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;
96 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
97 static void lwkt_setcpu_remote(void *arg);
100 extern void cpu_heavy_restore(void);
101 extern void cpu_lwkt_restore(void);
102 extern void cpu_kthread_restore(void);
103 extern void cpu_idle_restore(void);
106 * We can make all thread ports use the spin backend instead of the thread
107 * backend. This should only be set to debug the spin backend.
109 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
112 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
113 "Panic if attempting to switch lwkt's while mastering cpusync");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
116 "Number of switched threads");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
118 "Successful preemption events");
119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
120 "Failed preemption events");
121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
122 "Number of preempted threads.");
123 static int fairq_enable = 0;
124 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
125 &fairq_enable, 0, "Turn on fairq priority accumulators");
126 static int fairq_bypass = -1;
127 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
128 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
129 extern int lwkt_sched_debug;
130 int lwkt_sched_debug = 0;
131 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
132 &lwkt_sched_debug, 0, "Scheduler debug");
133 static int lwkt_spin_loops = 10;
134 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
135 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
136 static int lwkt_spin_reseq = 0;
137 SYSCTL_INT(_lwkt, OID_AUTO, spin_reseq, CTLFLAG_RW,
138 &lwkt_spin_reseq, 0, "Scheduler resequencer enable");
139 static int lwkt_spin_monitor = 0;
140 SYSCTL_INT(_lwkt, OID_AUTO, spin_monitor, CTLFLAG_RW,
141 &lwkt_spin_monitor, 0, "Scheduler uses monitor/mwait");
142 static int lwkt_spin_fatal = 0; /* disabled */
143 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
144 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
145 static int preempt_enable = 1;
146 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
147 &preempt_enable, 0, "Enable preemption");
148 static int lwkt_cache_threads = 0;
149 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
150 &lwkt_cache_threads, 0, "thread+kstack cache");
152 static __cachealign int lwkt_cseq_rindex;
153 static __cachealign int lwkt_cseq_windex;
156 * These helper procedures handle the runq, they can only be called from
157 * within a critical section.
159 * WARNING! Prior to SMP being brought up it is possible to enqueue and
160 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
161 * instead of 'mycpu' when referencing the globaldata structure. Once
162 * SMP live enqueuing and dequeueing only occurs on the current cpu.
166 _lwkt_dequeue(thread_t td)
168 if (td->td_flags & TDF_RUNQ) {
169 struct globaldata *gd = td->td_gd;
171 td->td_flags &= ~TDF_RUNQ;
172 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
173 --gd->gd_tdrunqcount;
174 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
175 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
182 * There are a limited number of lwkt threads runnable since user
183 * processes only schedule one at a time per cpu. However, there can
184 * be many user processes in kernel mode exiting from a tsleep() which
187 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
188 * will ignore user priority. This is to ensure that user threads in
189 * kernel mode get cpu at some point regardless of what the user
194 _lwkt_enqueue(thread_t td)
198 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
199 struct globaldata *gd = td->td_gd;
201 td->td_flags |= TDF_RUNQ;
202 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
204 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
205 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
208 * NOTE: td_upri - higher numbers more desireable, same sense
209 * as td_pri (typically reversed from lwp_upri).
211 * In the equal priority case we want the best selection
212 * at the beginning so the less desireable selections know
213 * that they have to setrunqueue/go-to-another-cpu, even
214 * though it means switching back to the 'best' selection.
215 * This also avoids degenerate situations when many threads
216 * are runnable or waking up at the same time.
218 * If upri matches exactly place at end/round-robin.
221 (xtd->td_pri >= td->td_pri ||
222 (xtd->td_pri == td->td_pri &&
223 xtd->td_upri >= td->td_upri))) {
224 xtd = TAILQ_NEXT(xtd, td_threadq);
227 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
229 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
231 ++gd->gd_tdrunqcount;
234 * Request a LWKT reschedule if we are now at the head of the queue.
236 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
242 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
244 struct thread *td = (struct thread *)obj;
246 td->td_kstack = NULL;
247 td->td_kstack_size = 0;
248 td->td_flags = TDF_ALLOCATED_THREAD;
254 _lwkt_thread_dtor(void *obj, void *privdata)
256 struct thread *td = (struct thread *)obj;
258 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
259 ("_lwkt_thread_dtor: not allocated from objcache"));
260 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
261 td->td_kstack_size > 0,
262 ("_lwkt_thread_dtor: corrupted stack"));
263 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
264 td->td_kstack = NULL;
269 * Initialize the lwkt s/system.
271 * Nominally cache up to 32 thread + kstack structures. Cache more on
272 * systems with a lot of cpu cores.
