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>
54 #include <sys/spinlock.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mplock2.h>
61 #include <sys/dsched.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
72 #include <machine/stdarg.h>
73 #include <machine/smp.h>
75 #if !defined(KTR_CTXSW)
76 #define KTR_CTXSW KTR_ALL
78 KTR_INFO_MASTER(ctxsw);
79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
81 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
82 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
84 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
87 static int panic_on_cscount = 0;
89 static __int64_t switch_count = 0;
90 static __int64_t preempt_hit = 0;
91 static __int64_t preempt_miss = 0;
92 static __int64_t preempt_weird = 0;
93 static int lwkt_use_spin_port;
94 static struct objcache *thread_cache;
97 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
98 static void lwkt_setcpu_remote(void *arg);
101 extern void cpu_heavy_restore(void);
102 extern void cpu_lwkt_restore(void);
103 extern void cpu_kthread_restore(void);
104 extern void cpu_idle_restore(void);
107 * We can make all thread ports use the spin backend instead of the thread
108 * backend. This should only be set to debug the spin backend.
110 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
113 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
114 "Panic if attempting to switch lwkt's while mastering cpusync");
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
186 * become runnable. We do a secondary comparison using td_upri to try
187 * to order these in the situation where several wake up at the same time
188 * to avoid excessive switching.
190 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
191 * will ignore user priority. This is to ensure that user threads in
192 * kernel mode get cpu at some point regardless of what the user
197 _lwkt_enqueue(thread_t td)
201 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
202 struct globaldata *gd = td->td_gd;
204 td->td_flags |= TDF_RUNQ;
205 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
211 (xtd->td_pri > td->td_pri ||
212 (xtd->td_pri == td->td_pri &&
213 xtd->td_upri >= td->td_pri))) {
214 xtd = TAILQ_NEXT(xtd, td_threadq);
217 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
219 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
221 ++gd->gd_tdrunqcount;
224 * Request a LWKT reschedule if we are now at the head of the queue.
226 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
232 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
234 struct thread *td = (struct thread *)obj;
236 td->td_kstack = NULL;
237 td->td_kstack_size = 0;
238 td->td_flags = TDF_ALLOCATED_THREAD;
244 _lwkt_thread_dtor(void *obj, void *privdata)
246 struct thread *td = (struct thread *)obj;
248 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
249 ("_lwkt_thread_dtor: not allocated from objcache"));
250 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
251 td->td_kstack_size > 0,
252 ("_lwkt_thread_dtor: corrupted stack"));
253 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
254 td->td_kstack = NULL;
259 * Initialize the lwkt s/system.
261 * Nominally cache up to 32 thread + kstack structures. Cache more on
262 * systems with a lot of cpu cores.
267 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
268 if (lwkt_cache_threads == 0) {
269 lwkt_cache_threads = ncpus * 4;
270 if (lwkt_cache_threads < 32)
271 lwkt_cache_threads = 32;
273 thread_cache = objcache_create_mbacked(
274 M_THREAD, sizeof(struct thread),
275 NULL, lwkt_cache_threads,
276 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
280 * Schedule a thread to run. As the current thread we can always safely
281 * schedule ourselves, and a shortcut procedure is provided for that
284 * (non-blocking, self contained on a per cpu basis)
287 lwkt_schedule_self(thread_t td)
289 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
290 crit_enter_quick(td);
291 KASSERT(td != &td->td_gd->gd_idlethread,
292 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
293 KKASSERT(td->td_lwp == NULL ||
294 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
300 * Deschedule a thread.
302 * (non-blocking, self contained on a per cpu basis)
305 lwkt_deschedule_self(thread_t td)
307 crit_enter_quick(td);
313 * LWKTs operate on a per-cpu basis
315 * WARNING! Called from early boot, 'mycpu' may not work yet.
318 lwkt_gdinit(struct globaldata *gd)
320 TAILQ_INIT(&gd->gd_tdrunq);
321 TAILQ_INIT(&gd->gd_tdallq);
325 * Create a new thread. The thread must be associated with a process context
326 * or LWKT start address before it can be scheduled. If the target cpu is
327 * -1 the thread will be created on the current cpu.
329 * If you intend to create a thread without a process context this function
330 * does everything except load the startup and switcher function.
333 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
335 static int cpu_rotator;
336 globaldata_t gd = mycpu;
340 * If static thread storage is not supplied allocate a thread. Reuse
341 * a cached free thread if possible. gd_freetd is used to keep an exiting
342 * thread intact through the exit.
