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 * NOTE: There are a limited number of lwkt threads runnable since user
184 * processes only schedule one at a time per cpu.
188 _lwkt_enqueue(thread_t td)
192 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
193 struct globaldata *gd = td->td_gd;
195 td->td_flags |= TDF_RUNQ;
196 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
198 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
199 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
201 while (xtd && xtd->td_pri >= td->td_pri)
202 xtd = TAILQ_NEXT(xtd, td_threadq);
204 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
206 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
208 ++gd->gd_tdrunqcount;
211 * Request a LWKT reschedule if we are now at the head of the queue.
213 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
219 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
221 struct thread *td = (struct thread *)obj;
223 td->td_kstack = NULL;
224 td->td_kstack_size = 0;
225 td->td_flags = TDF_ALLOCATED_THREAD;
231 _lwkt_thread_dtor(void *obj, void *privdata)
233 struct thread *td = (struct thread *)obj;
235 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
236 ("_lwkt_thread_dtor: not allocated from objcache"));
237 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
238 td->td_kstack_size > 0,
239 ("_lwkt_thread_dtor: corrupted stack"));
240 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
241 td->td_kstack = NULL;
246 * Initialize the lwkt s/system.
248 * Nominally cache up to 32 thread + kstack structures. Cache more on
249 * systems with a lot of cpu cores.
254 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
255 if (lwkt_cache_threads == 0) {
256 lwkt_cache_threads = ncpus * 4;
257 if (lwkt_cache_threads < 32)
258 lwkt_cache_threads = 32;
260 thread_cache = objcache_create_mbacked(
261 M_THREAD, sizeof(struct thread),
262 NULL, lwkt_cache_threads,
263 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
267 * Schedule a thread to run. As the current thread we can always safely
268 * schedule ourselves, and a shortcut procedure is provided for that
271 * (non-blocking, self contained on a per cpu basis)
274 lwkt_schedule_self(thread_t td)
276 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
277 crit_enter_quick(td);
278 KASSERT(td != &td->td_gd->gd_idlethread,
279 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
280 KKASSERT(td->td_lwp == NULL ||
281 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
287 * Deschedule a thread.
289 * (non-blocking, self contained on a per cpu basis)
292 lwkt_deschedule_self(thread_t td)
294 crit_enter_quick(td);
300 * LWKTs operate on a per-cpu basis
302 * WARNING! Called from early boot, 'mycpu' may not work yet.
305 lwkt_gdinit(struct globaldata *gd)
307 TAILQ_INIT(&gd->gd_tdrunq);
308 TAILQ_INIT(&gd->gd_tdallq);
312 * Create a new thread. The thread must be associated with a process context
313 * or LWKT start address before it can be scheduled. If the target cpu is
314 * -1 the thread will be created on the current cpu.
316 * If you intend to create a thread without a process context this function
317 * does everything except load the startup and switcher function.
320 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
322 static int cpu_rotator;
323 globaldata_t gd = mycpu;
327 * If static thread storage is not supplied allocate a thread. Reuse
328 * a cached free thread if possible. gd_freetd is used to keep an exiting
329 * thread intact through the exit.
333 if ((td = gd->gd_freetd) != NULL) {
334 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
336 gd->gd_freetd = NULL;
338 td = objcache_get(thread_cache, M_WAITOK);
339 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
343 KASSERT((td->td_flags &
344 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
345 TDF_ALLOCATED_THREAD,
346 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
347 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
351 * Try to reuse cached stack.
353 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
354 if (flags & TDF_ALLOCATED_STACK) {
355 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
360 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
361 flags |= TDF_ALLOCATED_STACK;
368 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
373 * Initialize a preexisting thread structure. This function is used by
374 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
376 * All threads start out in a critical section at a priority of
377 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
378 * appropriate. This function may send an IPI message when the
379 * requested cpu is not the current cpu and consequently gd_tdallq may
380 * not be initialized synchronously from the point of view of the originating
383 * NOTE! we have to be careful in regards to creating threads for other cpus
384 * if SMP has not yet been activated.
389 lwkt_init_thread_remote(void *arg)
394 * Protected by critical section held by IPI dispatch
396 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
402 * lwkt core thread structural initialization.
404 * NOTE: All threads are initialized as mpsafe threads.