277 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
278 if (lwkt_cache_threads == 0) {
279 lwkt_cache_threads = ncpus * 4;
280 if (lwkt_cache_threads < 32)
281 lwkt_cache_threads = 32;
283 thread_cache = objcache_create_mbacked(
284 M_THREAD, sizeof(struct thread),
285 0, lwkt_cache_threads,
286 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
290 * Schedule a thread to run. As the current thread we can always safely
291 * schedule ourselves, and a shortcut procedure is provided for that
294 * (non-blocking, self contained on a per cpu basis)
297 lwkt_schedule_self(thread_t td)
299 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
300 crit_enter_quick(td);
301 KASSERT(td != &td->td_gd->gd_idlethread,
302 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
303 KKASSERT(td->td_lwp == NULL ||
304 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
310 * Deschedule a thread.
312 * (non-blocking, self contained on a per cpu basis)
315 lwkt_deschedule_self(thread_t td)
317 crit_enter_quick(td);
323 * LWKTs operate on a per-cpu basis
325 * WARNING! Called from early boot, 'mycpu' may not work yet.
328 lwkt_gdinit(struct globaldata *gd)
330 TAILQ_INIT(&gd->gd_tdrunq);
331 TAILQ_INIT(&gd->gd_tdallq);
335 * Create a new thread. The thread must be associated with a process context
336 * or LWKT start address before it can be scheduled. If the target cpu is
337 * -1 the thread will be created on the current cpu.
339 * If you intend to create a thread without a process context this function
340 * does everything except load the startup and switcher function.
343 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
345 static int cpu_rotator;
346 globaldata_t gd = mycpu;
350 * If static thread storage is not supplied allocate a thread. Reuse
351 * a cached free thread if possible. gd_freetd is used to keep an exiting
352 * thread intact through the exit.
356 if ((td = gd->gd_freetd) != NULL) {
357 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
359 gd->gd_freetd = NULL;
361 td = objcache_get(thread_cache, M_WAITOK);
362 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
366 KASSERT((td->td_flags &
367 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
368 TDF_ALLOCATED_THREAD,
369 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
370 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
374 * Try to reuse cached stack.
376 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
377 if (flags & TDF_ALLOCATED_STACK) {
378 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
383 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
384 flags |= TDF_ALLOCATED_STACK;
391 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
396 * Initialize a preexisting thread structure. This function is used by
397 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
399 * All threads start out in a critical section at a priority of
400 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
401 * appropriate. This function may send an IPI message when the
402 * requested cpu is not the current cpu and consequently gd_tdallq may
403 * not be initialized synchronously from the point of view of the originating
406 * NOTE! we have to be careful in regards to creating threads for other cpus
407 * if SMP has not yet been activated.
412 lwkt_init_thread_remote(void *arg)
417 * Protected by critical section held by IPI dispatch
419 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
425 * lwkt core thread structural initialization.
427 * NOTE: All threads are initialized as mpsafe threads.
430 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
431 struct globaldata *gd)
433 globaldata_t mygd = mycpu;
435 bzero(td, sizeof(struct thread));
436 td->td_kstack = stack;
437 td->td_kstack_size = stksize;
438 td->td_flags = flags;
441 td->td_pri = TDPRI_KERN_DAEMON;
442 td->td_critcount = 1;
443 td->td_toks_have = NULL;
444 td->td_toks_stop = &td->td_toks_base;
445 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
446 lwkt_initport_spin(&td->td_msgport, td);
448 lwkt_initport_thread(&td->td_msgport, td);
449 pmap_init_thread(td);
452 * Normally initializing a thread for a remote cpu requires sending an
453 * IPI. However, the idlethread is setup before the other cpus are
454 * activated so we have to treat it as a special case. XXX manipulation
455 * of gd_tdallq requires the BGL.
457 if (gd == mygd || td == &gd->gd_idlethread) {
459 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
462 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
466 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
470 dsched_new_thread(td);
474 lwkt_set_comm(thread_t td, const char *ctl, ...)
479 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
481 KTR_LOG(ctxsw_newtd, td, td->td_comm);
485 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
486 * this does not prevent the thread from migrating to another cpu so the
487 * gd_tdallq state is not protected by this.
490 lwkt_hold(thread_t td)
492 atomic_add_int(&td->td_refs, 1);
496 lwkt_rele(thread_t td)
498 KKASSERT(td->td_refs > 0);
499 atomic_add_int(&td->td_refs, -1);
503 lwkt_free_thread(thread_t td)
505 KKASSERT(td->td_refs == 0);
506 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
507 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
508 if (td->td_flags & TDF_ALLOCATED_THREAD) {
509 objcache_put(thread_cache, td);
510 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
511 /* client-allocated struct with internally allocated stack */
512 KASSERT(td->td_kstack && td->td_kstack_size > 0,
513 ("lwkt_free_thread: corrupted stack"));
514 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
515 td->td_kstack = NULL;
516 td->td_kstack_size = 0;
518 KTR_LOG(ctxsw_deadtd, td);
523 * Switch to the next runnable lwkt. If no LWKTs are runnable then
524 * switch to the idlethread. Switching must occur within a critical
525 * section to avoid races with the scheduling queue.