346 if ((td = gd->gd_freetd) != NULL) {
347 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
349 gd->gd_freetd = NULL;
351 td = objcache_get(thread_cache, M_WAITOK);
352 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
356 KASSERT((td->td_flags &
357 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
358 TDF_ALLOCATED_THREAD,
359 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
360 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
364 * Try to reuse cached stack.
366 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
367 if (flags & TDF_ALLOCATED_STACK) {
368 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
373 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
374 flags |= TDF_ALLOCATED_STACK;
381 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
386 * Initialize a preexisting thread structure. This function is used by
387 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
389 * All threads start out in a critical section at a priority of
390 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
391 * appropriate. This function may send an IPI message when the
392 * requested cpu is not the current cpu and consequently gd_tdallq may
393 * not be initialized synchronously from the point of view of the originating
396 * NOTE! we have to be careful in regards to creating threads for other cpus
397 * if SMP has not yet been activated.
402 lwkt_init_thread_remote(void *arg)
407 * Protected by critical section held by IPI dispatch
409 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
415 * lwkt core thread structural initialization.
417 * NOTE: All threads are initialized as mpsafe threads.
420 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
421 struct globaldata *gd)
423 globaldata_t mygd = mycpu;
425 bzero(td, sizeof(struct thread));
426 td->td_kstack = stack;
427 td->td_kstack_size = stksize;
428 td->td_flags = flags;
431 td->td_pri = TDPRI_KERN_DAEMON;
432 td->td_critcount = 1;
433 td->td_toks_have = NULL;
434 td->td_toks_stop = &td->td_toks_base;
435 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
436 lwkt_initport_spin(&td->td_msgport, td);
438 lwkt_initport_thread(&td->td_msgport, td);
439 pmap_init_thread(td);
442 * Normally initializing a thread for a remote cpu requires sending an
443 * IPI. However, the idlethread is setup before the other cpus are
444 * activated so we have to treat it as a special case. XXX manipulation
445 * of gd_tdallq requires the BGL.
447 if (gd == mygd || td == &gd->gd_idlethread) {
449 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
452 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
456 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
460 dsched_new_thread(td);
464 lwkt_set_comm(thread_t td, const char *ctl, ...)
469 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
471 KTR_LOG(ctxsw_newtd, td, td->td_comm);
475 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
476 * this does not prevent the thread from migrating to another cpu so the
477 * gd_tdallq state is not protected by this.
480 lwkt_hold(thread_t td)
482 atomic_add_int(&td->td_refs, 1);
486 lwkt_rele(thread_t td)
488 KKASSERT(td->td_refs > 0);
489 atomic_add_int(&td->td_refs, -1);
493 lwkt_free_thread(thread_t td)
495 KKASSERT(td->td_refs == 0);
496 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
497 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
498 if (td->td_flags & TDF_ALLOCATED_THREAD) {
499 objcache_put(thread_cache, td);
500 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
501 /* client-allocated struct with internally allocated stack */
502 KASSERT(td->td_kstack && td->td_kstack_size > 0,
503 ("lwkt_free_thread: corrupted stack"));
504 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
505 td->td_kstack = NULL;
506 td->td_kstack_size = 0;
508 KTR_LOG(ctxsw_deadtd, td);
513 * Switch to the next runnable lwkt. If no LWKTs are runnable then
514 * switch to the idlethread. Switching must occur within a critical
515 * section to avoid races with the scheduling queue.
517 * We always have full control over our cpu's run queue. Other cpus
518 * that wish to manipulate our queue must use the cpu_*msg() calls to
519 * talk to our cpu, so a critical section is all that is needed and
520 * the result is very, very fast thread switching.
522 * The LWKT scheduler uses a fixed priority model and round-robins at
523 * each priority level. User process scheduling is a totally
524 * different beast and LWKT priorities should not be confused with
525 * user process priorities.
527 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
528 * is not called by the current thread in the preemption case, only when
529 * the preempting thread blocks (in order to return to the original thread).
531 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
532 * migration and tsleep deschedule the current lwkt thread and call
533 * lwkt_switch(). In particular, the target cpu of the migration fully
534 * expects the thread to become non-runnable and can deadlock against
535 * cpusync operations if we run any IPIs prior to switching the thread out.
537 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
538 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
543 globaldata_t gd = mycpu;
544 thread_t td = gd->gd_curthread;
548 KKASSERT(gd->gd_processing_ipiq == 0);
549 KKASSERT(td->td_flags & TDF_RUNNING);
552 * Switching from within a 'fast' (non thread switched) interrupt or IPI
553 * is illegal. However, we may have to do it anyway if we hit a fatal
554 * kernel trap or we have paniced.
