2 * Copyright (c) 2003-2010 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",
80 sizeof(int) + sizeof(struct thread *));
81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
90 static int panic_on_cscount = 0;
92 static __int64_t switch_count = 0;
93 static __int64_t preempt_hit = 0;
94 static __int64_t preempt_miss = 0;
95 static __int64_t preempt_weird = 0;
96 static __int64_t token_contention_count __debugvar = 0;
97 static int lwkt_use_spin_port;
98 static struct objcache *thread_cache;
101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
103 static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
105 extern void cpu_heavy_restore(void);
106 extern void cpu_lwkt_restore(void);
107 extern void cpu_kthread_restore(void);
108 extern void cpu_idle_restore(void);
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
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 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
132 static int fairq_enable = 1;
133 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0,
134 "Turn on fairq priority accumulators");
135 static int user_pri_sched = 0;
136 SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0,
138 static int preempt_enable = 1;
139 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0,
140 "Enable preemption");
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_fairq_total_pri -= td->td_pri;
162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
163 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
170 * NOTE: There are a limited number of lwkt threads runnable since user
171 * processes only schedule one at a time per cpu.
175 _lwkt_enqueue(thread_t td)
179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
180 struct globaldata *gd = td->td_gd;
182 td->td_flags |= TDF_RUNQ;
183 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
185 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
186 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
188 while (xtd && xtd->td_pri > td->td_pri)
189 xtd = TAILQ_NEXT(xtd, td_threadq);
191 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
195 gd->gd_fairq_total_pri += td->td_pri;
200 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
202 struct thread *td = (struct thread *)obj;
204 td->td_kstack = NULL;
205 td->td_kstack_size = 0;
206 td->td_flags = TDF_ALLOCATED_THREAD;
211 _lwkt_thread_dtor(void *obj, void *privdata)
213 struct thread *td = (struct thread *)obj;
215 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
216 ("_lwkt_thread_dtor: not allocated from objcache"));
217 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
218 td->td_kstack_size > 0,
219 ("_lwkt_thread_dtor: corrupted stack"));
220 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
224 * Initialize the lwkt s/system.
229 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
230 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
231 NULL, CACHE_NTHREADS/2,
232 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
236 * Schedule a thread to run. As the current thread we can always safely
237 * schedule ourselves, and a shortcut procedure is provided for that
240 * (non-blocking, self contained on a per cpu basis)
243 lwkt_schedule_self(thread_t td)
245 crit_enter_quick(td);
246 KASSERT(td != &td->td_gd->gd_idlethread,
247 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
248 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
254 * Deschedule a thread.
256 * (non-blocking, self contained on a per cpu basis)
259 lwkt_deschedule_self(thread_t td)
261 crit_enter_quick(td);
267 * LWKTs operate on a per-cpu basis
269 * WARNING! Called from early boot, 'mycpu' may not work yet.
272 lwkt_gdinit(struct globaldata *gd)
274 TAILQ_INIT(&gd->gd_tdrunq);
275 TAILQ_INIT(&gd->gd_tdallq);
279 * Create a new thread. The thread must be associated with a process context
280 * or LWKT start address before it can be scheduled. If the target cpu is
281 * -1 the thread will be created on the current cpu.
283 * If you intend to create a thread without a process context this function
284 * does everything except load the startup and switcher function.
287 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
289 globaldata_t gd = mycpu;
293 * If static thread storage is not supplied allocate a thread. Reuse
294 * a cached free thread if possible. gd_freetd is used to keep an exiting
295 * thread intact through the exit.
299 if ((td = gd->gd_freetd) != NULL) {
300 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
302 gd->gd_freetd = NULL;
304 td = objcache_get(thread_cache, M_WAITOK);
305 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
309 KASSERT((td->td_flags &
310 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
311 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
312 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
316 * Try to reuse cached stack.
