFFFFFF)
783 * lwkt_getalltokens() failed in sorted token mode, we can use
784 * monitor/mwait in this case.
786 if (spinning >= lwkt_spin_loops &&
787 (cpu_mi_feature & CPU_MI_MONITOR) &&
790 cpu_mmw_pause_int(&gd->gd_reqflags,
791 (gd->gd_reqflags | RQF_SPINNING) &
792 ~RQF_IDLECHECK_WK_MASK);
796 * We already checked that td is still scheduled so this should be
802 * This experimental resequencer is used as a fall-back to reduce
803 * hw cache line contention by placing each core's scheduler into a
804 * time-domain-multplexed slot.
806 * The resequencer is disabled by default. It's functionality has
807 * largely been superceeded by the token algorithm which limits races
808 * to a subset of cores.
810 * The resequencer algorithm tends to break down when more than
811 * 20 cores are contending. What appears to happen is that new
812 * tokens can be obtained out of address-sorted order by new cores
813 * while existing cores languish in long delays between retries and
814 * wind up being starved-out of the token acquisition.
816 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
817 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
820 while ((oseq = lwkt_cseq_rindex) != cseq) {
823 if (cpu_mi_feature & CPU_MI_MONITOR) {
824 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
834 atomic_add_int(&lwkt_cseq_rindex, 1);
836 /* highest level for(;;) loop */
841 * Clear gd_idle_repeat when doing a normal switch to a non-idle
844 ntd->td_wmesg = NULL;
845 ++gd->gd_cnt.v_swtch;
846 gd->gd_idle_repeat = 0;
848 havethread_preempted:
850 * If the new target does not need the MP lock and we are holding it,
851 * release the MP lock. If the new target requires the MP lock we have
852 * already acquired it for the target.
856 KASSERT(ntd->td_critcount,
857 ("priority problem in lwkt_switch %d %d",
858 td->td_critcount, ntd->td_critcount));
862 * Execute the actual thread switch operation. This function
863 * returns to the current thread and returns the previous thread
864 * (which may be different from the thread we switched to).
866 * We are responsible for marking ntd as TDF_RUNNING.
868 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
870 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
871 ntd->td_flags |= TDF_RUNNING;
872 lwkt_switch_return(td->td_switch(ntd));
873 /* ntd invalid, td_switch() can return a different thread_t */
877 * catch-all. XXX is this strictly needed?
881 /* NOTE: current cpu may have changed after switch */
886 * Called by assembly in the td_switch (thread restore path) for thread
887 * bootstrap cases which do not 'return' to lwkt_switch().
890 lwkt_switch_return(thread_t otd)
895 * Check if otd was migrating. Now that we are on ntd we can finish
896 * up the migration. This is a bit messy but it is the only place
897 * where td is known to be fully descheduled.
899 * We can only activate the migration if otd was migrating but not
900 * held on the cpu due to a preemption chain. We still have to
901 * clear TDF_RUNNING on the old thread either way.
903 * We are responsible for clearing the previously running thread's
906 if ((rgd = otd->td_migrate_gd) != NULL &&
907 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
908 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
909 (TDF_MIGRATING | TDF_RUNNING));
910 otd->td_migrate_gd = NULL;
911 otd->td_flags &= ~TDF_RUNNING;
912 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
914 otd->td_flags &= ~TDF_RUNNING;
918 * Final exit validations (see lwp_wait()). Note that otd becomes
919 * invalid the *instant* we set TDF_MP_EXITSIG.
921 while (otd->td_flags & TDF_EXITING) {
924 mpflags = otd->td_mpflags;
927 if (mpflags & TDF_MP_EXITWAIT) {
928 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
929 mpflags | TDF_MP_EXITSIG)) {
934 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
935 mpflags | TDF_MP_EXITSIG)) {
944 * Request that the target thread preempt the current thread. Preemption
945 * can only occur if our only critical section is the one that we were called
946 * with, the relative priority of the target thread is higher, and the target
947 * thread holds no tokens. This also only works if we are not holding any
948 * spinlocks (obviously).
950 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
951 * this is called via lwkt_schedule() through the td_preemptable callback.
952 * critcount is the managed critical priority that we should ignore in order
953 * to determine whether preemption is possible (aka usually just the crit
954 * priority of lwkt_schedule() itself).
956 * Preemption is typically limited to interrupt threads.