556 * If this case occurs save and restore the interrupt nesting level.
558 if (gd->gd_intr_nesting_level) {
562 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
563 panic("lwkt_switch: Attempt to switch from a "
564 "fast interrupt, ipi, or hard code section, "
568 savegdnest = gd->gd_intr_nesting_level;
569 savegdtrap = gd->gd_trap_nesting_level;
570 gd->gd_intr_nesting_level = 0;
571 gd->gd_trap_nesting_level = 0;
572 if ((td->td_flags & TDF_PANICWARN) == 0) {
573 td->td_flags |= TDF_PANICWARN;
574 kprintf("Warning: thread switch from interrupt, IPI, "
575 "or hard code section.\n"
576 "thread %p (%s)\n", td, td->td_comm);
580 gd->gd_intr_nesting_level = savegdnest;
581 gd->gd_trap_nesting_level = savegdtrap;
587 * Release our current user process designation if we are blocking
588 * or if a user reschedule was requested.
590 * NOTE: This function is NOT called if we are switching into or
591 * returning from a preemption.
593 * NOTE: Releasing our current user process designation may cause
594 * it to be assigned to another thread, which in turn will
595 * cause us to block in the usched acquire code when we attempt
596 * to return to userland.
598 * NOTE: On SMP systems this can be very nasty when heavy token
599 * contention is present so we want to be careful not to
600 * release the designation gratuitously.
602 if (td->td_release &&
603 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
611 if (TD_TOKS_HELD(td))
612 lwkt_relalltokens(td);
615 * We had better not be holding any spin locks, but don't get into an
616 * endless panic loop.
618 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
619 ("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");
635 * If we had preempted another thread on this cpu, resume the preempted
636 * thread. This occurs transparently, whether the preempted thread
637 * was scheduled or not (it may have been preempted after descheduling
640 * We have to setup the MP lock for the original thread after backing
641 * out the adjustment that was made to curthread when the original
644 if ((ntd = td->td_preempted) != NULL) {
645 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
646 ntd->td_flags |= TDF_PREEMPT_DONE;
649 * The interrupt may have woken a thread up, we need to properly
650 * set the reschedule flag if the originally interrupted thread is
651 * at a lower priority.
653 * The interrupt may not have descheduled.
655 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
657 goto havethread_preempted;
661 * If we cannot obtain ownership of the tokens we cannot immediately
662 * schedule the target thread.
664 * Reminder: Again, we cannot afford to run any IPIs in this path if
665 * the current thread has been descheduled.
668 clear_lwkt_resched();
671 * Hotpath - pull the head of the run queue and attempt to schedule
674 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
678 * Runq is empty, switch to idle to allow it to halt.
680 ntd = &gd->gd_idlethread;
682 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
683 ASSERT_NO_TOKENS_HELD(ntd);
685 cpu_time.cp_msg[0] = 0;
686 cpu_time.cp_stallpc = 0;
691 * Hotpath - schedule ntd.
693 * NOTE: For UP there is no mplock and lwkt_getalltokens()
696 if (TD_TOKS_NOT_HELD(ntd) ||
697 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
703 * Coldpath (SMP only since tokens always succeed on UP)
705 * We had some contention on the thread we wanted to schedule.
706 * What we do now is try to find a thread that we can schedule
709 * The coldpath scan does NOT rearrange threads in the run list.
710 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
711 * the next tick whenever the current head is not the current thread.
716 ++gd->gd_cnt.v_token_colls;
718 if (fairq_bypass > 0)
721 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
722 #ifdef LWKT_SPLIT_USERPRI
724 * Never schedule threads returning to userland or the
725 * user thread scheduler helper thread when higher priority
726 * threads are present. The runq is sorted by priority
727 * so we can give up traversing it when we find the first
728 * low priority thread.
730 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
739 if (TD_TOKS_NOT_HELD(ntd) ||
740 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
746 ++gd->gd_cnt.v_token_colls;
751 * We exhausted the run list, meaning that all runnable threads
755 ntd = &gd->gd_idlethread;
757 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
758 ASSERT_NO_TOKENS_HELD(ntd);
759 /* contention case, do not clear contention mask */
763 * We are going to have to retry but if the current thread is not
764 * on the runq we instead switch through the idle thread to get away
765 * from the current thread. We have to flag for lwkt reschedule
766 * to prevent the idle thread from halting.