527 * We always have full control over our cpu's run queue. Other cpus
528 * that wish to manipulate our queue must use the cpu_*msg() calls to
529 * talk to our cpu, so a critical section is all that is needed and
530 * the result is very, very fast thread switching.
532 * The LWKT scheduler uses a fixed priority model and round-robins at
533 * each priority level. User process scheduling is a totally
534 * different beast and LWKT priorities should not be confused with
535 * user process priorities.
537 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
538 * is not called by the current thread in the preemption case, only when
539 * the preempting thread blocks (in order to return to the original thread).
541 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
542 * migration and tsleep deschedule the current lwkt thread and call
543 * lwkt_switch(). In particular, the target cpu of the migration fully
544 * expects the thread to become non-runnable and can deadlock against
545 * cpusync operations if we run any IPIs prior to switching the thread out.
547 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
548 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
553 globaldata_t gd = mycpu;
554 thread_t td = gd->gd_curthread;
558 KKASSERT(gd->gd_processing_ipiq == 0);
559 KKASSERT(td->td_flags & TDF_RUNNING);
562 * Switching from within a 'fast' (non thread switched) interrupt or IPI
563 * is illegal. However, we may have to do it anyway if we hit a fatal
564 * kernel trap or we have paniced.
566 * If this case occurs save and restore the interrupt nesting level.
568 if (gd->gd_intr_nesting_level) {
572 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
573 panic("lwkt_switch: Attempt to switch from a "
574 "fast interrupt, ipi, or hard code section, "
578 savegdnest = gd->gd_intr_nesting_level;
579 savegdtrap = gd->gd_trap_nesting_level;
580 gd->gd_intr_nesting_level = 0;
581 gd->gd_trap_nesting_level = 0;
582 if ((td->td_flags & TDF_PANICWARN) == 0) {
583 td->td_flags |= TDF_PANICWARN;
584 kprintf("Warning: thread switch from interrupt, IPI, "
585 "or hard code section.\n"
586 "thread %p (%s)\n", td, td->td_comm);
590 gd->gd_intr_nesting_level = savegdnest;
591 gd->gd_trap_nesting_level = savegdtrap;
597 * Release our current user process designation if we are blocking
598 * or if a user reschedule was requested.
600 * NOTE: This function is NOT called if we are switching into or
601 * returning from a preemption.
603 * NOTE: Releasing our current user process designation may cause
604 * it to be assigned to another thread, which in turn will
605 * cause us to block in the usched acquire code when we attempt
606 * to return to userland.
608 * NOTE: On SMP systems this can be very nasty when heavy token
609 * contention is present so we want to be careful not to
610 * release the designation gratuitously.
612 if (td->td_release &&
613 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
621 if (TD_TOKS_HELD(td))
622 lwkt_relalltokens(td);
625 * We had better not be holding any spin locks, but don't get into an
626 * endless panic loop.
628 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
629 ("lwkt_switch: still holding %d exclusive spinlocks!",
635 if (td->td_cscount) {
636 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
638 if (panic_on_cscount)
639 panic("switching while mastering cpusync");
645 * If we had preempted another thread on this cpu, resume the preempted
646 * thread. This occurs transparently, whether the preempted thread
647 * was scheduled or not (it may have been preempted after descheduling
650 * We have to setup the MP lock for the original thread after backing
651 * out the adjustment that was made to curthread when the original
654 if ((ntd = td->td_preempted) != NULL) {
655 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
656 ntd->td_flags |= TDF_PREEMPT_DONE;
659 * The interrupt may have woken a thread up, we need to properly
660 * set the reschedule flag if the originally interrupted thread is
661 * at a lower priority.
663 * The interrupt may not have descheduled.
665 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
667 goto havethread_preempted;
671 * If we cannot obtain ownership of the tokens we cannot immediately
672 * schedule the target thread.
674 * Reminder: Again, we cannot afford to run any IPIs in this path if
675 * the current thread has been descheduled.
678 clear_lwkt_resched();
681 * Hotpath - pull the head of the run queue and attempt to schedule
684 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
688 * Runq is empty, switch to idle to allow it to halt.
690 ntd = &gd->gd_idlethread;
692 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
693 ASSERT_NO_TOKENS_HELD(ntd);
695 cpu_time.cp_msg[0] = 0;
696 cpu_time.cp_stallpc = 0;
701 * Hotpath - schedule ntd.