958 * Operation works in a fairly straight-forward manner. The normal
959 * scheduling code is bypassed and we switch directly to the target
960 * thread. When the target thread attempts to block or switch away
961 * code at the base of lwkt_switch() will switch directly back to our
962 * thread. Our thread is able to retain whatever tokens it holds and
963 * if the target needs one of them the target will switch back to us
964 * and reschedule itself normally.
967 lwkt_preempt(thread_t ntd, int critcount)
969 struct globaldata *gd = mycpu;
972 int save_gd_intr_nesting_level;
975 * The caller has put us in a critical section. We can only preempt
976 * if the caller of the caller was not in a critical section (basically
977 * a local interrupt), as determined by the 'critcount' parameter. We
978 * also can't preempt if the caller is holding any spinlocks (even if
979 * he isn't in a critical section). This also handles the tokens test.
981 * YYY The target thread must be in a critical section (else it must
982 * inherit our critical section? I dunno yet).
984 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
986 td = gd->gd_curthread;
987 if (preempt_enable == 0) {
991 if (ntd->td_pri <= td->td_pri) {
995 if (td->td_critcount > critcount) {
999 if (td->td_cscount) {
1003 if (ntd->td_gd != gd) {
1008 * We don't have to check spinlocks here as they will also bump
1011 * Do not try to preempt if the target thread is holding any tokens.
1012 * We could try to acquire the tokens but this case is so rare there
1013 * is no need to support it.
1015 KKASSERT(gd->gd_spinlocks == 0);
1017 if (TD_TOKS_HELD(ntd)) {
1021 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1025 if (ntd->td_preempted) {
1029 KKASSERT(gd->gd_processing_ipiq == 0);
1032 * Since we are able to preempt the current thread, there is no need to
1033 * call need_lwkt_resched().
1035 * We must temporarily clear gd_intr_nesting_level around the switch
1036 * since switchouts from the target thread are allowed (they will just
1037 * return to our thread), and since the target thread has its own stack.
1039 * A preemption must switch back to the original thread, assert the
1043 ntd->td_preempted = td;
1044 td->td_flags |= TDF_PREEMPT_LOCK;
1045 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1046 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1047 gd->gd_intr_nesting_level = 0;
1049 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1050 ntd->td_flags |= TDF_RUNNING;
1051 xtd = td->td_switch(ntd);
1052 KKASSERT(xtd == ntd);
1053 lwkt_switch_return(xtd);
1054 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1056 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1057 ntd->td_preempted = NULL;
1058 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1062 * Conditionally call splz() if gd_reqflags indicates work is pending.
1063 * This will work inside a critical section but not inside a hard code
1066 * (self contained on a per cpu basis)
1071 globaldata_t gd = mycpu;
1072 thread_t td = gd->gd_curthread;
1074 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1075 gd->gd_intr_nesting_level == 0 &&
1076 td->td_nest_count < 2)
1083 * This version is integrated into crit_exit, reqflags has already
1084 * been tested but td_critcount has not.
1086 * We only want to execute the splz() on the 1->0 transition of
1087 * critcount and not in a hard code section or if too deeply nested.
1089 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
1092 lwkt_maybe_splz(thread_t td)
1094 globaldata_t gd = td->td_gd;
1096 if (td->td_critcount == 0 &&
1097 gd->gd_intr_nesting_level == 0 &&
1098 td->td_nest_count < 2)
1105 * Drivers which set up processing co-threads can call this function to
1106 * run the co-thread at a higher priority and to allow it to preempt
1110 lwkt_set_interrupt_support_thread(void)
1112 thread_t td = curthread;
1114 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1115 td->td_flags |= TDF_INTTHREAD;
1116 td->td_preemptable = lwkt_preempt;
1121 * This function is used to negotiate a passive release of the current
1122 * process/lwp designation with the user scheduler, allowing the user
1123 * scheduler to schedule another user thread. The related kernel thread
1124 * (curthread) continues running in the released state.
1127 lwkt_passive_release(struct thread *td)
1129 struct lwp *lp = td->td_lwp;
1131 #ifndef NO_LWKT_SPLIT_USERPRI
1132 td->td_release = NULL;
1133 lwkt_setpri_self(TDPRI_KERN_USER);
1136 lp->lwp_proc->p_usched->release_curproc(lp);
1141 * This implements a LWKT yield, allowing a kernel thread to yield to other
1142 * kernel threads at the same or higher priority. This function can be
1143 * called in a tight loop and will typically only yield once per tick.