768 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
769 * instruct it to deal with the potential for deadlocks by
770 * ordering the tokens by address.
772 if ((td->td_flags & TDF_RUNQ) == 0) {
773 need_lwkt_resched(); /* prevent hlt */
776 #if defined(INVARIANTS) && defined(__amd64__)
777 if ((read_rflags() & PSL_I) == 0) {
779 panic("lwkt_switch() called with interrupts disabled");
784 * Number iterations so far. After a certain point we switch to
785 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
787 if (spinning < 0x7FFFFFFF)
792 * lwkt_getalltokens() failed in sorted token mode, we can use
793 * monitor/mwait in this case.
795 if (spinning >= lwkt_spin_loops &&
796 (cpu_mi_feature & CPU_MI_MONITOR) &&
799 cpu_mmw_pause_int(&gd->gd_reqflags,
800 (gd->gd_reqflags | RQF_SPINNING) &
801 ~RQF_IDLECHECK_WK_MASK);
806 * We already checked that td is still scheduled so this should be
812 * This experimental resequencer is used as a fall-back to reduce
813 * hw cache line contention by placing each core's scheduler into a
814 * time-domain-multplexed slot.
816 * The resequencer is disabled by default. It's functionality has
817 * largely been superceeded by the token algorithm which limits races
818 * to a subset of cores.
820 * The resequencer algorithm tends to break down when more than
821 * 20 cores are contending. What appears to happen is that new
822 * tokens can be obtained out of address-sorted order by new cores
823 * while existing cores languish in long delays between retries and
824 * wind up being starved-out of the token acquisition.
826 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
827 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
830 while ((oseq = lwkt_cseq_rindex) != cseq) {
833 if (cpu_mi_feature & CPU_MI_MONITOR) {
834 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
844 atomic_add_int(&lwkt_cseq_rindex, 1);
846 /* highest level for(;;) loop */
851 * Clear gd_idle_repeat when doing a normal switch to a non-idle
854 ntd->td_wmesg = NULL;
855 ++gd->gd_cnt.v_swtch;
856 gd->gd_idle_repeat = 0;
858 havethread_preempted:
860 * If the new target does not need the MP lock and we are holding it,
861 * release the MP lock. If the new target requires the MP lock we have
862 * already acquired it for the target.
866 KASSERT(ntd->td_critcount,
867 ("priority problem in lwkt_switch %d %d",
868 td->td_critcount, ntd->td_critcount));
872 * Execute the actual thread switch operation. This function
873 * returns to the current thread and returns the previous thread
874 * (which may be different from the thread we switched to).
876 * We are responsible for marking ntd as TDF_RUNNING.
878 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
880 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
881 ntd->td_flags |= TDF_RUNNING;
882 lwkt_switch_return(td->td_switch(ntd));
883 /* ntd invalid, td_switch() can return a different thread_t */
887 * catch-all. XXX is this strictly needed?
891 /* NOTE: current cpu may have changed after switch */
896 * Called by assembly in the td_switch (thread restore path) for thread
897 * bootstrap cases which do not 'return' to lwkt_switch().
900 lwkt_switch_return(thread_t otd)
906 * Check if otd was migrating. Now that we are on ntd we can finish
907 * up the migration. This is a bit messy but it is the only place
908 * where td is known to be fully descheduled.
910 * We can only activate the migration if otd was migrating but not
911 * held on the cpu due to a preemption chain. We still have to
912 * clear TDF_RUNNING on the old thread either way.
914 * We are responsible for clearing the previously running thread's
917 if ((rgd = otd->td_migrate_gd) != NULL &&
918 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
919 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
920 (TDF_MIGRATING | TDF_RUNNING));
921 otd->td_migrate_gd = NULL;
922 otd->td_flags &= ~TDF_RUNNING;
923 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
925 otd->td_flags &= ~TDF_RUNNING;
928 otd->td_flags &= ~TDF_RUNNING;
932 * Final exit validations (see lwp_wait()). Note that otd becomes
933 * invalid the *instant* we set TDF_MP_EXITSIG.
935 while (otd->td_flags & TDF_EXITING) {
938 mpflags = otd->td_mpflags;
941 if (mpflags & TDF_MP_EXITWAIT) {
942 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
943 mpflags | TDF_MP_EXITSIG)) {
948 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
949 mpflags | TDF_MP_EXITSIG)) {
958 * Request that the target thread preempt the current thread. Preemption
959 * can only occur if our only critical section is the one that we were called
960 * with, the relative priority of the target thread is higher, and the target
961 * thread holds no tokens. This also only works if we are not holding any
962 * spinlocks (obviously).
