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>
55 #include <sys/indefinite.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/indefinite2.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>
74 #include <machine/clock.h>
76 #ifdef _KERNEL_VIRTUAL
82 #if !defined(KTR_CTXSW)
83 #define KTR_CTXSW KTR_ALL
85 KTR_INFO_MASTER(ctxsw);
86 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
87 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
88 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
89 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
91 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
94 static int panic_on_cscount = 0;
96 #ifdef DEBUG_LWKT_THREAD
97 static int64_t switch_count = 0;
98 static int64_t preempt_hit = 0;
99 static int64_t preempt_miss = 0;
100 static int64_t preempt_weird = 0;
102 static int lwkt_use_spin_port;
103 __read_mostly static struct objcache *thread_cache;
104 int cpu_mwait_spin = 0;
106 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
107 static void lwkt_setcpu_remote(void *arg);
110 * We can make all thread ports use the spin backend instead of the thread
111 * backend. This should only be set to debug the spin backend.
113 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
116 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
117 "Panic if attempting to switch lwkt's while mastering cpusync");
119 #ifdef DEBUG_LWKT_THREAD
120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
123 "Successful preemption events");
124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
125 "Failed preemption events");
126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
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 __read_mostly static u_int lwkt_spin_loops = 10;
134 SYSCTL_UINT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
135 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
136 __read_mostly static int preempt_enable = 1;
137 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
138 &preempt_enable, 0, "Enable preemption");
139 static int lwkt_cache_threads = 0;
140 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
141 &lwkt_cache_threads, 0, "thread+kstack cache");
144 * These helper procedures handle the runq, they can only be called from
145 * within a critical section.
147 * WARNING! Prior to SMP being brought up it is possible to enqueue and
148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
149 * instead of 'mycpu' when referencing the globaldata structure. Once
150 * SMP live enqueuing and dequeueing only occurs on the current cpu.
154 _lwkt_dequeue(thread_t td)
156 if (td->td_flags & TDF_RUNQ) {
157 struct globaldata *gd = td->td_gd;
159 td->td_flags &= ~TDF_RUNQ;
160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
161 --gd->gd_tdrunqcount;
162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
163 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
170 * There are a limited number of lwkt threads runnable since user
171 * processes only schedule one at a time per cpu. However, there can
172 * be many user processes in kernel mode exiting from a tsleep() which
175 * We scan the queue in both directions to help deal with degenerate
176 * situations when hundreds or thousands (or more) threads are runnable.
178 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
179 * will ignore user priority. This is to ensure that user threads in
180 * kernel mode get cpu at some point regardless of what the user
185 _lwkt_enqueue(thread_t td)
187 thread_t xtd; /* forward scan */
188 thread_t rtd; /* reverse scan */
190 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
191 struct globaldata *gd = td->td_gd;
193 td->td_flags |= TDF_RUNQ;
194 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
196 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
197 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
200 * NOTE: td_upri - higher numbers more desireable, same sense
201 * as td_pri (typically reversed from lwp_upri).
203 * In the equal priority case we want the best selection
204 * at the beginning so the less desireable selections know
205 * that they have to setrunqueue/go-to-another-cpu, even
206 * though it means switching back to the 'best' selection.
207 * This also avoids degenerate situations when many threads
208 * are runnable or waking up at the same time.
210 * If upri matches exactly place at end/round-robin.
212 rtd = TAILQ_LAST(&gd->gd_tdrunq, lwkt_queue);
215 (xtd->td_pri > td->td_pri ||
216 (xtd->td_pri == td->td_pri &&
217 xtd->td_upri >= td->td_upri))) {
218 xtd = TAILQ_NEXT(xtd, td_threadq);
221 * Doing a reverse scan at the same time is an optimization
222 * for the insert-closer-to-tail case that avoids having to
223 * scan the entire list. This situation can occur when
224 * thousands of threads are woken up at the same time.
226 if (rtd->td_pri > td->td_pri ||
227 (rtd->td_pri == td->td_pri &&
228 rtd->td_upri >= td->td_upri)) {
229 TAILQ_INSERT_AFTER(&gd->gd_tdrunq, rtd, td, td_threadq);
232 rtd = TAILQ_PREV(rtd, lwkt_queue, td_threadq);
235 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
237 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
240 ++gd->gd_tdrunqcount;
243 * Request a LWKT reschedule if we are now at the head of the queue.
245 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
251 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
253 struct thread *td = (struct thread *)obj;
255 td->td_kstack = NULL;
256 td->td_kstack_size = 0;
257 td->td_flags = TDF_ALLOCATED_THREAD;
263 _lwkt_thread_dtor(void *obj, void *privdata)
265 struct thread *td = (struct thread *)obj;
267 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
268 ("_lwkt_thread_dtor: not allocated from objcache"));
269 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
270 td->td_kstack_size > 0,
271 ("_lwkt_thread_dtor: corrupted stack"));
272 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
273 td->td_kstack = NULL;
278 * Initialize the lwkt s/system.