1145 * Most kernel threads run at the same priority in order to allow equal
1148 * (self contained on a per cpu basis)
1153 globaldata_t gd = mycpu;
1154 thread_t td = gd->gd_curthread;
1156 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1158 if (lwkt_resched_wanted()) {
1159 lwkt_schedule_self(curthread);
1165 * The quick version processes pending interrupts and higher-priority
1166 * LWKT threads but will not round-robin same-priority LWKT threads.
1168 * When called while attempting to return to userland the only same-pri
1169 * threads are the ones which have already tried to become the current
1173 lwkt_yield_quick(void)
1175 globaldata_t gd = mycpu;
1176 thread_t td = gd->gd_curthread;
1178 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1180 if (lwkt_resched_wanted()) {
1182 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1183 clear_lwkt_resched();
1185 lwkt_schedule_self(curthread);
1193 * This yield is designed for kernel threads with a user context.
1195 * The kernel acting on behalf of the user is potentially cpu-bound,
1196 * this function will efficiently allow other threads to run and also
1197 * switch to other processes by releasing.
1199 * The lwkt_user_yield() function is designed to have very low overhead
1200 * if no yield is determined to be needed.
1203 lwkt_user_yield(void)
1205 globaldata_t gd = mycpu;
1206 thread_t td = gd->gd_curthread;
1209 * Always run any pending interrupts in case we are in a critical
1212 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1216 * Switch (which forces a release) if another kernel thread needs
1217 * the cpu, if userland wants us to resched, or if our kernel
1218 * quantum has run out.
1220 if (lwkt_resched_wanted() ||
1221 user_resched_wanted())
1228 * Reacquire the current process if we are released.
1230 * XXX not implemented atm. The kernel may be holding locks and such,
1231 * so we want the thread to continue to receive cpu.
1233 if (td->td_release == NULL && lp) {
1234 lp->lwp_proc->p_usched->acquire_curproc(lp);
1235 td->td_release = lwkt_passive_release;
1236 lwkt_setpri_self(TDPRI_USER_NORM);
1242 * Generic schedule. Possibly schedule threads belonging to other cpus and
1243 * deal with threads that might be blocked on a wait queue.
1245 * We have a little helper inline function which does additional work after
1246 * the thread has been enqueued, including dealing with preemption and
1247 * setting need_lwkt_resched() (which prevents the kernel from returning
1248 * to userland until it has processed higher priority threads).
1250 * It is possible for this routine to be called after a failed _enqueue
1251 * (due to the target thread migrating, sleeping, or otherwise blocked).
1252 * We have to check that the thread is actually on the run queue!
1256 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1258 if (ntd->td_flags & TDF_RUNQ) {
1259 if (ntd->td_preemptable) {
1260 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1267 _lwkt_schedule(thread_t td)
1269 globaldata_t mygd = mycpu;
1271 KASSERT(td != &td->td_gd->gd_idlethread,
1272 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1273 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1274 crit_enter_gd(mygd);
1275 KKASSERT(td->td_lwp == NULL ||
1276 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1278 if (td == mygd->gd_curthread) {
1282 * If we own the thread, there is no race (since we are in a
1283 * critical section). If we do not own the thread there might
1284 * be a race but the target cpu will deal with it.
1286 if (td->td_gd == mygd) {
1288 _lwkt_schedule_post(mygd, td, 1);
1290 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1297 lwkt_schedule(thread_t td)
1303 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1309 * When scheduled remotely if frame != NULL the IPIQ is being
1310 * run via doreti or an interrupt then preemption can be allowed.
1312 * To allow preemption we have to drop the critical section so only
1313 * one is present in _lwkt_schedule_post.
1316 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1318 thread_t td = curthread;
1321 if (frame && ntd->td_preemptable) {
1322 crit_exit_noyield(td);
1323 _lwkt_schedule(ntd);
1324 crit_enter_quick(td);
1326 _lwkt_schedule(ntd);
1331 * Thread migration using a 'Pull' method. The thread may or may not be
1332 * the current thread. It MUST be descheduled and in a stable state.
1333 * lwkt_giveaway() must be called on the cpu owning the thread.
1335 * At any point after lwkt_giveaway() is called, the target cpu may
1336 * 'pull' the thread by calling lwkt_acquire().