703 * NOTE: For UP there is no mplock and lwkt_getalltokens()
706 if (TD_TOKS_NOT_HELD(ntd) ||
707 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
713 * Coldpath (SMP only since tokens always succeed on UP)
715 * We had some contention on the thread we wanted to schedule.
716 * What we do now is try to find a thread that we can schedule
719 * The coldpath scan does NOT rearrange threads in the run list.
720 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
721 * the next tick whenever the current head is not the current thread.
726 ++gd->gd_cnt.v_token_colls;
728 if (fairq_bypass > 0)
731 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
732 #ifndef NO_LWKT_SPLIT_USERPRI
734 * Never schedule threads returning to userland or the
735 * user thread scheduler helper thread when higher priority
736 * threads are present. The runq is sorted by priority
737 * so we can give up traversing it when we find the first
738 * low priority thread.
740 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
749 if (TD_TOKS_NOT_HELD(ntd) ||
750 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
756 ++gd->gd_cnt.v_token_colls;
761 * We exhausted the run list, meaning that all runnable threads
765 ntd = &gd->gd_idlethread;
767 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
768 ASSERT_NO_TOKENS_HELD(ntd);
769 /* contention case, do not clear contention mask */
773 * We are going to have to retry but if the current thread is not
774 * on the runq we instead switch through the idle thread to get away
775 * from the current thread. We have to flag for lwkt reschedule
776 * to prevent the idle thread from halting.
778 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
779 * instruct it to deal with the potential for deadlocks by
780 * ordering the tokens by address.
782 if ((td->td_flags & TDF_RUNQ) == 0) {
783 need_lwkt_resched(); /* prevent hlt */
786 #if defined(INVARIANTS) && defined(__amd64__)
787 if ((read_rflags() & PSL_I) == 0) {
789 panic("lwkt_switch() called with interrupts disabled");
794 * Number iterations so far. After a certain point we switch to
795 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
797 if (spinning < 0x7FFFFFFF)
802 * lwkt_getalltokens() failed in sorted token mode, we can use
803 * monitor/mwait in this case.
805 if (spinning >= lwkt_spin_loops &&
806 (cpu_mi_feature & CPU_MI_MONITOR) &&
809 cpu_mmw_pause_int(&gd->gd_reqflags,
810 (gd->gd_reqflags | RQF_SPINNING) &
811 ~RQF_IDLECHECK_WK_MASK);
816 * We already checked that td is still scheduled so this should be
822 * This experimental resequencer is used as a fall-back to reduce
823 * hw cache line contention by placing each core's scheduler into a
824 * time-domain-multplexed slot.
826 * The resequencer is disabled by default. It's functionality has
827 * largely been superceeded by the token algorithm which limits races
828 * to a subset of cores.
830 * The resequencer algorithm tends to break down when more than
831 * 20 cores are contending. What appears to happen is that new
832 * tokens can be obtained out of address-sorted order by new cores
833 * while existing cores languish in long delays between retries and
834 * wind up being starved-out of the token acquisition.
836 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
837 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
840 while ((oseq = lwkt_cseq_rindex) != cseq) {
843 if (cpu_mi_feature & CPU_MI_MONITOR) {
844 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
854 atomic_add_int(&lwkt_cseq_rindex, 1);
856 /* highest level for(;;) loop */
861 * Clear gd_idle_repeat when doing a normal switch to a non-idle
864 ntd->td_wmesg = NULL;
865 ++gd->gd_cnt.v_swtch;
866 gd->gd_idle_repeat = 0;
868 havethread_preempted:
870 * If the new target does not need the MP lock and we are holding it,
871 * release the MP lock. If the new target requires the MP lock we have
872 * already acquired it for the target.
876 KASSERT(ntd->td_critcount,
877 ("priority problem in lwkt_switch %d %d",
878 td->td_critcount, ntd->td_critcount));
882 * Execute the actual thread switch operation. This function
883 * returns to the current thread and returns the previous thread
884 * (which may be different from the thread we switched to).
886 * We are responsible for marking ntd as TDF_RUNNING.
888 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
890 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
891 ntd->td_flags |= TDF_RUNNING;
892 lwkt_switch_return(td->td_switch(ntd));
893 /* ntd invalid, td_switch() can return a different thread_t */
897 * catch-all. XXX is this strictly needed?
901 /* NOTE: current cpu may have changed after switch */
906 * Called by assembly in the td_switch (thread restore path) for thread
907 * bootstrap cases which do not 'return' to lwkt_switch().
910 lwkt_switch_return(thread_t otd)
916 * Check if otd was migrating. Now that we are on ntd we can finish
917 * up the migration. This is a bit messy but it is the only place
918 * where td is known to be fully descheduled.