280 * Nominally cache up to 32 thread + kstack structures. Cache more on
281 * systems with a lot of cpu cores.
286 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
287 if (lwkt_cache_threads == 0) {
288 lwkt_cache_threads = ncpus * 4;
289 if (lwkt_cache_threads < 32)
290 lwkt_cache_threads = 32;
292 thread_cache = objcache_create_mbacked(
293 M_THREAD, sizeof(struct thread),
294 0, lwkt_cache_threads,
295 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
297 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
300 * Schedule a thread to run. As the current thread we can always safely
301 * schedule ourselves, and a shortcut procedure is provided for that
304 * (non-blocking, self contained on a per cpu basis)
307 lwkt_schedule_self(thread_t td)
309 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
310 crit_enter_quick(td);
311 KASSERT(td != &td->td_gd->gd_idlethread,
312 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
313 KKASSERT(td->td_lwp == NULL ||
314 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
320 * Deschedule a thread.
322 * (non-blocking, self contained on a per cpu basis)
325 lwkt_deschedule_self(thread_t td)
327 crit_enter_quick(td);
333 * LWKTs operate on a per-cpu basis
335 * WARNING! Called from early boot, 'mycpu' may not work yet.
338 lwkt_gdinit(struct globaldata *gd)
340 TAILQ_INIT(&gd->gd_tdrunq);
341 TAILQ_INIT(&gd->gd_tdallq);
342 lockinit(&gd->gd_sysctllock, "sysctl", 0, LK_CANRECURSE);
346 * Create a new thread. The thread must be associated with a process context
347 * or LWKT start address before it can be scheduled. If the target cpu is
348 * -1 the thread will be created on the current cpu.
350 * If you intend to create a thread without a process context this function
351 * does everything except load the startup and switcher function.
354 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
356 static int cpu_rotator;
357 globaldata_t gd = mycpu;
361 * If static thread storage is not supplied allocate a thread. Reuse
362 * a cached free thread if possible. gd_freetd is used to keep an exiting
363 * thread intact through the exit.
367 if ((td = gd->gd_freetd) != NULL) {
368 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
370 gd->gd_freetd = NULL;
372 td = objcache_get(thread_cache, M_WAITOK);
373 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
377 KASSERT((td->td_flags &
378 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
379 TDF_ALLOCATED_THREAD,
380 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
381 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
385 * Try to reuse cached stack.
387 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
388 if (flags & TDF_ALLOCATED_STACK) {
389 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
395 stack = (void *)kmem_alloc_stack(&kernel_map, stksize, 0);
397 stack = (void *)kmem_alloc_stack(&kernel_map, stksize,
399 flags |= TDF_ALLOCATED_STACK;
404 cpu = (uint32_t)cpu % (uint32_t)ncpus;
406 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
411 * Initialize a preexisting thread structure. This function is used by
412 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
414 * All threads start out in a critical section at a priority of
415 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
416 * appropriate. This function may send an IPI message when the
417 * requested cpu is not the current cpu and consequently gd_tdallq may
418 * not be initialized synchronously from the point of view of the originating
421 * NOTE! we have to be careful in regards to creating threads for other cpus
422 * if SMP has not yet been activated.
425 lwkt_init_thread_remote(void *arg)
430 * Protected by critical section held by IPI dispatch
432 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
436 * lwkt core thread structural initialization.
438 * NOTE: All threads are initialized as mpsafe threads.
441 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
442 struct globaldata *gd)
444 globaldata_t mygd = mycpu;
446 bzero(td, sizeof(struct thread));
447 td->td_kstack = stack;
448 td->td_kstack_size = stksize;
449 td->td_flags = flags;
451 td->td_type = TD_TYPE_GENERIC;
453 td->td_pri = TDPRI_KERN_DAEMON;
454 td->td_critcount = 1;
455 td->td_toks_have = NULL;
456 td->td_toks_stop = &td->td_toks_base;
457 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
458 lwkt_initport_spin(&td->td_msgport, td,
459 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
461 lwkt_initport_thread(&td->td_msgport, td);
463 pmap_init_thread(td);
465 * Normally initializing a thread for a remote cpu requires sending an
466 * IPI. However, the idlethread is setup before the other cpus are
467 * activated so we have to treat it as a special case. XXX manipulation
468 * of gd_tdallq requires the BGL.
470 if (gd == mygd || td == &gd->gd_idlethread) {
472 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
475 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
477 dsched_enter_thread(td);
481 lwkt_set_comm(thread_t td, const char *ctl, ...)
486 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
488 KTR_LOG(ctxsw_newtd, td, td->td_comm);
492 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
493 * this does not prevent the thread from migrating to another cpu so the
494 * gd_tdallq state is not protected by this.