407 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
408 struct globaldata *gd)
410 globaldata_t mygd = mycpu;
412 bzero(td, sizeof(struct thread));
413 td->td_kstack = stack;
414 td->td_kstack_size = stksize;
415 td->td_flags = flags;
418 td->td_pri = TDPRI_KERN_DAEMON;
419 td->td_critcount = 1;
420 td->td_toks_have = NULL;
421 td->td_toks_stop = &td->td_toks_base;
422 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT))
423 lwkt_initport_spin(&td->td_msgport, td);
425 lwkt_initport_thread(&td->td_msgport, td);
426 pmap_init_thread(td);
429 * Normally initializing a thread for a remote cpu requires sending an
430 * IPI. However, the idlethread is setup before the other cpus are
431 * activated so we have to treat it as a special case. XXX manipulation
432 * of gd_tdallq requires the BGL.
434 if (gd == mygd || td == &gd->gd_idlethread) {
436 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
439 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
443 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
447 dsched_new_thread(td);
451 lwkt_set_comm(thread_t td, const char *ctl, ...)
456 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
458 KTR_LOG(ctxsw_newtd, td, td->td_comm);
462 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
463 * this does not prevent the thread from migrating to another cpu so the
464 * gd_tdallq state is not protected by this.
467 lwkt_hold(thread_t td)
469 atomic_add_int(&td->td_refs, 1);
473 lwkt_rele(thread_t td)
475 KKASSERT(td->td_refs > 0);
476 atomic_add_int(&td->td_refs, -1);
480 lwkt_free_thread(thread_t td)
482 KKASSERT(td->td_refs == 0);
483 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
484 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
485 if (td->td_flags & TDF_ALLOCATED_THREAD) {
486 objcache_put(thread_cache, td);
487 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
488 /* client-allocated struct with internally allocated stack */
489 KASSERT(td->td_kstack && td->td_kstack_size > 0,
490 ("lwkt_free_thread: corrupted stack"));
491 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
492 td->td_kstack = NULL;
493 td->td_kstack_size = 0;
495 KTR_LOG(ctxsw_deadtd, td);
500 * Switch to the next runnable lwkt. If no LWKTs are runnable then
501 * switch to the idlethread. Switching must occur within a critical
502 * section to avoid races with the scheduling queue.
504 * We always have full control over our cpu's run queue. Other cpus
505 * that wish to manipulate our queue must use the cpu_*msg() calls to
506 * talk to our cpu, so a critical section is all that is needed and
507 * the result is very, very fast thread switching.
509 * The LWKT scheduler uses a fixed priority model and round-robins at
510 * each priority level. User process scheduling is a totally
511 * different beast and LWKT priorities should not be confused with
512 * user process priorities.
514 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
515 * is not called by the current thread in the preemption case, only when
516 * the preempting thread blocks (in order to return to the original thread).
518 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
519 * migration and tsleep deschedule the current lwkt thread and call
520 * lwkt_switch(). In particular, the target cpu of the migration fully
521 * expects the thread to become non-runnable and can deadlock against
522 * cpusync operations if we run any IPIs prior to switching the thread out.
524 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
525 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
530 globaldata_t gd = mycpu;
531 thread_t td = gd->gd_curthread;
535 KKASSERT(gd->gd_processing_ipiq == 0);
536 KKASSERT(td->td_flags & TDF_RUNNING);
539 * Switching from within a 'fast' (non thread switched) interrupt or IPI
540 * is illegal. However, we may have to do it anyway if we hit a fatal
541 * kernel trap or we have paniced.
543 * If this case occurs save and restore the interrupt nesting level.
545 if (gd->gd_intr_nesting_level) {
549 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
550 panic("lwkt_switch: Attempt to switch from a "
551 "fast interrupt, ipi, or hard code section, "
555 savegdnest = gd->gd_intr_nesting_level;
556 savegdtrap = gd->gd_trap_nesting_level;
557 gd->gd_intr_nesting_level = 0;
558 gd->gd_trap_nesting_level = 0;
559 if ((td->td_flags & TDF_PANICWARN) == 0) {
560 td->td_flags |= TDF_PANICWARN;
561 kprintf("Warning: thread switch from interrupt, IPI, "
562 "or hard code section.\n"
563 "thread %p (%s)\n", td, td->td_comm);
567 gd->gd_intr_nesting_level = savegdnest;
568 gd->gd_trap_nesting_level = savegdtrap;
574 * Release our current user process designation if we are blocking
575 * or if a user reschedule was requested.