920 * We can only activate the migration if otd was migrating but not
921 * held on the cpu due to a preemption chain. We still have to
922 * clear TDF_RUNNING on the old thread either way.
924 * We are responsible for clearing the previously running thread's
927 if ((rgd = otd->td_migrate_gd) != NULL &&
928 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
929 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
930 (TDF_MIGRATING | TDF_RUNNING));
931 otd->td_migrate_gd = NULL;
932 otd->td_flags &= ~TDF_RUNNING;
933 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
935 otd->td_flags &= ~TDF_RUNNING;
938 otd->td_flags &= ~TDF_RUNNING;
942 * Final exit validations (see lwp_wait()). Note that otd becomes
943 * invalid the *instant* we set TDF_MP_EXITSIG.
945 while (otd->td_flags & TDF_EXITING) {
948 mpflags = otd->td_mpflags;
951 if (mpflags & TDF_MP_EXITWAIT) {
952 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
953 mpflags | TDF_MP_EXITSIG)) {
958 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
959 mpflags | TDF_MP_EXITSIG)) {
968 * Request that the target thread preempt the current thread. Preemption
969 * can only occur if our only critical section is the one that we were called
970 * with, the relative priority of the target thread is higher, and the target
971 * thread holds no tokens. This also only works if we are not holding any
972 * spinlocks (obviously).
974 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
975 * this is called via lwkt_schedule() through the td_preemptable callback.
976 * critcount is the managed critical priority that we should ignore in order
977 * to determine whether preemption is possible (aka usually just the crit
978 * priority of lwkt_schedule() itself).
980 * Preemption is typically limited to interrupt threads.
982 * Operation works in a fairly straight-forward manner. The normal
983 * scheduling code is bypassed and we switch directly to the target
984 * thread. When the target thread attempts to block or switch away
985 * code at the base of lwkt_switch() will switch directly back to our
986 * thread. Our thread is able to retain whatever tokens it holds and
987 * if the target needs one of them the target will switch back to us
988 * and reschedule itself normally.
991 lwkt_preempt(thread_t ntd, int critcount)
993 struct globaldata *gd = mycpu;
996 int save_gd_intr_nesting_level;
999 * The caller has put us in a critical section. We can only preempt
1000 * if the caller of the caller was not in a critical section (basically
1001 * a local interrupt), as determined by the 'critcount' parameter. We
1002 * also can't preempt if the caller is holding any spinlocks (even if
1003 * he isn't in a critical section). This also handles the tokens test.
1005 * YYY The target thread must be in a critical section (else it must
1006 * inherit our critical section? I dunno yet).
1008 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1010 td = gd->gd_curthread;
1011 if (preempt_enable == 0) {
1015 if (ntd->td_pri <= td->td_pri) {
1019 if (td->td_critcount > critcount) {
1024 if (td->td_cscount) {
1028 if (ntd->td_gd != gd) {
1034 * We don't have to check spinlocks here as they will also bump
1037 * Do not try to preempt if the target thread is holding any tokens.
1038 * We could try to acquire the tokens but this case is so rare there
1039 * is no need to support it.
1041 KKASSERT(gd->gd_spinlocks == 0);
1043 if (TD_TOKS_HELD(ntd)) {
1047 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1051 if (ntd->td_preempted) {
1055 KKASSERT(gd->gd_processing_ipiq == 0);
1058 * Since we are able to preempt the current thread, there is no need to
1059 * call need_lwkt_resched().
1061 * We must temporarily clear gd_intr_nesting_level around the switch
1062 * since switchouts from the target thread are allowed (they will just
1063 * return to our thread), and since the target thread has its own stack.
1065 * A preemption must switch back to the original thread, assert the
1069 ntd->td_preempted = td;
1070 td->td_flags |= TDF_PREEMPT_LOCK;
1071 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1072 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1073 gd->gd_intr_nesting_level = 0;
1075 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1076 ntd->td_flags |= TDF_RUNNING;
1077 xtd = td->td_switch(ntd);
1078 KKASSERT(xtd == ntd);
1079 lwkt_switch_return(xtd);
1080 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1082 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1083 ntd->td_preempted = NULL;
1084 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1088 * Conditionally call splz() if gd_reqflags indicates work is pending.
1089 * This will work inside a critical section but not inside a hard code
1092 * (self contained on a per cpu basis)
1097 globaldata_t gd = mycpu;
1098 thread_t td = gd->gd_curthread;
1100 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1101 gd->gd_intr_nesting_level == 0 &&
1102 td->td_nest_count < 2)
1109 * This version is integrated into crit_exit, reqflags has already
1110 * been tested but td_critcount has not.
1112 * We only want to execute the splz() on the 1->0 transition of
1113 * critcount and not in a hard code section or if too deeply nested.