318 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
319 if (flags & TDF_ALLOCATED_STACK) {
320 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
325 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
326 flags |= TDF_ALLOCATED_STACK;
329 lwkt_init_thread(td, stack, stksize, flags, gd);
331 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
336 * Initialize a preexisting thread structure. This function is used by
337 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
339 * All threads start out in a critical section at a priority of
340 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
341 * appropriate. This function may send an IPI message when the
342 * requested cpu is not the current cpu and consequently gd_tdallq may
343 * not be initialized synchronously from the point of view of the originating
346 * NOTE! we have to be careful in regards to creating threads for other cpus
347 * if SMP has not yet been activated.
352 lwkt_init_thread_remote(void *arg)
357 * Protected by critical section held by IPI dispatch
359 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
365 * lwkt core thread structural initialization.
367 * NOTE: All threads are initialized as mpsafe threads.
370 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
371 struct globaldata *gd)
373 globaldata_t mygd = mycpu;
375 bzero(td, sizeof(struct thread));
376 td->td_kstack = stack;
377 td->td_kstack_size = stksize;
378 td->td_flags = flags;
380 td->td_pri = TDPRI_KERN_DAEMON;
381 td->td_critcount = 1;
382 td->td_toks_stop = &td->td_toks_base;
383 if (lwkt_use_spin_port)
384 lwkt_initport_spin(&td->td_msgport);
386 lwkt_initport_thread(&td->td_msgport, td);
387 pmap_init_thread(td);
390 * Normally initializing a thread for a remote cpu requires sending an
391 * IPI. However, the idlethread is setup before the other cpus are
392 * activated so we have to treat it as a special case. XXX manipulation
393 * of gd_tdallq requires the BGL.
395 if (gd == mygd || td == &gd->gd_idlethread) {
397 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
400 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
404 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
408 dsched_new_thread(td);
412 lwkt_set_comm(thread_t td, const char *ctl, ...)
417 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
419 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
423 lwkt_hold(thread_t td)
429 lwkt_rele(thread_t td)
431 KKASSERT(td->td_refs > 0);
436 lwkt_wait_free(thread_t td)
439 tsleep(td, 0, "tdreap", hz);
443 lwkt_free_thread(thread_t td)
445 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
446 if (td->td_flags & TDF_ALLOCATED_THREAD) {
447 objcache_put(thread_cache, td);
448 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
449 /* client-allocated struct with internally allocated stack */
450 KASSERT(td->td_kstack && td->td_kstack_size > 0,
451 ("lwkt_free_thread: corrupted stack"));
452 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
453 td->td_kstack = NULL;
454 td->td_kstack_size = 0;
456 KTR_LOG(ctxsw_deadtd, td);
461 * Switch to the next runnable lwkt. If no LWKTs are runnable then
462 * switch to the idlethread. Switching must occur within a critical
463 * section to avoid races with the scheduling queue.
465 * We always have full control over our cpu's run queue. Other cpus
466 * that wish to manipulate our queue must use the cpu_*msg() calls to
467 * talk to our cpu, so a critical section is all that is needed and
468 * the result is very, very fast thread switching.
470 * The LWKT scheduler uses a fixed priority model and round-robins at
471 * each priority level. User process scheduling is a totally
472 * different beast and LWKT priorities should not be confused with
473 * user process priorities.
475 * The MP lock may be out of sync with the thread's td_mpcount + td_xpcount.
476 * lwkt_switch() cleans it up.
478 * Note that the td_switch() function cannot do anything that requires
479 * the MP lock since the MP lock will have already been setup for
480 * the target thread (not the current thread). It's nice to have a scheduler
481 * that does not need the MP lock to work because it allows us to do some
482 * really cool high-performance MP lock optimizations.
484 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
485 * is not called by the current thread in the preemption case, only when
486 * the preempting thread blocks (in order to return to the original thread).