577 * NOTE: This function is NOT called if we are switching into or
578 * returning from a preemption.
580 * NOTE: Releasing our current user process designation may cause
581 * it to be assigned to another thread, which in turn will
582 * cause us to block in the usched acquire code when we attempt
583 * to return to userland.
585 * NOTE: On SMP systems this can be very nasty when heavy token
586 * contention is present so we want to be careful not to
587 * release the designation gratuitously.
589 if (td->td_release &&
590 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
598 if (TD_TOKS_HELD(td))
599 lwkt_relalltokens(td);
602 * We had better not be holding any spin locks, but don't get into an
603 * endless panic loop.
605 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
606 ("lwkt_switch: still holding %d exclusive spinlocks!",
607 gd->gd_spinlocks_wr));
612 if (td->td_cscount) {
613 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
615 if (panic_on_cscount)
616 panic("switching while mastering cpusync");
622 * If we had preempted another thread on this cpu, resume the preempted
623 * thread. This occurs transparently, whether the preempted thread
624 * was scheduled or not (it may have been preempted after descheduling
627 * We have to setup the MP lock for the original thread after backing
628 * out the adjustment that was made to curthread when the original
631 if ((ntd = td->td_preempted) != NULL) {
632 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
633 ntd->td_flags |= TDF_PREEMPT_DONE;
636 * The interrupt may have woken a thread up, we need to properly
637 * set the reschedule flag if the originally interrupted thread is
638 * at a lower priority.
640 * The interrupt may not have descheduled.
642 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
644 goto havethread_preempted;
648 * If we cannot obtain ownership of the tokens we cannot immediately
649 * schedule the target thread.
651 * Reminder: Again, we cannot afford to run any IPIs in this path if
652 * the current thread has been descheduled.
655 clear_lwkt_resched();
658 * Hotpath - pull the head of the run queue and attempt to schedule
661 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
665 * Runq is empty, switch to idle to allow it to halt.
667 ntd = &gd->gd_idlethread;
669 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
670 ASSERT_NO_TOKENS_HELD(ntd);
672 cpu_time.cp_msg[0] = 0;
673 cpu_time.cp_stallpc = 0;
678 * Hotpath - schedule ntd.
680 * NOTE: For UP there is no mplock and lwkt_getalltokens()
683 if (TD_TOKS_NOT_HELD(ntd) ||
684 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops)))
690 * Coldpath (SMP only since tokens always succeed on UP)
692 * We had some contention on the thread we wanted to schedule.
693 * What we do now is try to find a thread that we can schedule
696 * The coldpath scan does NOT rearrange threads in the run list.
697 * The lwkt_schedulerclock() will assert need_lwkt_resched() on
698 * the next tick whenever the current head is not the current thread.
703 ++gd->gd_cnt.v_token_colls;
705 if (fairq_bypass > 0)
708 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
710 * Never schedule threads returning to userland or the
711 * user thread scheduler helper thread when higher priority
712 * threads are present. The runq is sorted by priority
713 * so we can give up traversing it when we find the first
714 * low priority thread.
716 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
724 if (TD_TOKS_NOT_HELD(ntd) ||
725 lwkt_getalltokens(ntd, (spinning >= lwkt_spin_loops))) {
731 ++gd->gd_cnt.v_token_colls;
736 * We exhausted the run list, meaning that all runnable threads
740 ntd = &gd->gd_idlethread;
742 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
743 ASSERT_NO_TOKENS_HELD(ntd);
744 /* contention case, do not clear contention mask */
748 * We are going to have to retry but if the current thread is not
749 * on the runq we instead switch through the idle thread to get away
750 * from the current thread. We have to flag for lwkt reschedule
751 * to prevent the idle thread from halting.
753 * NOTE: A non-zero spinning is passed to lwkt_getalltokens() to
754 * instruct it to deal with the potential for deadlocks by
755 * ordering the tokens by address.
757 if ((td->td_flags & TDF_RUNQ) == 0) {
758 need_lwkt_resched(); /* prevent hlt */
761 #if defined(INVARIANTS) && defined(__amd64__)
762 if ((read_rflags() & PSL_I) == 0) {
764 panic("lwkt_switch() called with interrupts disabled");
769 * Number iterations so far. After a certain point we switch to
770 * a sorted-address/monitor/mwait version of lwkt_getalltokens()
772 if (spinning < 0x7FFFFFFF)
777 * lwkt_getalltokens() failed in sorted token mode, we can use
778 * monitor/mwait in this case.