1338 * We have to make sure the thread is not sitting on a per-cpu tsleep
1339 * queue or it will blow up when it moves to another cpu.
1341 * MPSAFE - must be called under very specific conditions.
1344 lwkt_giveaway(thread_t td)
1346 globaldata_t gd = mycpu;
1349 if (td->td_flags & TDF_TSLEEPQ)
1351 KKASSERT(td->td_gd == gd);
1352 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1353 td->td_flags |= TDF_MIGRATING;
1358 lwkt_acquire(thread_t td)
1362 int retry = 10000000;
1364 KKASSERT(td->td_flags & TDF_MIGRATING);
1369 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1370 crit_enter_gd(mygd);
1371 DEBUG_PUSH_INFO("lwkt_acquire");
1372 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1373 lwkt_process_ipiq();
1376 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1384 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1385 td->td_flags &= ~TDF_MIGRATING;
1388 crit_enter_gd(mygd);
1389 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1390 td->td_flags &= ~TDF_MIGRATING;
1396 * Generic deschedule. Descheduling threads other then your own should be
1397 * done only in carefully controlled circumstances. Descheduling is
1400 * This function may block if the cpu has run out of messages.
1403 lwkt_deschedule(thread_t td)
1406 if (td == curthread) {
1409 if (td->td_gd == mycpu) {
1412 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1419 * Set the target thread's priority. This routine does not automatically
1420 * switch to a higher priority thread, LWKT threads are not designed for
1421 * continuous priority changes. Yield if you want to switch.
1424 lwkt_setpri(thread_t td, int pri)
1426 if (td->td_pri != pri) {
1429 if (td->td_flags & TDF_RUNQ) {
1430 KKASSERT(td->td_gd == mycpu);
1442 * Set the initial priority for a thread prior to it being scheduled for
1443 * the first time. The thread MUST NOT be scheduled before or during
1444 * this call. The thread may be assigned to a cpu other then the current
1447 * Typically used after a thread has been created with TDF_STOPPREQ,
1448 * and before the thread is initially scheduled.
1451 lwkt_setpri_initial(thread_t td, int pri)
1454 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1459 lwkt_setpri_self(int pri)
1461 thread_t td = curthread;
1463 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1465 if (td->td_flags & TDF_RUNQ) {
1476 * hz tick scheduler clock for LWKT threads
1479 lwkt_schedulerclock(thread_t td)
1481 globaldata_t gd = td->td_gd;
1484 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1486 * If the current thread is at the head of the runq shift it to the
1487 * end of any equal-priority threads and request a LWKT reschedule
1490 * Ignore upri in this situation. There will only be one user thread
1491 * in user mode, all others will be user threads running in kernel
1492 * mode and we have to make sure they get some cpu.
1494 xtd = TAILQ_NEXT(td, td_threadq);
1495 if (xtd && xtd->td_pri == td->td_pri) {
1496 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1497 while (xtd && xtd->td_pri == td->td_pri)
1498 xtd = TAILQ_NEXT(xtd, td_threadq);
1500 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1502 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1503 need_lwkt_resched();
1507 * If we scheduled a thread other than the one at the head of the
1508 * queue always request a reschedule every tick.
1510 need_lwkt_resched();
1515 * Migrate the current thread to the specified cpu.
1517 * This is accomplished by descheduling ourselves from the current cpu
1518 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1519 * 'old' thread wants to migrate after it has been completely switched out
1520 * and will complete the migration.
1522 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1524 * We must be sure to release our current process designation (if a user
1525 * process) before clearing out any tsleepq we are on because the release
1526 * code may re-add us.
1528 * We must be sure to remove ourselves from the current cpu's tsleepq
1529 * before potentially moving to another queue. The thread can be on
1530 * a tsleepq due to a left-over tsleep_interlock().
1534 lwkt_setcpu_self(globaldata_t rgd)
1536 thread_t td = curthread;
1538 if (td->td_gd != rgd) {
1539 crit_enter_quick(td);
1543 if (td->td_flags & TDF_TSLEEPQ)
1547 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1548 * trying to deschedule ourselves and switch away, then deschedule
1549 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1550 * call lwkt_switch() to complete the operation.