491 globaldata_t gd = mycpu;
492 thread_t td = gd->gd_curthread;
501 const char *lmsg; /* diagnostic - 'systat -pv 1' */
505 * Switching from within a 'fast' (non thread switched) interrupt or IPI
506 * is illegal. However, we may have to do it anyway if we hit a fatal
507 * kernel trap or we have paniced.
509 * If this case occurs save and restore the interrupt nesting level.
511 if (gd->gd_intr_nesting_level) {
515 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
516 panic("lwkt_switch: Attempt to switch from a "
517 "a fast interrupt, ipi, or hard code section, "
521 savegdnest = gd->gd_intr_nesting_level;
522 savegdtrap = gd->gd_trap_nesting_level;
523 gd->gd_intr_nesting_level = 0;
524 gd->gd_trap_nesting_level = 0;
525 if ((td->td_flags & TDF_PANICWARN) == 0) {
526 td->td_flags |= TDF_PANICWARN;
527 kprintf("Warning: thread switch from interrupt, IPI, "
528 "or hard code section.\n"
529 "thread %p (%s)\n", td, td->td_comm);
533 gd->gd_intr_nesting_level = savegdnest;
534 gd->gd_trap_nesting_level = savegdtrap;
540 * Passive release (used to transition from user to kernel mode
541 * when we block or switch rather then when we enter the kernel).
542 * This function is NOT called if we are switching into a preemption
543 * or returning from a preemption. Typically this causes us to lose
544 * our current process designation (if we have one) and become a true
545 * LWKT thread, and may also hand the current process designation to
546 * another process and schedule thread.
552 if (TD_TOKS_HELD(td))
553 lwkt_relalltokens(td);
556 * We had better not be holding any spin locks, but don't get into an
557 * endless panic loop.
559 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
560 ("lwkt_switch: still holding %d exclusive spinlocks!",
561 gd->gd_spinlocks_wr));
566 * td_mpcount + td_xpcount cannot be used to determine if we currently
567 * hold the MP lock because get_mplock() will increment it prior to
568 * attempting to get the lock, and switch out if it can't. Our
569 * ownership of the actual lock will remain stable while we are
570 * in a critical section, and once we actually acquire the underlying
571 * lock as long as the count is greater than 0.
573 mpheld = MP_LOCK_HELD(gd);
575 if (td->td_cscount) {
576 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
578 if (panic_on_cscount)
579 panic("switching while mastering cpusync");
585 * If we had preempted another thread on this cpu, resume the preempted
586 * thread. This occurs transparently, whether the preempted thread
587 * was scheduled or not (it may have been preempted after descheduling
590 * We have to setup the MP lock for the original thread after backing
591 * out the adjustment that was made to curthread when the original
594 if ((ntd = td->td_preempted) != NULL) {
595 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
597 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
598 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
599 td, ntd, td->td_mpcount, ntd->td_mpcount + ntd->td_xpcount);
603 ntd->td_flags |= TDF_PREEMPT_DONE;
606 * The interrupt may have woken a thread up, we need to properly
607 * set the reschedule flag if the originally interrupted thread is
608 * at a lower priority.
610 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
611 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
614 /* YYY release mp lock on switchback if original doesn't need it */
615 goto havethread_preempted;
619 * Implement round-robin fairq with priority insertion. The priority
620 * insertion is handled by _lwkt_enqueue()
622 * We have to adjust the MP lock for the target thread. If we
623 * need the MP lock and cannot obtain it we try to locate a
624 * thread that does not need the MP lock. If we cannot, we spin
627 * A similar issue exists for the tokens held by the target thread.
628 * If we cannot obtain ownership of the tokens we cannot immediately
629 * schedule the thread.
632 clear_lwkt_resched();
634 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
637 * Hotpath if we can get all necessary resources.