780 if (spinning >= lwkt_spin_loops &&
781 (cpu_mi_feature & CPU_MI_MONITOR) &&
784 cpu_mmw_pause_int(&gd->gd_reqflags,
785 (gd->gd_reqflags | RQF_SPINNING) &
786 ~RQF_IDLECHECK_WK_MASK);
791 * We already checked that td is still scheduled so this should be
797 * This experimental resequencer is used as a fall-back to reduce
798 * hw cache line contention by placing each core's scheduler into a
799 * time-domain-multplexed slot.
801 * The resequencer is disabled by default. It's functionality has
802 * largely been superceeded by the token algorithm which limits races
803 * to a subset of cores.
805 * The resequencer algorithm tends to break down when more than
806 * 20 cores are contending. What appears to happen is that new
807 * tokens can be obtained out of address-sorted order by new cores
808 * while existing cores languish in long delays between retries and
809 * wind up being starved-out of the token acquisition.
811 if (lwkt_spin_reseq && spinning >= lwkt_spin_reseq) {
812 int cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
815 while ((oseq = lwkt_cseq_rindex) != cseq) {
818 if (cpu_mi_feature & CPU_MI_MONITOR) {
819 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
829 atomic_add_int(&lwkt_cseq_rindex, 1);
831 /* highest level for(;;) loop */
836 * Clear gd_idle_repeat when doing a normal switch to a non-idle
839 ntd->td_wmesg = NULL;
840 ++gd->gd_cnt.v_swtch;
841 gd->gd_idle_repeat = 0;
843 havethread_preempted:
845 * If the new target does not need the MP lock and we are holding it,
846 * release the MP lock. If the new target requires the MP lock we have
847 * already acquired it for the target.
851 KASSERT(ntd->td_critcount,
852 ("priority problem in lwkt_switch %d %d",
853 td->td_critcount, ntd->td_critcount));
857 * Execute the actual thread switch operation. This function
858 * returns to the current thread and returns the previous thread
859 * (which may be different from the thread we switched to).
861 * We are responsible for marking ntd as TDF_RUNNING.
863 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
865 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
866 ntd->td_flags |= TDF_RUNNING;
867 lwkt_switch_return(td->td_switch(ntd));
868 /* ntd invalid, td_switch() can return a different thread_t */
872 * catch-all. XXX is this strictly needed?
876 /* NOTE: current cpu may have changed after switch */
881 * Called by assembly in the td_switch (thread restore path) for thread
882 * bootstrap cases which do not 'return' to lwkt_switch().
885 lwkt_switch_return(thread_t otd)
891 * Check if otd was migrating. Now that we are on ntd we can finish
892 * up the migration. This is a bit messy but it is the only place
893 * where td is known to be fully descheduled.
895 * We can only activate the migration if otd was migrating but not
896 * held on the cpu due to a preemption chain. We still have to
897 * clear TDF_RUNNING on the old thread either way.
899 * We are responsible for clearing the previously running thread's
902 if ((rgd = otd->td_migrate_gd) != NULL &&
903 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
904 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
905 (TDF_MIGRATING | TDF_RUNNING));
906 otd->td_migrate_gd = NULL;
907 otd->td_flags &= ~TDF_RUNNING;
908 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
910 otd->td_flags &= ~TDF_RUNNING;
913 otd->td_flags &= ~TDF_RUNNING;
917 * Final exit validations (see lwp_wait()). Note that otd becomes
918 * invalid the *instant* we set TDF_MP_EXITSIG.
920 while (otd->td_flags & TDF_EXITING) {
923 mpflags = otd->td_mpflags;
926 if (mpflags & TDF_MP_EXITWAIT) {
927 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
928 mpflags | TDF_MP_EXITSIG)) {
933 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
934 mpflags | TDF_MP_EXITSIG)) {
943 * Request that the target thread preempt the current thread. Preemption
944 * can only occur if our only critical section is the one that we were called
945 * with, the relative priority of the target thread is higher, and the target
946 * thread holds no tokens. This also only works if we are not holding any
947 * spinlocks (obviously).
949 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
950 * this is called via lwkt_schedule() through the td_preemptable callback.