1115 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
1118 lwkt_maybe_splz(thread_t td)
1120 globaldata_t gd = td->td_gd;
1122 if (td->td_critcount == 0 &&
1123 gd->gd_intr_nesting_level == 0 &&
1124 td->td_nest_count < 2)
1131 * Drivers which set up processing co-threads can call this function to
1132 * run the co-thread at a higher priority and to allow it to preempt
1136 lwkt_set_interrupt_support_thread(void)
1138 thread_t td = curthread;
1140 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1141 td->td_flags |= TDF_INTTHREAD;
1142 td->td_preemptable = lwkt_preempt;
1147 * This function is used to negotiate a passive release of the current
1148 * process/lwp designation with the user scheduler, allowing the user
1149 * scheduler to schedule another user thread. The related kernel thread
1150 * (curthread) continues running in the released state.
1153 lwkt_passive_release(struct thread *td)
1155 struct lwp *lp = td->td_lwp;
1157 #ifndef NO_LWKT_SPLIT_USERPRI
1158 td->td_release = NULL;
1159 lwkt_setpri_self(TDPRI_KERN_USER);
1162 lp->lwp_proc->p_usched->release_curproc(lp);
1167 * This implements a LWKT yield, allowing a kernel thread to yield to other
1168 * kernel threads at the same or higher priority. This function can be
1169 * called in a tight loop and will typically only yield once per tick.
1171 * Most kernel threads run at the same priority in order to allow equal
1174 * (self contained on a per cpu basis)
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()) {
1185 lwkt_schedule_self(curthread);
1191 * The quick version processes pending interrupts and higher-priority
1192 * LWKT threads but will not round-robin same-priority LWKT threads.
1194 * When called while attempting to return to userland the only same-pri
1195 * threads are the ones which have already tried to become the current
1199 lwkt_yield_quick(void)
1201 globaldata_t gd = mycpu;
1202 thread_t td = gd->gd_curthread;
1204 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1206 if (lwkt_resched_wanted()) {
1207 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1208 clear_lwkt_resched();
1210 lwkt_schedule_self(curthread);
1217 * This yield is designed for kernel threads with a user context.
1219 * The kernel acting on behalf of the user is potentially cpu-bound,
1220 * this function will efficiently allow other threads to run and also
1221 * switch to other processes by releasing.
1223 * The lwkt_user_yield() function is designed to have very low overhead
1224 * if no yield is determined to be needed.
1227 lwkt_user_yield(void)
1229 globaldata_t gd = mycpu;
1230 thread_t td = gd->gd_curthread;
1233 * Always run any pending interrupts in case we are in a critical
1236 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1240 * Switch (which forces a release) if another kernel thread needs
1241 * the cpu, if userland wants us to resched, or if our kernel
1242 * quantum has run out.
1244 if (lwkt_resched_wanted() ||
1245 user_resched_wanted())
1252 * Reacquire the current process if we are released.
1254 * XXX not implemented atm. The kernel may be holding locks and such,
1255 * so we want the thread to continue to receive cpu.
1257 if (td->td_release == NULL && lp) {
1258 lp->lwp_proc->p_usched->acquire_curproc(lp);
1259 td->td_release = lwkt_passive_release;
1260 lwkt_setpri_self(TDPRI_USER_NORM);
1266 * Generic schedule. Possibly schedule threads belonging to other cpus and
1267 * deal with threads that might be blocked on a wait queue.
1269 * We have a little helper inline function which does additional work after
1270 * the thread has been enqueued, including dealing with preemption and
1271 * setting need_lwkt_resched() (which prevents the kernel from returning
1272 * to userland until it has processed higher priority threads).
1274 * It is possible for this routine to be called after a failed _enqueue
1275 * (due to the target thread migrating, sleeping, or otherwise blocked).
1276 * We have to check that the thread is actually on the run queue!
1280 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1282 if (ntd->td_flags & TDF_RUNQ) {
1283 if (ntd->td_preemptable) {
1284 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1291 _lwkt_schedule(thread_t td)
1293 globaldata_t mygd = mycpu;
1295 KASSERT(td != &td->td_gd->gd_idlethread,
1296 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1297 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1298 crit_enter_gd(mygd);
1299 KKASSERT(td->td_lwp == NULL ||
1300 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1302 if (td == mygd->gd_curthread) {
1306 * If we own the thread, there is no race (since we are in a
1307 * critical section). If we do not own the thread there might
1308 * be a race but the target cpu will deal with it.
1311 if (td->td_gd == mygd) {
1313 _lwkt_schedule_post(mygd, td, 1);
1315 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1319 _lwkt_schedule_post(mygd, td, 1);
1326 lwkt_schedule(thread_t td)
1332 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1340 * When scheduled remotely if frame != NULL the IPIQ is being
1341 * run via doreti or an interrupt then preemption can be allowed.