497 lwkt_hold(thread_t td)
499 atomic_add_int(&td->td_refs, 1);
503 lwkt_rele(thread_t td)
505 KKASSERT(td->td_refs > 0);
506 atomic_add_int(&td->td_refs, -1);
510 lwkt_free_thread(thread_t td)
512 KKASSERT(td->td_refs == 0);
513 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
514 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
515 if (td->td_flags & TDF_ALLOCATED_THREAD) {
516 objcache_put(thread_cache, td);
517 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
518 /* client-allocated struct with internally allocated stack */
519 KASSERT(td->td_kstack && td->td_kstack_size > 0,
520 ("lwkt_free_thread: corrupted stack"));
521 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
522 td->td_kstack = NULL;
523 td->td_kstack_size = 0;
526 KTR_LOG(ctxsw_deadtd, td);
531 * Switch to the next runnable lwkt. If no LWKTs are runnable then
532 * switch to the idlethread. Switching must occur within a critical
533 * section to avoid races with the scheduling queue.
535 * We always have full control over our cpu's run queue. Other cpus
536 * that wish to manipulate our queue must use the cpu_*msg() calls to
537 * talk to our cpu, so a critical section is all that is needed and
538 * the result is very, very fast thread switching.
540 * The LWKT scheduler uses a fixed priority model and round-robins at
541 * each priority level. User process scheduling is a totally
542 * different beast and LWKT priorities should not be confused with
543 * user process priorities.
545 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
546 * is not called by the current thread in the preemption case, only when
547 * the preempting thread blocks (in order to return to the original thread).
549 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
550 * migration and tsleep deschedule the current lwkt thread and call
551 * lwkt_switch(). In particular, the target cpu of the migration fully
552 * expects the thread to become non-runnable and can deadlock against
553 * cpusync operations if we run any IPIs prior to switching the thread out.
555 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
556 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
561 globaldata_t gd = mycpu;
562 thread_t td = gd->gd_curthread;
567 uint64_t tsc_base = rdtsc();
570 KKASSERT(gd->gd_processing_ipiq == 0);
571 KKASSERT(td->td_flags & TDF_RUNNING);
574 * Switching from within a 'fast' (non thread switched) interrupt or IPI
575 * is illegal. However, we may have to do it anyway if we hit a fatal
576 * kernel trap or we have paniced.
578 * If this case occurs save and restore the interrupt nesting level.
580 if (gd->gd_intr_nesting_level) {
584 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
585 panic("lwkt_switch: Attempt to switch from a "
586 "fast interrupt, ipi, or hard code section, "
590 savegdnest = gd->gd_intr_nesting_level;
591 savegdtrap = gd->gd_trap_nesting_level;
592 gd->gd_intr_nesting_level = 0;
593 gd->gd_trap_nesting_level = 0;
594 if ((td->td_flags & TDF_PANICWARN) == 0) {
595 td->td_flags |= TDF_PANICWARN;
596 kprintf("Warning: thread switch from interrupt, IPI, "
597 "or hard code section.\n"
598 "thread %p (%s)\n", td, td->td_comm);
602 gd->gd_intr_nesting_level = savegdnest;
603 gd->gd_trap_nesting_level = savegdtrap;
609 * Release our current user process designation if we are blocking
610 * or if a user reschedule was requested.
612 * NOTE: This function is NOT called if we are switching into or
613 * returning from a preemption.
615 * NOTE: Releasing our current user process designation may cause
616 * it to be assigned to another thread, which in turn will
617 * cause us to block in the usched acquire code when we attempt
618 * to return to userland.
620 * NOTE: On SMP systems this can be very nasty when heavy token
621 * contention is present so we want to be careful not to
622 * release the designation gratuitously.
624 if (td->td_release &&
625 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
630 * Release all tokens. Once we do this we must remain in the critical
631 * section and cannot run IPIs or other interrupts until we switch away
632 * because they may implode if they try to get a token using our thread
636 if (TD_TOKS_HELD(td))
637 lwkt_relalltokens(td);
640 * We had better not be holding any spin locks, but don't get into an
641 * endless panic loop.
643 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
644 ("lwkt_switch: still holding %d exclusive spinlocks!",
648 if (td->td_cscount) {
649 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
651 if (panic_on_cscount)
652 panic("switching while mastering cpusync");
657 * If we had preempted another thread on this cpu, resume the preempted
658 * thread. This occurs transparently, whether the preempted thread
659 * was scheduled or not (it may have been preempted after descheduling
662 * We have to setup the MP lock for the original thread after backing
663 * out the adjustment that was made to curthread when the original
666 if ((ntd = td->td_preempted) != NULL) {
667 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
668 ntd->td_flags |= TDF_PREEMPT_DONE;
669 ntd->td_contended = 0; /* reset contended */
672 * The interrupt may have woken a thread up, we need to properly
673 * set the reschedule flag if the originally interrupted thread is
674 * at a lower priority.
676 * NOTE: The interrupt may not have descheduled ntd.