951 * critcount is the managed critical priority that we should ignore in order
952 * to determine whether preemption is possible (aka usually just the crit
953 * priority of lwkt_schedule() itself).
955 * Preemption is typically limited to interrupt threads.
957 * Operation works in a fairly straight-forward manner. The normal
958 * scheduling code is bypassed and we switch directly to the target
959 * thread. When the target thread attempts to block or switch away
960 * code at the base of lwkt_switch() will switch directly back to our
961 * thread. Our thread is able to retain whatever tokens it holds and
962 * if the target needs one of them the target will switch back to us
963 * and reschedule itself normally.
966 lwkt_preempt(thread_t ntd, int critcount)
968 struct globaldata *gd = mycpu;
971 int save_gd_intr_nesting_level;
974 * The caller has put us in a critical section. We can only preempt
975 * if the caller of the caller was not in a critical section (basically
976 * a local interrupt), as determined by the 'critcount' parameter. We
977 * also can't preempt if the caller is holding any spinlocks (even if
978 * he isn't in a critical section). This also handles the tokens test.
980 * YYY The target thread must be in a critical section (else it must
981 * inherit our critical section? I dunno yet).
983 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
985 td = gd->gd_curthread;
986 if (preempt_enable == 0) {
990 if (ntd->td_pri <= td->td_pri) {
994 if (td->td_critcount > critcount) {
999 if (td->td_cscount) {
1003 if (ntd->td_gd != gd) {
1009 * We don't have to check spinlocks here as they will also bump
1012 * Do not try to preempt if the target thread is holding any tokens.
1013 * We could try to acquire the tokens but this case is so rare there
1014 * is no need to support it.
1016 KKASSERT(gd->gd_spinlocks_wr == 0);
1018 if (TD_TOKS_HELD(ntd)) {
1022 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1026 if (ntd->td_preempted) {
1030 KKASSERT(gd->gd_processing_ipiq == 0);
1033 * Since we are able to preempt the current thread, there is no need to
1034 * call need_lwkt_resched().
1036 * We must temporarily clear gd_intr_nesting_level around the switch
1037 * since switchouts from the target thread are allowed (they will just
1038 * return to our thread), and since the target thread has its own stack.
1040 * A preemption must switch back to the original thread, assert the
1044 ntd->td_preempted = td;
1045 td->td_flags |= TDF_PREEMPT_LOCK;
1046 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1047 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1048 gd->gd_intr_nesting_level = 0;
1050 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
1051 ntd->td_flags |= TDF_RUNNING;
1052 xtd = td->td_switch(ntd);
1053 KKASSERT(xtd == ntd);
1054 lwkt_switch_return(xtd);
1055 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1057 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1058 ntd->td_preempted = NULL;
1059 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1063 * Conditionally call splz() if gd_reqflags indicates work is pending.
1064 * This will work inside a critical section but not inside a hard code
1067 * (self contained on a per cpu basis)
1072 globaldata_t gd = mycpu;
1073 thread_t td = gd->gd_curthread;
1075 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1076 gd->gd_intr_nesting_level == 0 &&
1077 td->td_nest_count < 2)
1084 * This version is integrated into crit_exit, reqflags has already
1085 * been tested but td_critcount has not.
1087 * We only want to execute the splz() on the 1->0 transition of
1088 * critcount and not in a hard code section or if too deeply nested.
1090 * NOTE: gd->gd_spinlocks_wr is implied to be 0 when td_critcount is 0.
1093 lwkt_maybe_splz(thread_t td)
1095 globaldata_t gd = td->td_gd;
1097 if (td->td_critcount == 0 &&
1098 gd->gd_intr_nesting_level == 0 &&
1099 td->td_nest_count < 2)
1106 * Drivers which set up processing co-threads can call this function to
1107 * run the co-thread at a higher priority and to allow it to preempt
1111 lwkt_set_interrupt_support_thread(void)
1113 thread_t td = curthread;
1115 lwkt_setpri_self(TDPRI_INT_SUPPORT);
1116 td->td_flags |= TDF_INTTHREAD;
1117 td->td_preemptable = lwkt_preempt;
1122 * This function is used to negotiate a passive release of the current
1123 * process/lwp designation with the user scheduler, allowing the user
1124 * scheduler to schedule another user thread. The related kernel thread
1125 * (curthread) continues running in the released state.