1343 * To allow preemption we have to drop the critical section so only
1344 * one is present in _lwkt_schedule_post.
1347 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1349 thread_t td = curthread;
1352 if (frame && ntd->td_preemptable) {
1353 crit_exit_noyield(td);
1354 _lwkt_schedule(ntd);
1355 crit_enter_quick(td);
1357 _lwkt_schedule(ntd);
1362 * Thread migration using a 'Pull' method. The thread may or may not be
1363 * the current thread. It MUST be descheduled and in a stable state.
1364 * lwkt_giveaway() must be called on the cpu owning the thread.
1366 * At any point after lwkt_giveaway() is called, the target cpu may
1367 * 'pull' the thread by calling lwkt_acquire().
1369 * We have to make sure the thread is not sitting on a per-cpu tsleep
1370 * queue or it will blow up when it moves to another cpu.
1372 * MPSAFE - must be called under very specific conditions.
1375 lwkt_giveaway(thread_t td)
1377 globaldata_t gd = mycpu;
1380 if (td->td_flags & TDF_TSLEEPQ)
1382 KKASSERT(td->td_gd == gd);
1383 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1384 td->td_flags |= TDF_MIGRATING;
1389 lwkt_acquire(thread_t td)
1393 int retry = 10000000;
1395 KKASSERT(td->td_flags & TDF_MIGRATING);
1400 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1401 crit_enter_gd(mygd);
1402 DEBUG_PUSH_INFO("lwkt_acquire");
1403 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1405 lwkt_process_ipiq();
1409 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1417 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1418 td->td_flags &= ~TDF_MIGRATING;
1421 crit_enter_gd(mygd);
1422 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1423 td->td_flags &= ~TDF_MIGRATING;
1431 * Generic deschedule. Descheduling threads other then your own should be
1432 * done only in carefully controlled circumstances. Descheduling is
1435 * This function may block if the cpu has run out of messages.
1438 lwkt_deschedule(thread_t td)
1442 if (td == curthread) {
1445 if (td->td_gd == mycpu) {
1448 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1458 * Set the target thread's priority. This routine does not automatically
1459 * switch to a higher priority thread, LWKT threads are not designed for
1460 * continuous priority changes. Yield if you want to switch.
1463 lwkt_setpri(thread_t td, int pri)
1465 if (td->td_pri != pri) {
1468 if (td->td_flags & TDF_RUNQ) {
1469 KKASSERT(td->td_gd == mycpu);
1481 * Set the initial priority for a thread prior to it being scheduled for
1482 * the first time. The thread MUST NOT be scheduled before or during
1483 * this call. The thread may be assigned to a cpu other then the current
1486 * Typically used after a thread has been created with TDF_STOPPREQ,
1487 * and before the thread is initially scheduled.
1490 lwkt_setpri_initial(thread_t td, int pri)
1493 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1498 lwkt_setpri_self(int pri)
1500 thread_t td = curthread;
1502 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1504 if (td->td_flags & TDF_RUNQ) {
1515 * hz tick scheduler clock for LWKT threads
1518 lwkt_schedulerclock(thread_t td)
1520 globaldata_t gd = td->td_gd;
1523 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1525 * If the current thread is at the head of the runq shift it to the
1526 * end of any equal-priority threads and request a LWKT reschedule
1529 * Ignore upri in this situation. There will only be one user thread
1530 * in user mode, all others will be user threads running in kernel
1531 * mode and we have to make sure they get some cpu.
1533 xtd = TAILQ_NEXT(td, td_threadq);
1534 if (xtd && xtd->td_pri == td->td_pri) {
1535 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1536 while (xtd && xtd->td_pri == td->td_pri)
1537 xtd = TAILQ_NEXT(xtd, td_threadq);
1539 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1541 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1542 need_lwkt_resched();
1546 * If we scheduled a thread other than the one at the head of the
1547 * queue always request a reschedule every tick.
1549 need_lwkt_resched();
1554 * Migrate the current thread to the specified cpu.
1556 * This is accomplished by descheduling ourselves from the current cpu
1557 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1558 * 'old' thread wants to migrate after it has been completely switched out
1559 * and will complete the migration.
1561 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1563 * We must be sure to release our current process designation (if a user
1564 * process) before clearing out any tsleepq we are on because the release
1565 * code may re-add us.
1567 * We must be sure to remove ourselves from the current cpu's tsleepq
1568 * before potentially moving to another queue. The thread can be on
1569 * a tsleepq due to a left-over tsleep_interlock().
1573 lwkt_setcpu_self(globaldata_t rgd)
1576 thread_t td = curthread;
1578 if (td->td_gd != rgd) {
1579 crit_enter_quick(td);
1583 if (td->td_flags & TDF_TSLEEPQ)
1587 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1588 * trying to deschedule ourselves and switch away, then deschedule
1589 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1590 * call lwkt_switch() to complete the operation.