639 * If nothing is runnable switch to the idle thread
642 ntd = &gd->gd_idlethread;
643 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
644 ntd->td_flags |= TDF_IDLE_NOHLT;
646 KKASSERT(ntd->td_xpcount == 0);
647 if (ntd->td_mpcount) {
648 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
649 panic("Idle thread %p was holding the BGL!", ntd);
651 set_cpu_contention_mask(gd);
652 handle_cpu_contention_mask();
654 mpheld = MP_LOCK_HELD(gd);
659 clr_cpu_contention_mask(gd);
661 cpu_time.cp_msg[0] = 0;
662 cpu_time.cp_stallpc = 0;
669 * NOTE: For UP there is no mplock and lwkt_getalltokens()
672 if (ntd->td_fairq_accum >= 0 &&
674 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
675 mpheld || cpu_try_mplock()) &&
677 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
680 clr_cpu_contention_mask(gd);
689 if (ntd->td_fairq_accum >= 0)
690 set_cpu_contention_mask(gd);
691 /* Reload mpheld (it become stale after mplock/token ops) */
692 mpheld = MP_LOCK_HELD(gd);
693 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
695 laddr = ntd->td_mplock_stallpc;
700 * Coldpath - unable to schedule ntd, continue looking for threads
701 * to schedule. This is only allowed of the (presumably) kernel
702 * thread exhausted its fair share. A kernel thread stuck on
703 * resources does not currently allow a user thread to get in
707 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
708 (ntd->td_fairq_accum < 0));
716 * If the fair-share scheduler ran out ntd gets moved to the
717 * end and its accumulator will be bumped, if it didn't we
718 * maintain the same queue position.
720 * nlast keeps track of the last element prior to any moves.
722 if (ntd->td_fairq_accum < 0) {
723 lwkt_fairq_accumulate(gd, ntd);
729 xtd = TAILQ_NEXT(ntd, td_threadq);
730 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
731 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
734 * Set terminal element (nlast)
743 ntd = TAILQ_NEXT(ntd, td_threadq);
747 * If we exhausted the run list switch to the idle thread.
748 * Since one or more threads had resource acquisition issues
749 * we do not allow the idle thread to halt.
751 * NOTE: nlast can be NULL.
755 ntd = &gd->gd_idlethread;
756 ntd->td_flags |= TDF_IDLE_NOHLT;
758 KKASSERT(ntd->td_xpcount == 0);
759 if (ntd->td_mpcount) {
760 mpheld = MP_LOCK_HELD(gd);
761 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
762 panic("Idle thread %p was holding the BGL!", ntd);
764 set_cpu_contention_mask(gd);
765 handle_cpu_contention_mask();
767 mpheld = MP_LOCK_HELD(gd);
769 break; /* try again from the top, almost */
775 * If fairq accumulations occured we do not schedule the
776 * idle thread. This will cause us to try again from
780 break; /* try again from the top, almost */
782 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
783 cpu_time.cp_stallpc = (uintptr_t)laddr;
788 * Try to switch to this thread.
790 * NOTE: For UP there is no mplock and lwkt_getalltokens()
793 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
794 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
796 (ntd->td_mpcount + ntd->td_xpcount == 0 ||
797 mpheld || cpu_try_mplock()) &&
799 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
802 clr_cpu_contention_mask(gd);
808 * Thread was runnable but we were unable to get the required
809 * resources (tokens and/or mplock).
812 ntd->td_wmesg = lmsg;
813 if (ntd->td_fairq_accum >= 0)
814 set_cpu_contention_mask(gd);
816 * Reload mpheld (it become stale after mplock/token ops).
818 mpheld = MP_LOCK_HELD(gd);
819 if (ntd->td_mpcount + ntd->td_xpcount && mpheld == 0) {
821 laddr = ntd->td_mplock_stallpc;
823 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
829 * All threads exhausted but we can loop due to a negative
832 * While we are looping in the scheduler be sure to service
833 * any interrupts which were made pending due to our critical
834 * section, otherwise we could livelock (e.g.) IPIs.
836 * NOTE: splz can enter and exit the mplock so mpheld is
837 * stale after this call.
843 * Our mplock can be cached and cause other cpus to livelock
844 * if we loop due to e.g. not being able to acquire tokens.