1128 lwkt_passive_release(struct thread *td)
1130 struct lwp *lp = td->td_lwp;
1132 td->td_release = NULL;
1133 lwkt_setpri_self(TDPRI_KERN_USER);
1134 lp->lwp_proc->p_usched->release_curproc(lp);
1139 * This implements a LWKT yield, allowing a kernel thread to yield to other
1140 * kernel threads at the same or higher priority. This function can be
1141 * called in a tight loop and will typically only yield once per tick.
1143 * Most kernel threads run at the same priority in order to allow equal
1146 * (self contained on a per cpu basis)
1151 globaldata_t gd = mycpu;
1152 thread_t td = gd->gd_curthread;
1154 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1156 if (lwkt_resched_wanted()) {
1157 lwkt_schedule_self(curthread);
1163 * The quick version processes pending interrupts and higher-priority
1164 * LWKT threads but will not round-robin same-priority LWKT threads.
1166 * When called while attempting to return to userland the only same-pri
1167 * threads are the ones which have already tried to become the current
1171 lwkt_yield_quick(void)
1173 globaldata_t gd = mycpu;
1174 thread_t td = gd->gd_curthread;
1176 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1178 if (lwkt_resched_wanted()) {
1179 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1180 clear_lwkt_resched();
1182 lwkt_schedule_self(curthread);
1189 * This yield is designed for kernel threads with a user context.
1191 * The kernel acting on behalf of the user is potentially cpu-bound,
1192 * this function will efficiently allow other threads to run and also
1193 * switch to other processes by releasing.
1195 * The lwkt_user_yield() function is designed to have very low overhead
1196 * if no yield is determined to be needed.
1199 lwkt_user_yield(void)
1201 globaldata_t gd = mycpu;
1202 thread_t td = gd->gd_curthread;
1205 * Always run any pending interrupts in case we are in a critical
1208 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1212 * Switch (which forces a release) if another kernel thread needs
1213 * the cpu, if userland wants us to resched, or if our kernel
1214 * quantum has run out.
1216 if (lwkt_resched_wanted() ||
1217 user_resched_wanted())
1224 * Reacquire the current process if we are released.
1226 * XXX not implemented atm. The kernel may be holding locks and such,
1227 * so we want the thread to continue to receive cpu.
1229 if (td->td_release == NULL && lp) {
1230 lp->lwp_proc->p_usched->acquire_curproc(lp);
1231 td->td_release = lwkt_passive_release;
1232 lwkt_setpri_self(TDPRI_USER_NORM);
1238 * Generic schedule. Possibly schedule threads belonging to other cpus and
1239 * deal with threads that might be blocked on a wait queue.
1241 * We have a little helper inline function which does additional work after
1242 * the thread has been enqueued, including dealing with preemption and
1243 * setting need_lwkt_resched() (which prevents the kernel from returning
1244 * to userland until it has processed higher priority threads).
1246 * It is possible for this routine to be called after a failed _enqueue
1247 * (due to the target thread migrating, sleeping, or otherwise blocked).
1248 * We have to check that the thread is actually on the run queue!
1252 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1254 if (ntd->td_flags & TDF_RUNQ) {
1255 if (ntd->td_preemptable) {
1256 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1263 _lwkt_schedule(thread_t td)
1265 globaldata_t mygd = mycpu;
1267 KASSERT(td != &td->td_gd->gd_idlethread,
1268 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1269 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1270 crit_enter_gd(mygd);
1271 KKASSERT(td->td_lwp == NULL ||
1272 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1274 if (td == mygd->gd_curthread) {
1278 * If we own the thread, there is no race (since we are in a
1279 * critical section). If we do not own the thread there might
1280 * be a race but the target cpu will deal with it.
1283 if (td->td_gd == mygd) {
1285 _lwkt_schedule_post(mygd, td, 1);
1287 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1291 _lwkt_schedule_post(mygd, td, 1);
1298 lwkt_schedule(thread_t td)
1304 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1312 * When scheduled remotely if frame != NULL the IPIQ is being
1313 * run via doreti or an interrupt then preemption can be allowed.
1315 * To allow preemption we have to drop the critical section so only
1316 * one is present in _lwkt_schedule_post.
1319 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1321 thread_t td = curthread;
1324 if (frame && ntd->td_preemptable) {
1325 crit_exit_noyield(td);
1326 _lwkt_schedule(ntd);
1327 crit_enter_quick(td);
1329 _lwkt_schedule(ntd);
1334 * Thread migration using a 'Pull' method. The thread may or may not be
1335 * the current thread. It MUST be descheduled and in a stable state.