678 * NOTE: We do not reschedule if there are no threads on the runq.
679 * (ntd could be the idlethread).
681 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
682 if (xtd && xtd != ntd)
684 goto havethread_preempted;
688 * Figure out switch target. If we cannot switch to our desired target
689 * look for a thread that we can switch to.
691 * NOTE! The limited spin loop and related parameters are extremely
692 * important for system performance, particularly for pipes and
693 * concurrent conflicting VM faults.
695 clear_lwkt_resched();
696 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
700 if (TD_TOKS_NOT_HELD(ntd) ||
701 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
705 ++ntd->td_contended; /* overflow ok */
706 if (gd->gd_indefinite.type == 0)
707 indefinite_init(&gd->gd_indefinite, NULL, 0, 't');
709 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
710 kprintf("lwkt_switch: excessive contended %d "
711 "thread %p\n", ntd->td_contended, ntd);
715 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
719 * Bleh, the thread we wanted to switch to has a contended token.
720 * See if we can switch to another thread.
722 * We generally don't want to do this because it represents a
723 * priority inversion, but contending tokens on the same cpu can
724 * cause real problems if we don't now that we have an exclusive
725 * priority mechanism over shared for tokens.
727 * The solution is to allow threads with pending tokens to compete
728 * for them (a lower priority thread will get less cpu once it
729 * returns from the kernel anyway). If a thread does not have
730 * any contending tokens, we go by td_pri and upri.
732 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
733 if (TD_TOKS_NOT_HELD(ntd) &&
734 ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri) {
737 if (upri < ntd->td_upri)
743 if (TD_TOKS_NOT_HELD(ntd) ||
744 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
747 ++ntd->td_contended; /* overflow ok */
751 * Fall through, switch to idle thread to get us out of the current
752 * context. Since we were contended, prevent HLT by flagging a
759 * We either contended on ntd or the runq is empty. We must switch
760 * through the idle thread to get out of the current context.
762 ntd = &gd->gd_idlethread;
763 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
764 ASSERT_NO_TOKENS_HELD(ntd);
765 cpu_time.cp_msg[0] = 0;
770 * Clear gd_idle_repeat when doing a normal switch to a non-idle
773 ntd->td_wmesg = NULL;
774 ntd->td_contended = 0; /* reset once scheduled */
775 ++gd->gd_cnt.v_swtch;
776 gd->gd_idle_repeat = 0;
779 * If we were busy waiting record final disposition
781 if (gd->gd_indefinite.type)
782 indefinite_done(&gd->gd_indefinite);
784 havethread_preempted:
786 * If the new target does not need the MP lock and we are holding it,
787 * release the MP lock. If the new target requires the MP lock we have
788 * already acquired it for the target.
792 KASSERT(ntd->td_critcount,
793 ("priority problem in lwkt_switch %d %d",
794 td->td_critcount, ntd->td_critcount));
798 * Execute the actual thread switch operation. This function
799 * returns to the current thread and returns the previous thread
800 * (which may be different from the thread we switched to).
802 * We are responsible for marking ntd as TDF_RUNNING.
804 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
805 #ifdef DEBUG_LWKT_THREAD
808 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
809 ntd->td_flags |= TDF_RUNNING;
810 lwkt_switch_return(td->td_switch(ntd));
811 /* ntd invalid, td_switch() can return a different thread_t */
815 * catch-all. XXX is this strictly needed?
819 /* NOTE: current cpu may have changed after switch */
824 * Called by assembly in the td_switch (thread restore path) for thread
825 * bootstrap cases which do not 'return' to lwkt_switch().
828 lwkt_switch_return(thread_t otd)
832 uint64_t tsc_base = rdtsc();
836 exiting = otd->td_flags & TDF_EXITING;
840 * Check if otd was migrating. Now that we are on ntd we can finish
841 * up the migration. This is a bit messy but it is the only place
842 * where td is known to be fully descheduled.
844 * We can only activate the migration if otd was migrating but not
845 * held on the cpu due to a preemption chain. We still have to
846 * clear TDF_RUNNING on the old thread either way.
848 * We are responsible for clearing the previously running thread's
851 if ((rgd = otd->td_migrate_gd) != NULL &&
852 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
853 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
854 (TDF_MIGRATING | TDF_RUNNING));
855 otd->td_migrate_gd = NULL;
856 otd->td_flags &= ~TDF_RUNNING;
857 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
859 otd->td_flags &= ~TDF_RUNNING;
863 * Final exit validations (see lwp_wait()). Note that otd becomes
864 * invalid the *instant* we set TDF_MP_EXITSIG.
866 * Use the EXITING status loaded from before we clear TDF_RUNNING,
867 * because if it is not set otd becomes invalid the instant we clear
868 * TDF_RUNNING on it (otherwise, if the system is fast enough, we
869 * might 'steal' TDF_EXITING from another switch-return!).