964 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
965 * this is called via lwkt_schedule() through the td_preemptable callback.
966 * critcount is the managed critical priority that we should ignore in order
967 * to determine whether preemption is possible (aka usually just the crit
968 * priority of lwkt_schedule() itself).
970 * Preemption is typically limited to interrupt threads.
972 * Operation works in a fairly straight-forward manner. The normal
973 * scheduling code is bypassed and we switch directly to the target
974 * thread. When the target thread attempts to block or switch away
975 * code at the base of lwkt_switch() will switch directly back to our
976 * thread. Our thread is able to retain whatever tokens it holds and
977 * if the target needs one of them the target will switch back to us
978 * and reschedule itself normally.
981 lwkt_preempt(thread_t ntd, int critcount)
983 struct globaldata *gd = mycpu;
986 int save_gd_intr_nesting_level;
989 * The caller has put us in a critical section. We can only preempt
990 * if the caller of the caller was not in a critical section (basically
991 * a local interrupt), as determined by the 'critcount' parameter. We
992 * also can't preempt if the caller is holding any spinlocks (even if
993 * he isn't in a critical section). This also handles the tokens test.
995 * YYY The target thread must be in a critical section (else it must
996 * inherit our critical section? I dunno yet).
998 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1000 td = gd->gd_curthread;
1001 if (preempt_enable == 0) {
1005 if (ntd->td_pri <= td->td_pri) {
1009 if (td->td_critcount > critcount) {
1014 if (td->td_cscount) {
1018 if (ntd->td_gd != gd) {
1024 * We don't have to check spinlocks here as they will also bump
1027 * Do not try to preempt if the target thread is holding any tokens.
1028 * We could try to acquire the tokens but this case is so rare there
1029 * is no need to support it.
1031 KKASSERT(gd->gd_spinlocks == 0);
1033 if (TD_TOKS_HELD(ntd)) {
1037 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1041 if (ntd->td_preempted) {
1045 KKASSERT(gd->gd_processing_ipiq == 0);
1048 * Since we are able to preempt the current thread, there is no need to
1049 * call need_lwkt_resched().
1051 * We must temporarily clear gd_intr_nesting_level around the switch
1052 * since switchouts from the target thread are allowed (they will just
1053 * return to our thread), and since the target thread has its own stack.
1055 * A preemption must switch back to the original thread, assert the
1059 ntd->td_preempted = td;
1060 td->td_flags |= TDF_PREEMPT_LOCK;
1061 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1062 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1063 gd->gd_intr_nesting_level = 0;
1065 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1066 ntd->td_flags |= TDF_RUNNING;
1067 xtd = td->td_switch(ntd);
1068 KKASSERT(xtd == ntd);
1069 lwkt_switch_return(xtd);
1070 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1072 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1073 ntd->td_preempted = NULL;
1074 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1078 * Conditionally call splz() if gd_reqflags indicates work is pending.
1079 * This will work inside a critical section but not inside a hard code
1082 * (self contained on a per cpu basis)
1087 globaldata_t gd = mycpu;
1088 thread_t td = gd->gd_curthread;
1090 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1091 gd->gd_intr_nesting_level == 0 &&
1092 td->td_nest_count < 2)
1099 * This version is integrated into crit_exit, reqflags has already
1100 * been tested but td_critcount has not.
1102 * We only want to execute the splz() on the 1->0 transition of
1103 * critcount and not in a hard code section or if too deeply nested.
1105 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
1108 lwkt_maybe_splz(thread_t td)
1110 globaldata_t gd = td->td_gd;
1112 if (td->td_critcount == 0 &&
1113 gd->gd_intr_nesting_level == 0 &&
1114 td->td_nest_count < 2)
1121 * Drivers which set up processing co-threads can call this function to
1122 * run the co-thread at a higher priority and to allow it to preempt
1126 lwkt_set_interrupt_support_thread(void)
1128 thread_t td = curthread;
1130 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1131 td->td_flags |= TDF_INTTHREAD;
1132 td->td_preemptable = lwkt_preempt;
1137 * This function is used to negotiate a passive release of the current
1138 * process/lwp designation with the user scheduler, allowing the user
1139 * scheduler to schedule another user thread. The related kernel thread
1140 * (curthread) continues running in the released state.