1336 * lwkt_giveaway() must be called on the cpu owning the thread.
1338 * At any point after lwkt_giveaway() is called, the target cpu may
1339 * 'pull' the thread by calling lwkt_acquire().
1341 * We have to make sure the thread is not sitting on a per-cpu tsleep
1342 * queue or it will blow up when it moves to another cpu.
1344 * MPSAFE - must be called under very specific conditions.
1347 lwkt_giveaway(thread_t td)
1349 globaldata_t gd = mycpu;
1352 if (td->td_flags & TDF_TSLEEPQ)
1354 KKASSERT(td->td_gd == gd);
1355 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1356 td->td_flags |= TDF_MIGRATING;
1361 lwkt_acquire(thread_t td)
1365 int retry = 10000000;
1367 KKASSERT(td->td_flags & TDF_MIGRATING);
1372 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1373 crit_enter_gd(mygd);
1374 DEBUG_PUSH_INFO("lwkt_acquire");
1375 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1377 lwkt_process_ipiq();
1381 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1389 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1390 td->td_flags &= ~TDF_MIGRATING;
1393 crit_enter_gd(mygd);
1394 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1395 td->td_flags &= ~TDF_MIGRATING;
1403 * Generic deschedule. Descheduling threads other then your own should be
1404 * done only in carefully controlled circumstances. Descheduling is
1407 * This function may block if the cpu has run out of messages.
1410 lwkt_deschedule(thread_t td)
1414 if (td == curthread) {
1417 if (td->td_gd == mycpu) {
1420 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1430 * Set the target thread's priority. This routine does not automatically
1431 * switch to a higher priority thread, LWKT threads are not designed for
1432 * continuous priority changes. Yield if you want to switch.
1435 lwkt_setpri(thread_t td, int pri)
1437 if (td->td_pri != pri) {
1440 if (td->td_flags & TDF_RUNQ) {
1441 KKASSERT(td->td_gd == mycpu);
1453 * Set the initial priority for a thread prior to it being scheduled for
1454 * the first time. The thread MUST NOT be scheduled before or during
1455 * this call. The thread may be assigned to a cpu other then the current
1458 * Typically used after a thread has been created with TDF_STOPPREQ,
1459 * and before the thread is initially scheduled.
1462 lwkt_setpri_initial(thread_t td, int pri)
1465 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1470 lwkt_setpri_self(int pri)
1472 thread_t td = curthread;
1474 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1476 if (td->td_flags & TDF_RUNQ) {
1487 * hz tick scheduler clock for LWKT threads
1490 lwkt_schedulerclock(thread_t td)
1492 globaldata_t gd = td->td_gd;
1495 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1497 * If the current thread is at the head of the runq shift it to the
1498 * end of any equal-priority threads and request a LWKT reschedule
1501 xtd = TAILQ_NEXT(td, td_threadq);
1502 if (xtd && xtd->td_pri == td->td_pri) {
1503 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1504 while (xtd && xtd->td_pri == td->td_pri)
1505 xtd = TAILQ_NEXT(xtd, td_threadq);
1507 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1509 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1510 need_lwkt_resched();
1514 * If we scheduled a thread other than the one at the head of the
1515 * queue always request a reschedule every tick.
1517 need_lwkt_resched();
1522 * Migrate the current thread to the specified cpu.
1524 * This is accomplished by descheduling ourselves from the current cpu
1525 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1526 * 'old' thread wants to migrate after it has been completely switched out
1527 * and will complete the migration.
1529 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1531 * We must be sure to release our current process designation (if a user
1532 * process) before clearing out any tsleepq we are on because the release
1533 * code may re-add us.
1535 * We must be sure to remove ourselves from the current cpu's tsleepq
1536 * before potentially moving to another queue. The thread can be on
1537 * a tsleepq due to a left-over tsleep_interlock().
1541 lwkt_setcpu_self(globaldata_t rgd)
1544 thread_t td = curthread;
1546 if (td->td_gd != rgd) {
1547 crit_enter_quick(td);
1551 if (td->td_flags & TDF_TSLEEPQ)
1555 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1556 * trying to deschedule ourselves and switch away, then deschedule
1557 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1558 * call lwkt_switch() to complete the operation.