874 mpflags = otd->td_mpflags;
877 if (mpflags & TDF_MP_EXITWAIT) {
878 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
879 mpflags | TDF_MP_EXITSIG)) {
884 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
885 mpflags | TDF_MP_EXITSIG)) {
892 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
893 kprintf("lwkt_switch_return: excessive TDF_EXITING "
902 * Request that the target thread preempt the current thread. Preemption
903 * can only occur only:
905 * - If our critical section is the one that we were called with
906 * - The relative priority of the target thread is higher
907 * - The target is not excessively interrupt-nested via td_nest_count
908 * - The target thread holds no tokens.
909 * - The target thread is not already scheduled and belongs to the
911 * - The current thread is not holding any spin-locks.
913 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
914 * this is called via lwkt_schedule() through the td_preemptable callback.
915 * critcount is the managed critical priority that we should ignore in order
916 * to determine whether preemption is possible (aka usually just the crit
917 * priority of lwkt_schedule() itself).
919 * Preemption is typically limited to interrupt threads.
921 * Operation works in a fairly straight-forward manner. The normal
922 * scheduling code is bypassed and we switch directly to the target
923 * thread. When the target thread attempts to block or switch away
924 * code at the base of lwkt_switch() will switch directly back to our
925 * thread. Our thread is able to retain whatever tokens it holds and
926 * if the target needs one of them the target will switch back to us
927 * and reschedule itself normally.
930 lwkt_preempt(thread_t ntd, int critcount)
932 struct globaldata *gd = mycpu;
935 int save_gd_intr_nesting_level;
938 * The caller has put us in a critical section. We can only preempt
939 * if the caller of the caller was not in a critical section (basically
940 * a local interrupt), as determined by the 'critcount' parameter. We
941 * also can't preempt if the caller is holding any spinlocks (even if
942 * he isn't in a critical section). This also handles the tokens test.
944 * YYY The target thread must be in a critical section (else it must
945 * inherit our critical section? I dunno yet).
947 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
949 td = gd->gd_curthread;
950 if (preempt_enable == 0) {
951 #ifdef DEBUG_LWKT_THREAD
956 if (ntd->td_pri <= td->td_pri) {
957 #ifdef DEBUG_LWKT_THREAD
962 if (td->td_critcount > critcount) {
963 #ifdef DEBUG_LWKT_THREAD
968 if (td->td_nest_count >= 2) {
969 #ifdef DEBUG_LWKT_THREAD
974 if (td->td_cscount) {
975 #ifdef DEBUG_LWKT_THREAD
980 if (ntd->td_gd != gd) {
981 #ifdef DEBUG_LWKT_THREAD
988 * We don't have to check spinlocks here as they will also bump
991 * Do not try to preempt if the target thread is holding any tokens.
992 * We could try to acquire the tokens but this case is so rare there
993 * is no need to support it.
995 KKASSERT(gd->gd_spinlocks == 0);
997 if (TD_TOKS_HELD(ntd)) {
998 #ifdef DEBUG_LWKT_THREAD
1003 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1004 #ifdef DEBUG_LWKT_THREAD
1009 if (ntd->td_preempted) {
1010 #ifdef DEBUG_LWKT_THREAD
1015 KKASSERT(gd->gd_processing_ipiq == 0);
1018 * Since we are able to preempt the current thread, there is no need to
1019 * call need_lwkt_resched().
1021 * We must temporarily clear gd_intr_nesting_level around the switch
1022 * since switchouts from the target thread are allowed (they will just
1023 * return to our thread), and since the target thread has its own stack.
1025 * A preemption must switch back to the original thread, assert the
1028 #ifdef DEBUG_LWKT_THREAD
1031 ntd->td_preempted = td;
1032 td->td_flags |= TDF_PREEMPT_LOCK;
1033 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1034 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1035 gd->gd_intr_nesting_level = 0;
1037 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1038 ntd->td_flags |= TDF_RUNNING;
1039 xtd = td->td_switch(ntd);
1040 KKASSERT(xtd == ntd);
1041 lwkt_switch_return(xtd);
1042 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1044 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1045 ntd->td_preempted = NULL;
1046 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1050 * Conditionally call splz() if gd_reqflags indicates work is pending.
1051 * This will work inside a critical section but not inside a hard code
1054 * (self contained on a per cpu basis)
1059 globaldata_t gd = mycpu;
1060 thread_t td = gd->gd_curthread;
1062 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1063 gd->gd_intr_nesting_level == 0 &&
1064 td->td_nest_count < 2)
1071 * This version is integrated into crit_exit, reqflags has already
1072 * been tested but td_critcount has not.
1074 * We only want to execute the splz() on the 1->0 transition of
1075 * critcount and not in a hard code section or if too deeply nested.