1552 td->td_flags |= TDF_MIGRATING;
1553 lwkt_deschedule_self(td);
1554 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1555 td->td_migrate_gd = rgd;
1559 * We are now on the target cpu
1561 KKASSERT(rgd == mycpu);
1562 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1563 crit_exit_quick(td);
1568 lwkt_migratecpu(int cpuid)
1572 rgd = globaldata_find(cpuid);
1573 lwkt_setcpu_self(rgd);
1577 * Remote IPI for cpu migration (called while in a critical section so we
1578 * do not have to enter another one).
1580 * The thread (td) has already been completely descheduled from the
1581 * originating cpu and we can simply assert the case. The thread is
1582 * assigned to the new cpu and enqueued.
1584 * The thread will re-add itself to tdallq when it resumes execution.
1587 lwkt_setcpu_remote(void *arg)
1590 globaldata_t gd = mycpu;
1592 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1595 td->td_flags &= ~TDF_MIGRATING;
1596 KKASSERT(td->td_migrate_gd == NULL);
1597 KKASSERT(td->td_lwp == NULL ||
1598 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1603 lwkt_preempted_proc(void)
1605 thread_t td = curthread;
1606 while (td->td_preempted)
1607 td = td->td_preempted;
1612 * Create a kernel process/thread/whatever. It shares it's address space
1613 * with proc0 - ie: kernel only.
1615 * If the cpu is not specified one will be selected. In the future
1616 * specifying a cpu of -1 will enable kernel thread migration between
1620 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1621 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1626 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1630 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1633 * Set up arg0 for 'ps' etc
1635 __va_start(ap, fmt);
1636 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1640 * Schedule the thread to run
1642 if (td->td_flags & TDF_NOSTART)
1643 td->td_flags &= ~TDF_NOSTART;
1650 * Destroy an LWKT thread. Warning! This function is not called when
1651 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1652 * uses a different reaping mechanism.
1657 thread_t td = curthread;
1662 * Do any cleanup that might block here
1664 if (td->td_flags & TDF_VERBOSE)
1665 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1667 dsched_exit_thread(td);
1670 * Get us into a critical section to interlock gd_freetd and loop
1671 * until we can get it freed.
1673 * We have to cache the current td in gd_freetd because objcache_put()ing
1674 * it would rip it out from under us while our thread is still active.
1676 * We are the current thread so of course our own TDF_RUNNING bit will
1677 * be set, so unlike the lwp reap code we don't wait for it to clear.
1680 crit_enter_quick(td);
1683 tsleep(td, 0, "tdreap", 1);
1686 if ((std = gd->gd_freetd) != NULL) {
1687 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1688 gd->gd_freetd = NULL;
1689 objcache_put(thread_cache, std);
1696 * Remove thread resources from kernel lists and deschedule us for
1697 * the last time. We cannot block after this point or we may end
1698 * up with a stale td on the tsleepq.
1700 * None of this may block, the critical section is the only thing
1701 * protecting tdallq and the only thing preventing new lwkt_hold()
1704 if (td->td_flags & TDF_TSLEEPQ)
1706 lwkt_deschedule_self(td);
1707 lwkt_remove_tdallq(td);
1708 KKASSERT(td->td_refs == 0);
1713 KKASSERT(gd->gd_freetd == NULL);
1714 if (td->td_flags & TDF_ALLOCATED_THREAD)
1720 lwkt_remove_tdallq(thread_t td)
1722 KKASSERT(td->td_gd == mycpu);
1723 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1727 * Code reduction and branch prediction improvements. Call/return
1728 * overhead on modern cpus often degenerates into 0 cycles due to
1729 * the cpu's branch prediction hardware and return pc cache. We
1730 * can take advantage of this by not inlining medium-complexity
1731 * functions and we can also reduce the branch prediction impact
1732 * by collapsing perfectly predictable branches into a single
1733 * procedure instead of duplicating it.
1735 * Is any of this noticeable? Probably not, so I'll take the
1736 * smaller code size.
1739 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1741 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1747 thread_t td = curthread;
1748 int lcrit = td->td_critcount;
1750 td->td_critcount = 0;
1751 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1756 * Called from debugger/panic on cpus which have been stopped. We must still
1757 * process the IPIQ while stopped, even if we were stopped while in a critical
1760 * If we are dumping also try to process any pending interrupts. This may
1761 * or may not work depending on the state of the cpu at the point it was
1765 lwkt_smp_stopped(void)
1767 globaldata_t gd = mycpu;
1771 lwkt_process_ipiq();
1774 lwkt_process_ipiq();