846 if (MP_LOCK_HELD(gd))
847 cpu_rel_mplock(gd->gd_cpuid);
853 * Do the actual switch. WARNING: mpheld is stale here.
855 * We must always decrement td_fairq_accum on non-idle threads just
856 * in case a thread never gets a tick due to being in a continuous
857 * critical section. The page-zeroing code does that.
859 * If the thread we came up with is a higher or equal priority verses
860 * the thread at the head of the queue we move our thread to the
861 * front. This way we can always check the front of the queue.
864 ++gd->gd_cnt.v_swtch;
865 --ntd->td_fairq_accum;
866 ntd->td_wmesg = NULL;
867 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
868 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
869 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
870 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
872 havethread_preempted:
875 * If the new target does not need the MP lock and we are holding it,
876 * release the MP lock. If the new target requires the MP lock we have
877 * already acquired it for the target.
879 * WARNING: mpheld is stale here.
882 KASSERT(ntd->td_critcount,
883 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
885 if (ntd->td_mpcount + ntd->td_xpcount == 0 ) {
886 if (MP_LOCK_HELD(gd))
887 cpu_rel_mplock(gd->gd_cpuid);
889 ASSERT_MP_LOCK_HELD(ntd);
894 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
897 /* NOTE: current cpu may have changed after switch */
902 * Request that the target thread preempt the current thread. Preemption
903 * only works under a specific set of conditions:
905 * - We are not preempting ourselves
906 * - The target thread is owned by the current cpu
907 * - We are not currently being preempted
908 * - The target is not currently being preempted
909 * - We are not holding any spin locks
910 * - The target thread is not holding any tokens
911 * - We are able to satisfy the target's MP lock requirements (if any).
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 * XXX at the moment we run the target thread in a critical section during
920 * the preemption in order to prevent the target from taking interrupts
921 * that *WE* can't. Preemption is strictly limited to interrupt threads
922 * and interrupt-like threads, outside of a critical section, and the
923 * preempted source thread will be resumed the instant the target blocks
924 * whether or not the source is scheduled (i.e. preemption is supposed to
925 * be as transparent as possible).
927 * The target thread inherits our MP count (added to its own) for the
928 * duration of the preemption in order to preserve the atomicy of the
929 * MP lock during the preemption. Therefore, any preempting targets must be
930 * careful in regards to MP assertions. Note that the MP count may be
931 * out of sync with the physical mp_lock, but we do not have to preserve
932 * the original ownership of the lock if it was out of synch (that is, we
933 * can leave it synchronized on return).
936 lwkt_preempt(thread_t ntd, int critcount)
938 struct globaldata *gd = mycpu;
944 int save_gd_intr_nesting_level;
947 * The caller has put us in a critical section. We can only preempt
948 * if the caller of the caller was not in a critical section (basically
949 * a local interrupt), as determined by the 'critcount' parameter. We
950 * also can't preempt if the caller is holding any spinlocks (even if
951 * he isn't in a critical section). This also handles the tokens test.
953 * YYY The target thread must be in a critical section (else it must
954 * inherit our critical section? I dunno yet).
956 * Set need_lwkt_resched() unconditionally for now YYY.
958 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
960 if (preempt_enable == 0) {
965 td = gd->gd_curthread;
966 if (ntd->td_pri <= td->td_pri) {
970 if (td->td_critcount > critcount) {
976 if (ntd->td_gd != gd) {
983 * We don't have to check spinlocks here as they will also bump
986 * Do not try to preempt if the target thread is holding any tokens.
987 * We could try to acquire the tokens but this case is so rare there
988 * is no need to support it.
990 KKASSERT(gd->gd_spinlocks_wr == 0);
992 if (TD_TOKS_HELD(ntd)) {
997 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1002 if (ntd->td_preempted) {
1004 need_lwkt_resched();
1009 * NOTE: An interrupt might have occured just as we were transitioning
1010 * to or from the MP lock. In this case td_mpcount will be pre-disposed
1011 * (non-zero) but not actually synchronized with the mp_lock itself.