1077 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
1080 lwkt_maybe_splz(thread_t td)
1082 globaldata_t gd = td->td_gd;
1084 if (td->td_critcount == 0 &&
1085 gd->gd_intr_nesting_level == 0 &&
1086 td->td_nest_count < 2)
1093 * Drivers which set up processing co-threads can call this function to
1094 * run the co-thread at a higher priority and to allow it to preempt
1098 lwkt_set_interrupt_support_thread(void)
1100 thread_t td = curthread;
1102 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1103 td->td_flags |= TDF_INTTHREAD;
1104 td->td_preemptable = lwkt_preempt;
1109 * This function is used to negotiate a passive release of the current
1110 * process/lwp designation with the user scheduler, allowing the user
1111 * scheduler to schedule another user thread. The related kernel thread
1112 * (curthread) continues running in the released state.
1115 lwkt_passive_release(struct thread *td)
1117 struct lwp *lp = td->td_lwp;
1119 td->td_release = NULL;
1120 lwkt_setpri_self(TDPRI_KERN_USER);
1122 lp->lwp_proc->p_usched->release_curproc(lp);
1127 * This implements a LWKT yield, allowing a kernel thread to yield to other
1128 * kernel threads at the same or higher priority. This function can be
1129 * called in a tight loop and will typically only yield once per tick.
1131 * Most kernel threads run at the same priority in order to allow equal
1134 * (self contained on a per cpu basis)
1139 globaldata_t gd = mycpu;
1140 thread_t td = gd->gd_curthread;
1143 * Should never be called with spinlocks held but there is a path
1144 * via ACPI where it might happen.
1146 if (gd->gd_spinlocks)
1150 * Safe to call splz if we are not too-heavily nested.
1152 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1156 * Caller allows switching
1158 if (lwkt_resched_wanted()) {
1159 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
1160 lwkt_schedule_self(td);
1166 * The quick version processes pending interrupts and higher-priority
1167 * LWKT threads but will not round-robin same-priority LWKT threads.
1169 * When called while attempting to return to userland the only same-pri
1170 * threads are the ones which have already tried to become the current
1174 lwkt_yield_quick(void)
1176 globaldata_t gd = mycpu;
1177 thread_t td = gd->gd_curthread;
1179 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1181 if (lwkt_resched_wanted()) {
1183 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1184 clear_lwkt_resched();
1186 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD);
1187 lwkt_schedule_self(curthread);
1195 * This yield is designed for kernel threads with a user context.
1197 * The kernel acting on behalf of the user is potentially cpu-bound,
1198 * this function will efficiently allow other threads to run and also
1199 * switch to other processes by releasing.
1201 * The lwkt_user_yield() function is designed to have very low overhead
1202 * if no yield is determined to be needed.
1205 lwkt_user_yield(void)
1207 globaldata_t gd = mycpu;
1208 thread_t td = gd->gd_curthread;
1211 * Should never be called with spinlocks held but there is a path
1212 * via ACPI where it might happen.
1214 if (gd->gd_spinlocks)
1218 * Always run any pending interrupts in case we are in a critical
1221 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1225 * Switch (which forces a release) if another kernel thread needs
1226 * the cpu, if userland wants us to resched, or if our kernel
1227 * quantum has run out.
1229 if (lwkt_resched_wanted() ||
1230 user_resched_wanted())
1237 * Reacquire the current process if we are released.
1239 * XXX not implemented atm. The kernel may be holding locks and such,
1240 * so we want the thread to continue to receive cpu.
1242 if (td->td_release == NULL && lp) {
1243 lp->lwp_proc->p_usched->acquire_curproc(lp);
1244 td->td_release = lwkt_passive_release;
1245 lwkt_setpri_self(TDPRI_USER_NORM);
1251 * Generic schedule. Possibly schedule threads belonging to other cpus and
1252 * deal with threads that might be blocked on a wait queue.
1254 * We have a little helper inline function which does additional work after
1255 * the thread has been enqueued, including dealing with preemption and
1256 * setting need_lwkt_resched() (which prevents the kernel from returning
1257 * to userland until it has processed higher priority threads).
1259 * It is possible for this routine to be called after a failed _enqueue
1260 * (due to the target thread migrating, sleeping, or otherwise blocked).
1261 * We have to check that the thread is actually on the run queue!
1265 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1267 if (ntd->td_flags & TDF_RUNQ) {
1268 if (ntd->td_preemptable) {
1269 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1276 _lwkt_schedule(thread_t td)
1278 globaldata_t mygd = mycpu;
1280 KASSERT(td != &td->td_gd->gd_idlethread,
1281 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1282 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1283 crit_enter_gd(mygd);
1284 KKASSERT(td->td_lwp == NULL ||
1285 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1287 if (td == mygd->gd_curthread) {
1291 * If we own the thread, there is no race (since we are in a
1292 * critical section). If we do not own the thread there might
1293 * be a race but the target cpu will deal with it.
1295 if (td->td_gd == mygd) {
1297 _lwkt_schedule_post(mygd, td, 1);
1299 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1306 lwkt_schedule(thread_t td)
1312 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1318 * When scheduled remotely if frame != NULL the IPIQ is being
1319 * run via doreti or an interrupt then preemption can be allowed.