1143 lwkt_passive_release(struct thread *td)
1145 struct lwp *lp = td->td_lwp;
1147 #ifdef LWKT_SPLIT_USERPRI
1148 td->td_release = NULL;
1149 lwkt_setpri_self(TDPRI_KERN_USER);
1152 lp->lwp_proc->p_usched->release_curproc(lp);
1157 * This implements a LWKT yield, allowing a kernel thread to yield to other
1158 * kernel threads at the same or higher priority. This function can be
1159 * called in a tight loop and will typically only yield once per tick.
1161 * Most kernel threads run at the same priority in order to allow equal
1164 * (self contained on a per cpu basis)
1169 globaldata_t gd = mycpu;
1170 thread_t td = gd->gd_curthread;
1172 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1174 if (lwkt_resched_wanted()) {
1175 lwkt_schedule_self(curthread);
1181 * The quick version processes pending interrupts and higher-priority
1182 * LWKT threads but will not round-robin same-priority LWKT threads.
1184 * When called while attempting to return to userland the only same-pri
1185 * threads are the ones which have already tried to become the current
1189 lwkt_yield_quick(void)
1191 globaldata_t gd = mycpu;
1192 thread_t td = gd->gd_curthread;
1194 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1196 if (lwkt_resched_wanted()) {
1197 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1198 clear_lwkt_resched();
1200 lwkt_schedule_self(curthread);
1207 * This yield is designed for kernel threads with a user context.
1209 * The kernel acting on behalf of the user is potentially cpu-bound,
1210 * this function will efficiently allow other threads to run and also
1211 * switch to other processes by releasing.
1213 * The lwkt_user_yield() function is designed to have very low overhead
1214 * if no yield is determined to be needed.
1217 lwkt_user_yield(void)
1219 globaldata_t gd = mycpu;
1220 thread_t td = gd->gd_curthread;
1223 * Always run any pending interrupts in case we are in a critical
1226 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1230 * Switch (which forces a release) if another kernel thread needs
1231 * the cpu, if userland wants us to resched, or if our kernel
1232 * quantum has run out.
1234 if (lwkt_resched_wanted() ||
1235 user_resched_wanted())
1242 * Reacquire the current process if we are released.
1244 * XXX not implemented atm. The kernel may be holding locks and such,
1245 * so we want the thread to continue to receive cpu.
1247 if (td->td_release == NULL && lp) {
1248 lp->lwp_proc->p_usched->acquire_curproc(lp);
1249 td->td_release = lwkt_passive_release;
1250 lwkt_setpri_self(TDPRI_USER_NORM);
1256 * Generic schedule. Possibly schedule threads belonging to other cpus and
1257 * deal with threads that might be blocked on a wait queue.
1259 * We have a little helper inline function which does additional work after
1260 * the thread has been enqueued, including dealing with preemption and
1261 * setting need_lwkt_resched() (which prevents the kernel from returning
1262 * to userland until it has processed higher priority threads).
1264 * It is possible for this routine to be called after a failed _enqueue
1265 * (due to the target thread migrating, sleeping, or otherwise blocked).
1266 * We have to check that the thread is actually on the run queue!
1270 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1272 if (ntd->td_flags & TDF_RUNQ) {
1273 if (ntd->td_preemptable) {
1274 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1281 _lwkt_schedule(thread_t td)
1283 globaldata_t mygd = mycpu;
1285 KASSERT(td != &td->td_gd->gd_idlethread,
1286 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1287 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1288 crit_enter_gd(mygd);
1289 KKASSERT(td->td_lwp == NULL ||
1290 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1292 if (td == mygd->gd_curthread) {
1296 * If we own the thread, there is no race (since we are in a
1297 * critical section). If we do not own the thread there might
1298 * be a race but the target cpu will deal with it.
1301 if (td->td_gd == mygd) {
1303 _lwkt_schedule_post(mygd, td, 1);
1305 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1309 _lwkt_schedule_post(mygd, td, 1);
1316 lwkt_schedule(thread_t td)
1322 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1330 * When scheduled remotely if frame != NULL the IPIQ is being
1331 * run via doreti or an interrupt then preemption can be allowed.
1333 * To allow preemption we have to drop the critical section so only
1334 * one is present in _lwkt_schedule_post.
1337 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1339 thread_t td = curthread;
1342 if (frame && ntd->td_preemptable) {
1343 crit_exit_noyield(td);
1344 _lwkt_schedule(ntd);
1345 crit_enter_quick(td);
1347 _lwkt_schedule(ntd);
1352 * Thread migration using a 'Pull' method. The thread may or may not be
1353 * the current thread. It MUST be descheduled and in a stable state.