1012 * We can use it to imply an MP lock requirement for the preemption but
1013 * we cannot use it to test whether we hold the MP lock or not.
1015 savecnt = td->td_mpcount;
1016 mpheld = MP_LOCK_HELD(gd);
1017 ntd->td_xpcount = td->td_mpcount + td->td_xpcount;
1018 if (mpheld == 0 && ntd->td_mpcount + ntd->td_xpcount && !cpu_try_mplock()) {
1019 ntd->td_xpcount = 0;
1021 need_lwkt_resched();
1027 * Since we are able to preempt the current thread, there is no need to
1028 * call need_lwkt_resched().
1030 * We must temporarily clear gd_intr_nesting_level around the switch
1031 * since switchouts from the target thread are allowed (they will just
1032 * return to our thread), and since the target thread has its own stack.
1035 ntd->td_preempted = td;
1036 td->td_flags |= TDF_PREEMPT_LOCK;
1037 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1038 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1039 gd->gd_intr_nesting_level = 0;
1041 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1043 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1045 KKASSERT(savecnt == td->td_mpcount);
1046 mpheld = MP_LOCK_HELD(gd);
1047 if (mpheld && td->td_mpcount == 0)
1048 cpu_rel_mplock(gd->gd_cpuid);
1049 else if (mpheld == 0 && td->td_mpcount + td->td_xpcount)
1050 panic("lwkt_preempt(): MP lock was not held through");
1052 ntd->td_preempted = NULL;
1053 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1057 * Conditionally call splz() if gd_reqflags indicates work is pending.
1058 * This will work inside a critical section but not inside a hard code
1061 * (self contained on a per cpu basis)
1066 globaldata_t gd = mycpu;
1067 thread_t td = gd->gd_curthread;
1069 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1070 gd->gd_intr_nesting_level == 0 &&
1071 td->td_nest_count < 2)
1078 * This version is integrated into crit_exit, reqflags has already
1079 * been tested but td_critcount has not.
1081 * We only want to execute the splz() on the 1->0 transition of
1082 * critcount and not in a hard code section or if too deeply nested.
1085 lwkt_maybe_splz(thread_t td)
1087 globaldata_t gd = td->td_gd;
1089 if (td->td_critcount == 0 &&
1090 gd->gd_intr_nesting_level == 0 &&
1091 td->td_nest_count < 2)
1098 * This function is used to negotiate a passive release of the current
1099 * process/lwp designation with the user scheduler, allowing the user
1100 * scheduler to schedule another user thread. The related kernel thread
1101 * (curthread) continues running in the released state.
1104 lwkt_passive_release(struct thread *td)
1106 struct lwp *lp = td->td_lwp;
1108 td->td_release = NULL;
1109 lwkt_setpri_self(TDPRI_KERN_USER);
1110 lp->lwp_proc->p_usched->release_curproc(lp);
1115 * This implements a normal yield. This routine is virtually a nop if
1116 * there is nothing to yield to but it will always run any pending interrupts
1117 * if called from a critical section.
1119 * This yield is designed for kernel threads without a user context.
1121 * (self contained on a per cpu basis)
1126 globaldata_t gd = mycpu;
1127 thread_t td = gd->gd_curthread;
1130 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1132 if (td->td_fairq_accum < 0) {
1133 lwkt_schedule_self(curthread);
1136 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1137 if (xtd && xtd->td_pri > td->td_pri) {
1138 lwkt_schedule_self(curthread);
1145 * This yield is designed for kernel threads with a user context.
1147 * The kernel acting on behalf of the user is potentially cpu-bound,
1148 * this function will efficiently allow other threads to run and also
1149 * switch to other processes by releasing.
1151 * The lwkt_user_yield() function is designed to have very low overhead
1152 * if no yield is determined to be needed.