1321 * To allow preemption we have to drop the critical section so only
1322 * one is present in _lwkt_schedule_post.
1325 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1327 thread_t td = curthread;
1330 if (frame && ntd->td_preemptable) {
1331 crit_exit_noyield(td);
1332 _lwkt_schedule(ntd);
1333 crit_enter_quick(td);
1335 _lwkt_schedule(ntd);
1340 * Thread migration using a 'Pull' method. The thread may or may not be
1341 * the current thread. It MUST be descheduled and in a stable state.
1342 * lwkt_giveaway() must be called on the cpu owning the thread.
1344 * At any point after lwkt_giveaway() is called, the target cpu may
1345 * 'pull' the thread by calling lwkt_acquire().
1347 * We have to make sure the thread is not sitting on a per-cpu tsleep
1348 * queue or it will blow up when it moves to another cpu.
1350 * MPSAFE - must be called under very specific conditions.
1353 lwkt_giveaway(thread_t td)
1355 globaldata_t gd = mycpu;
1358 if (td->td_flags & TDF_TSLEEPQ)
1360 KKASSERT(td->td_gd == gd);
1361 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1362 td->td_flags |= TDF_MIGRATING;
1367 lwkt_acquire(thread_t td)
1372 KKASSERT(td->td_flags & TDF_MIGRATING);
1377 uint64_t tsc_base = rdtsc();
1380 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1381 crit_enter_gd(mygd);
1382 DEBUG_PUSH_INFO("lwkt_acquire");
1383 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1384 lwkt_process_ipiq();
1386 #ifdef _KERNEL_VIRTUAL
1390 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1391 kprintf("lwkt_acquire: stuck td %p td->td_flags %08x\n",
1400 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1401 td->td_flags &= ~TDF_MIGRATING;
1404 crit_enter_gd(mygd);
1405 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1406 td->td_flags &= ~TDF_MIGRATING;
1412 * Generic deschedule. Descheduling threads other then your own should be
1413 * done only in carefully controlled circumstances. Descheduling is
1416 * This function may block if the cpu has run out of messages.
1419 lwkt_deschedule(thread_t td)
1422 if (td == curthread) {
1425 if (td->td_gd == mycpu) {
1428 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1435 * Set the target thread's priority. This routine does not automatically
1436 * switch to a higher priority thread, LWKT threads are not designed for
1437 * continuous priority changes. Yield if you want to switch.
1440 lwkt_setpri(thread_t td, int pri)
1442 if (td->td_pri != pri) {
1445 if (td->td_flags & TDF_RUNQ) {
1446 KKASSERT(td->td_gd == mycpu);
1458 * Set the initial priority for a thread prior to it being scheduled for
1459 * the first time. The thread MUST NOT be scheduled before or during
1460 * this call. The thread may be assigned to a cpu other then the current
1463 * Typically used after a thread has been created with TDF_STOPPREQ,
1464 * and before the thread is initially scheduled.
1467 lwkt_setpri_initial(thread_t td, int pri)
1470 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1475 lwkt_setpri_self(int pri)
1477 thread_t td = curthread;
1479 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1481 if (td->td_flags & TDF_RUNQ) {
1492 * hz tick scheduler clock for LWKT threads
1495 lwkt_schedulerclock(thread_t td)
1497 globaldata_t gd = td->td_gd;
1500 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1503 * If the current thread is at the head of the runq shift it to the
1504 * end of any equal-priority threads and request a LWKT reschedule
1507 * Ignore upri in this situation. There will only be one user thread
1508 * in user mode, all others will be user threads running in kernel
1509 * mode and we have to make sure they get some cpu.
1511 xtd = TAILQ_NEXT(td, td_threadq);
1512 if (xtd && xtd->td_pri == td->td_pri) {
1513 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1514 while (xtd && xtd->td_pri == td->td_pri)
1515 xtd = TAILQ_NEXT(xtd, td_threadq);
1517 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1519 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1520 need_lwkt_resched();
1524 * If we scheduled a thread other than the one at the head of the
1525 * queue always request a reschedule every tick.
1527 need_lwkt_resched();
1529 /* else curthread probably the idle thread, no need to reschedule */
1533 * Migrate the current thread to the specified cpu.
1535 * This is accomplished by descheduling ourselves from the current cpu
1536 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1537 * 'old' thread wants to migrate after it has been completely switched out
1538 * and will complete the migration.
1540 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1542 * We must be sure to release our current process designation (if a user
1543 * process) before clearing out any tsleepq we are on because the release
1544 * code may re-add us.
1546 * We must be sure to remove ourselves from the current cpu's tsleepq
1547 * before potentially moving to another queue. The thread can be on
1548 * a tsleepq due to a left-over tsleep_interlock().