1354 * lwkt_giveaway() must be called on the cpu owning the thread.
1356 * At any point after lwkt_giveaway() is called, the target cpu may
1357 * 'pull' the thread by calling lwkt_acquire().
1359 * We have to make sure the thread is not sitting on a per-cpu tsleep
1360 * queue or it will blow up when it moves to another cpu.
1362 * MPSAFE - must be called under very specific conditions.
1365 lwkt_giveaway(thread_t td)
1367 globaldata_t gd = mycpu;
1370 if (td->td_flags & TDF_TSLEEPQ)
1372 KKASSERT(td->td_gd == gd);
1373 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1374 td->td_flags |= TDF_MIGRATING;
1379 lwkt_acquire(thread_t td)
1383 int retry = 10000000;
1385 KKASSERT(td->td_flags & TDF_MIGRATING);
1390 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1391 crit_enter_gd(mygd);
1392 DEBUG_PUSH_INFO("lwkt_acquire");
1393 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1395 lwkt_process_ipiq();
1399 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1407 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1408 td->td_flags &= ~TDF_MIGRATING;
1411 crit_enter_gd(mygd);
1412 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1413 td->td_flags &= ~TDF_MIGRATING;
1421 * Generic deschedule. Descheduling threads other then your own should be
1422 * done only in carefully controlled circumstances. Descheduling is
1425 * This function may block if the cpu has run out of messages.
1428 lwkt_deschedule(thread_t td)
1432 if (td == curthread) {
1435 if (td->td_gd == mycpu) {
1438 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1448 * Set the target thread's priority. This routine does not automatically
1449 * switch to a higher priority thread, LWKT threads are not designed for
1450 * continuous priority changes. Yield if you want to switch.
1453 lwkt_setpri(thread_t td, int pri)
1455 if (td->td_pri != pri) {
1458 if (td->td_flags & TDF_RUNQ) {
1459 KKASSERT(td->td_gd == mycpu);
1471 * Set the initial priority for a thread prior to it being scheduled for
1472 * the first time. The thread MUST NOT be scheduled before or during
1473 * this call. The thread may be assigned to a cpu other then the current
1476 * Typically used after a thread has been created with TDF_STOPPREQ,
1477 * and before the thread is initially scheduled.
1480 lwkt_setpri_initial(thread_t td, int pri)
1483 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1488 lwkt_setpri_self(int pri)
1490 thread_t td = curthread;
1492 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1494 if (td->td_flags & TDF_RUNQ) {
1505 * hz tick scheduler clock for LWKT threads
1508 lwkt_schedulerclock(thread_t td)
1510 globaldata_t gd = td->td_gd;
1513 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1515 * If the current thread is at the head of the runq shift it to the
1516 * end of any equal-priority threads and request a LWKT reschedule
1519 * Ignore upri in this situation. There will only be one user thread
1520 * in user mode, all others will be user threads running in kernel
1521 * mode and we have to make sure they get some cpu.
1523 xtd = TAILQ_NEXT(td, td_threadq);
1524 if (xtd && xtd->td_pri == td->td_pri) {
1525 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1526 while (xtd && xtd->td_pri == td->td_pri)
1527 xtd = TAILQ_NEXT(xtd, td_threadq);
1529 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1531 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1532 need_lwkt_resched();
1536 * If we scheduled a thread other than the one at the head of the
1537 * queue always request a reschedule every tick.
1539 need_lwkt_resched();
1544 * Migrate the current thread to the specified cpu.
1546 * This is accomplished by descheduling ourselves from the current cpu
1547 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1548 * 'old' thread wants to migrate after it has been completely switched out
1549 * and will complete the migration.
1551 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1553 * We must be sure to release our current process designation (if a user
1554 * process) before clearing out any tsleepq we are on because the release
1555 * code may re-add us.
1557 * We must be sure to remove ourselves from the current cpu's tsleepq
1558 * before potentially moving to another queue. The thread can be on
1559 * a tsleepq due to a left-over tsleep_interlock().
1563 lwkt_setcpu_self(globaldata_t rgd)
1566 thread_t td = curthread;
1568 if (td->td_gd != rgd) {
1569 crit_enter_quick(td);
1573 if (td->td_flags & TDF_TSLEEPQ)
1577 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1578 * trying to deschedule ourselves and switch away, then deschedule
1579 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1580 * call lwkt_switch() to complete the operation.