1560 td->td_flags |= TDF_MIGRATING;
1561 lwkt_deschedule_self(td);
1562 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1563 td->td_migrate_gd = rgd;
1567 * We are now on the target cpu
1569 KKASSERT(rgd == mycpu);
1570 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1571 crit_exit_quick(td);
1577 lwkt_migratecpu(int cpuid)
1582 rgd = globaldata_find(cpuid);
1583 lwkt_setcpu_self(rgd);
1589 * Remote IPI for cpu migration (called while in a critical section so we
1590 * do not have to enter another one).
1592 * The thread (td) has already been completely descheduled from the
1593 * originating cpu and we can simply assert the case. The thread is
1594 * assigned to the new cpu and enqueued.
1596 * The thread will re-add itself to tdallq when it resumes execution.
1599 lwkt_setcpu_remote(void *arg)
1602 globaldata_t gd = mycpu;
1604 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1607 td->td_flags &= ~TDF_MIGRATING;
1608 KKASSERT(td->td_migrate_gd == NULL);
1609 KKASSERT(td->td_lwp == NULL ||
1610 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1616 lwkt_preempted_proc(void)
1618 thread_t td = curthread;
1619 while (td->td_preempted)
1620 td = td->td_preempted;
1625 * Create a kernel process/thread/whatever. It shares it's address space
1626 * with proc0 - ie: kernel only.
1628 * If the cpu is not specified one will be selected. In the future
1629 * specifying a cpu of -1 will enable kernel thread migration between
1633 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1634 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1639 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1643 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1646 * Set up arg0 for 'ps' etc
1648 __va_start(ap, fmt);
1649 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1653 * Schedule the thread to run
1655 if (td->td_flags & TDF_NOSTART)
1656 td->td_flags &= ~TDF_NOSTART;
1663 * Destroy an LWKT thread. Warning! This function is not called when
1664 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1665 * uses a different reaping mechanism.
1670 thread_t td = curthread;
1675 * Do any cleanup that might block here
1677 if (td->td_flags & TDF_VERBOSE)
1678 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1681 dsched_exit_thread(td);
1684 * Get us into a critical section to interlock gd_freetd and loop
1685 * until we can get it freed.
1687 * We have to cache the current td in gd_freetd because objcache_put()ing
1688 * it would rip it out from under us while our thread is still active.
1690 * We are the current thread so of course our own TDF_RUNNING bit will
1691 * be set, so unlike the lwp reap code we don't wait for it to clear.
1694 crit_enter_quick(td);
1697 tsleep(td, 0, "tdreap", 1);
1700 if ((std = gd->gd_freetd) != NULL) {
1701 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1702 gd->gd_freetd = NULL;
1703 objcache_put(thread_cache, std);
1710 * Remove thread resources from kernel lists and deschedule us for
1711 * the last time. We cannot block after this point or we may end
1712 * up with a stale td on the tsleepq.
1714 * None of this may block, the critical section is the only thing
1715 * protecting tdallq and the only thing preventing new lwkt_hold()
1718 if (td->td_flags & TDF_TSLEEPQ)
1720 lwkt_deschedule_self(td);
1721 lwkt_remove_tdallq(td);
1722 KKASSERT(td->td_refs == 0);
1727 KKASSERT(gd->gd_freetd == NULL);
1728 if (td->td_flags & TDF_ALLOCATED_THREAD)
1734 lwkt_remove_tdallq(thread_t td)
1736 KKASSERT(td->td_gd == mycpu);
1737 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1741 * Code reduction and branch prediction improvements. Call/return
1742 * overhead on modern cpus often degenerates into 0 cycles due to
1743 * the cpu's branch prediction hardware and return pc cache. We
1744 * can take advantage of this by not inlining medium-complexity
1745 * functions and we can also reduce the branch prediction impact
1746 * by collapsing perfectly predictable branches into a single
1747 * procedure instead of duplicating it.
1749 * Is any of this noticeable? Probably not, so I'll take the
1750 * smaller code size.
1753 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1755 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1761 thread_t td = curthread;
1762 int lcrit = td->td_critcount;
1764 td->td_critcount = 0;
1765 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1772 * Called from debugger/panic on cpus which have been stopped. We must still
1773 * process the IPIQ while stopped, even if we were stopped while in a critical
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;
1787 lwkt_process_ipiq();
1790 lwkt_process_ipiq();