1592 td->td_flags |= TDF_MIGRATING;
1593 lwkt_deschedule_self(td);
1594 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1595 td->td_migrate_gd = rgd;
1599 * We are now on the target cpu
1601 KKASSERT(rgd == mycpu);
1602 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1603 crit_exit_quick(td);
1609 lwkt_migratecpu(int cpuid)
1614 rgd = globaldata_find(cpuid);
1615 lwkt_setcpu_self(rgd);
1621 * Remote IPI for cpu migration (called while in a critical section so we
1622 * do not have to enter another one).
1624 * The thread (td) has already been completely descheduled from the
1625 * originating cpu and we can simply assert the case. The thread is
1626 * assigned to the new cpu and enqueued.
1628 * The thread will re-add itself to tdallq when it resumes execution.
1631 lwkt_setcpu_remote(void *arg)
1634 globaldata_t gd = mycpu;
1636 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1639 td->td_flags &= ~TDF_MIGRATING;
1640 KKASSERT(td->td_migrate_gd == NULL);
1641 KKASSERT(td->td_lwp == NULL ||
1642 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1648 lwkt_preempted_proc(void)
1650 thread_t td = curthread;
1651 while (td->td_preempted)
1652 td = td->td_preempted;
1657 * Create a kernel process/thread/whatever. It shares it's address space
1658 * with proc0 - ie: kernel only.
1660 * If the cpu is not specified one will be selected. In the future
1661 * specifying a cpu of -1 will enable kernel thread migration between
1665 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1666 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1671 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1675 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1678 * Set up arg0 for 'ps' etc
1680 __va_start(ap, fmt);
1681 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1685 * Schedule the thread to run
1687 if (td->td_flags & TDF_NOSTART)
1688 td->td_flags &= ~TDF_NOSTART;
1695 * Destroy an LWKT thread. Warning! This function is not called when
1696 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1697 * uses a different reaping mechanism.
1702 thread_t td = curthread;
1707 * Do any cleanup that might block here
1709 if (td->td_flags & TDF_VERBOSE)
1710 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1712 dsched_exit_thread(td);
1715 * Get us into a critical section to interlock gd_freetd and loop
1716 * until we can get it freed.
1718 * We have to cache the current td in gd_freetd because objcache_put()ing
1719 * it would rip it out from under us while our thread is still active.
1721 * We are the current thread so of course our own TDF_RUNNING bit will
1722 * be set, so unlike the lwp reap code we don't wait for it to clear.
1725 crit_enter_quick(td);
1728 tsleep(td, 0, "tdreap", 1);
1731 if ((std = gd->gd_freetd) != NULL) {
1732 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1733 gd->gd_freetd = NULL;
1734 objcache_put(thread_cache, std);
1741 * Remove thread resources from kernel lists and deschedule us for
1742 * the last time. We cannot block after this point or we may end
1743 * up with a stale td on the tsleepq.
1745 * None of this may block, the critical section is the only thing
1746 * protecting tdallq and the only thing preventing new lwkt_hold()
1749 if (td->td_flags & TDF_TSLEEPQ)
1751 lwkt_deschedule_self(td);
1752 lwkt_remove_tdallq(td);
1753 KKASSERT(td->td_refs == 0);
1758 KKASSERT(gd->gd_freetd == NULL);
1759 if (td->td_flags & TDF_ALLOCATED_THREAD)
1765 lwkt_remove_tdallq(thread_t td)
1767 KKASSERT(td->td_gd == mycpu);
1768 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1772 * Code reduction and branch prediction improvements. Call/return
1773 * overhead on modern cpus often degenerates into 0 cycles due to
1774 * the cpu's branch prediction hardware and return pc cache. We
1775 * can take advantage of this by not inlining medium-complexity
1776 * functions and we can also reduce the branch prediction impact
1777 * by collapsing perfectly predictable branches into a single
1778 * procedure instead of duplicating it.
1780 * Is any of this noticeable? Probably not, so I'll take the
1781 * smaller code size.
1784 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1786 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1792 thread_t td = curthread;
1793 int lcrit = td->td_critcount;
1795 td->td_critcount = 0;
1796 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1803 * Called from debugger/panic on cpus which have been stopped. We must still
1804 * process the IPIQ while stopped, even if we were stopped while in a critical
1807 * If we are dumping also try to process any pending interrupts. This may
1808 * or may not work depending on the state of the cpu at the point it was
1812 lwkt_smp_stopped(void)
1814 globaldata_t gd = mycpu;
1818 lwkt_process_ipiq();
1821 lwkt_process_ipiq();