1582 td->td_flags |= TDF_MIGRATING;
1583 lwkt_deschedule_self(td);
1584 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1585 td->td_migrate_gd = rgd;
1589 * We are now on the target cpu
1591 KKASSERT(rgd == mycpu);
1592 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1593 crit_exit_quick(td);
1599 lwkt_migratecpu(int cpuid)
1604 rgd = globaldata_find(cpuid);
1605 lwkt_setcpu_self(rgd);
1611 * Remote IPI for cpu migration (called while in a critical section so we
1612 * do not have to enter another one).
1614 * The thread (td) has already been completely descheduled from the
1615 * originating cpu and we can simply assert the case. The thread is
1616 * assigned to the new cpu and enqueued.
1618 * The thread will re-add itself to tdallq when it resumes execution.
1621 lwkt_setcpu_remote(void *arg)
1624 globaldata_t gd = mycpu;
1626 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1629 td->td_flags &= ~TDF_MIGRATING;
1630 KKASSERT(td->td_migrate_gd == NULL);
1631 KKASSERT(td->td_lwp == NULL ||
1632 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1638 lwkt_preempted_proc(void)
1640 thread_t td = curthread;
1641 while (td->td_preempted)
1642 td = td->td_preempted;
1647 * Create a kernel process/thread/whatever. It shares it's address space
1648 * with proc0 - ie: kernel only.
1650 * If the cpu is not specified one will be selected. In the future
1651 * specifying a cpu of -1 will enable kernel thread migration between
1655 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1656 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1661 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1665 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1668 * Set up arg0 for 'ps' etc
1670 __va_start(ap, fmt);
1671 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1675 * Schedule the thread to run
1677 if (td->td_flags & TDF_NOSTART)
1678 td->td_flags &= ~TDF_NOSTART;
1685 * Destroy an LWKT thread. Warning! This function is not called when
1686 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1687 * uses a different reaping mechanism.
1692 thread_t td = curthread;
1697 * Do any cleanup that might block here
1699 if (td->td_flags & TDF_VERBOSE)
1700 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1703 dsched_exit_thread(td);
1706 * Get us into a critical section to interlock gd_freetd and loop
1707 * until we can get it freed.
1709 * We have to cache the current td in gd_freetd because objcache_put()ing
1710 * it would rip it out from under us while our thread is still active.
1712 * We are the current thread so of course our own TDF_RUNNING bit will
1713 * be set, so unlike the lwp reap code we don't wait for it to clear.
1716 crit_enter_quick(td);
1719 tsleep(td, 0, "tdreap", 1);
1722 if ((std = gd->gd_freetd) != NULL) {
1723 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1724 gd->gd_freetd = NULL;
1725 objcache_put(thread_cache, std);
1732 * Remove thread resources from kernel lists and deschedule us for
1733 * the last time. We cannot block after this point or we may end
1734 * up with a stale td on the tsleepq.
1736 * None of this may block, the critical section is the only thing
1737 * protecting tdallq and the only thing preventing new lwkt_hold()
1740 if (td->td_flags & TDF_TSLEEPQ)
1742 lwkt_deschedule_self(td);
1743 lwkt_remove_tdallq(td);
1744 KKASSERT(td->td_refs == 0);
1749 KKASSERT(gd->gd_freetd == NULL);
1750 if (td->td_flags & TDF_ALLOCATED_THREAD)
1756 lwkt_remove_tdallq(thread_t td)
1758 KKASSERT(td->td_gd == mycpu);
1759 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1763 * Code reduction and branch prediction improvements. Call/return
1764 * overhead on modern cpus often degenerates into 0 cycles due to
1765 * the cpu's branch prediction hardware and return pc cache. We
1766 * can take advantage of this by not inlining medium-complexity
1767 * functions and we can also reduce the branch prediction impact
1768 * by collapsing perfectly predictable branches into a single
1769 * procedure instead of duplicating it.
1771 * Is any of this noticeable? Probably not, so I'll take the
1772 * smaller code size.
1775 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1777 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1783 thread_t td = curthread;
1784 int lcrit = td->td_critcount;
1786 td->td_critcount = 0;
1787 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1794 * Called from debugger/panic on cpus which have been stopped. We must still
1795 * process the IPIQ while stopped, even if we were stopped while in a critical
1798 * If we are dumping also try to process any pending interrupts. This may
1799 * or may not work depending on the state of the cpu at the point it was
1803 lwkt_smp_stopped(void)
1805 globaldata_t gd = mycpu;
1809 lwkt_process_ipiq();
1812 lwkt_process_ipiq();