1552 lwkt_setcpu_self(globaldata_t rgd)
1554 thread_t td = curthread;
1556 if (td->td_gd != rgd) {
1557 crit_enter_quick(td);
1561 if (td->td_flags & TDF_TSLEEPQ)
1565 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1566 * trying to deschedule ourselves and switch away, then deschedule
1567 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1568 * call lwkt_switch() to complete the operation.
1570 td->td_flags |= TDF_MIGRATING;
1571 lwkt_deschedule_self(td);
1572 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1573 td->td_migrate_gd = rgd;
1577 * We are now on the target cpu
1579 KKASSERT(rgd == mycpu);
1580 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1581 crit_exit_quick(td);
1586 lwkt_migratecpu(int cpuid)
1590 rgd = globaldata_find(cpuid);
1591 lwkt_setcpu_self(rgd);
1595 * Remote IPI for cpu migration (called while in a critical section so we
1596 * do not have to enter another one).
1598 * The thread (td) has already been completely descheduled from the
1599 * originating cpu and we can simply assert the case. The thread is
1600 * assigned to the new cpu and enqueued.
1602 * The thread will re-add itself to tdallq when it resumes execution.
1605 lwkt_setcpu_remote(void *arg)
1608 globaldata_t gd = mycpu;
1610 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1613 td->td_flags &= ~TDF_MIGRATING;
1614 KKASSERT(td->td_migrate_gd == NULL);
1615 KKASSERT(td->td_lwp == NULL ||
1616 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1621 lwkt_preempted_proc(void)
1623 thread_t td = curthread;
1624 while (td->td_preempted)
1625 td = td->td_preempted;
1630 * Create a kernel process/thread/whatever. It shares it's address space
1631 * with proc0 - ie: kernel only.
1633 * If the cpu is not specified one will be selected. In the future
1634 * specifying a cpu of -1 will enable kernel thread migration between
1638 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1639 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1644 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1648 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1651 * Set up arg0 for 'ps' etc
1653 __va_start(ap, fmt);
1654 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1658 * Schedule the thread to run
1660 if (td->td_flags & TDF_NOSTART)
1661 td->td_flags &= ~TDF_NOSTART;
1668 * Destroy an LWKT thread. Warning! This function is not called when
1669 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1670 * uses a different reaping mechanism.
1675 thread_t td = curthread;
1680 * Do any cleanup that might block here
1683 dsched_exit_thread(td);
1686 * Get us into a critical section to interlock gd_freetd and loop
1687 * until we can get it freed.
1689 * We have to cache the current td in gd_freetd because objcache_put()ing
1690 * it would rip it out from under us while our thread is still active.
1692 * We are the current thread so of course our own TDF_RUNNING bit will
1693 * be set, so unlike the lwp reap code we don't wait for it to clear.
1696 crit_enter_quick(td);
1699 tsleep(td, 0, "tdreap", 1);
1702 if ((std = gd->gd_freetd) != NULL) {
1703 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1704 gd->gd_freetd = NULL;
1705 objcache_put(thread_cache, std);
1712 * Remove thread resources from kernel lists and deschedule us for
1713 * the last time. We cannot block after this point or we may end
1714 * up with a stale td on the tsleepq.
1716 * None of this may block, the critical section is the only thing
1717 * protecting tdallq and the only thing preventing new lwkt_hold()
1720 if (td->td_flags & TDF_TSLEEPQ)
1722 lwkt_deschedule_self(td);
1723 lwkt_remove_tdallq(td);
1724 KKASSERT(td->td_refs == 0);
1729 KKASSERT(gd->gd_freetd == NULL);
1730 if (td->td_flags & TDF_ALLOCATED_THREAD)
1736 lwkt_remove_tdallq(thread_t td)
1738 KKASSERT(td->td_gd == mycpu);
1739 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1743 * Code reduction and branch prediction improvements. Call/return
1744 * overhead on modern cpus often degenerates into 0 cycles due to
1745 * the cpu's branch prediction hardware and return pc cache. We
1746 * can take advantage of this by not inlining medium-complexity
1747 * functions and we can also reduce the branch prediction impact
1748 * by collapsing perfectly predictable branches into a single
1749 * procedure instead of duplicating it.
1751 * Is any of this noticeable? Probably not, so I'll take the
1752 * smaller code size.
1755 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1757 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1763 thread_t td = curthread;
1764 int lcrit = td->td_critcount;
1766 td->td_critcount = 0;
1768 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1773 * Called from debugger/panic on cpus which have been stopped. We must still
1774 * process the IPIQ while stopped.
1776 * If we are dumping also try to process any pending interrupts. This may
1777 * or may not work depending on the state of the cpu at the point it was
1781 lwkt_smp_stopped(void)
1783 globaldata_t gd = mycpu;
1786 lwkt_process_ipiq();
1787 --gd->gd_intr_nesting_level;
1789 ++gd->gd_intr_nesting_level;
1791 lwkt_process_ipiq();