1155 lwkt_user_yield(void)
1157 globaldata_t gd = mycpu;
1158 thread_t td = gd->gd_curthread;
1161 * Always run any pending interrupts in case we are in a critical
1164 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1169 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1170 * kernel can prevent other cpus from servicing interrupt threads
1171 * which still require the MP lock (which is a lot of them). This
1172 * has a chaining effect since if the interrupt is blocked, so is
1173 * the event, so normal scheduling will not pick up on the problem.
1175 if (cpu_contention_mask && td->td_mpcount + td->td_xpcount) {
1181 * Switch (which forces a release) if another kernel thread needs
1182 * the cpu, if userland wants us to resched, or if our kernel
1183 * quantum has run out.
1185 if (lwkt_resched_wanted() ||
1186 user_resched_wanted() ||
1187 td->td_fairq_accum < 0)
1194 * Reacquire the current process if we are released.
1196 * XXX not implemented atm. The kernel may be holding locks and such,
1197 * so we want the thread to continue to receive cpu.
1199 if (td->td_release == NULL && lp) {
1200 lp->lwp_proc->p_usched->acquire_curproc(lp);
1201 td->td_release = lwkt_passive_release;
1202 lwkt_setpri_self(TDPRI_USER_NORM);
1208 * Generic schedule. Possibly schedule threads belonging to other cpus and
1209 * deal with threads that might be blocked on a wait queue.
1211 * We have a little helper inline function which does additional work after
1212 * the thread has been enqueued, including dealing with preemption and
1213 * setting need_lwkt_resched() (which prevents the kernel from returning
1214 * to userland until it has processed higher priority threads).
1216 * It is possible for this routine to be called after a failed _enqueue
1217 * (due to the target thread migrating, sleeping, or otherwise blocked).
1218 * We have to check that the thread is actually on the run queue!
1220 * reschedok is an optimized constant propagated from lwkt_schedule() or
1221 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1222 * reschedule to be requested if the target thread has a higher priority.
1223 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1224 * be 0, prevented undesired reschedules.
1228 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1232 if (ntd->td_flags & TDF_RUNQ) {
1233 if (ntd->td_preemptable && reschedok) {
1234 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1235 } else if (reschedok) {
1237 if (ntd->td_pri > otd->td_pri)
1238 need_lwkt_resched();
1242 * Give the thread a little fair share scheduler bump if it
1243 * has been asleep for a while. This is primarily to avoid
1244 * a degenerate case for interrupt threads where accumulator
1245 * crosses into negative territory unnecessarily.
1247 if (ntd->td_fairq_lticks != ticks) {
1248 ntd->td_fairq_lticks = ticks;
1249 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1250 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1251 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1258 _lwkt_schedule(thread_t td, int reschedok)
1260 globaldata_t mygd = mycpu;
1262 KASSERT(td != &td->td_gd->gd_idlethread,
1263 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1264 crit_enter_gd(mygd);
1265 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1266 if (td == mygd->gd_curthread) {
1270 * If we own the thread, there is no race (since we are in a
1271 * critical section). If we do not own the thread there might
1272 * be a race but the target cpu will deal with it.
1275 if (td->td_gd == mygd) {
1277 _lwkt_schedule_post(mygd, td, 1, reschedok);
1279 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1283 _lwkt_schedule_post(mygd, td, 1, reschedok);
1290 lwkt_schedule(thread_t td)
1292 _lwkt_schedule(td, 1);
1296 lwkt_schedule_noresched(thread_t td)
1298 _lwkt_schedule(td, 0);
1304 * When scheduled remotely if frame != NULL the IPIQ is being
1305 * run via doreti or an interrupt then preemption can be allowed.
1307 * To allow preemption we have to drop the critical section so only
1308 * one is present in _lwkt_schedule_post.
1311 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1313 thread_t td = curthread;
1316 if (frame && ntd->td_preemptable) {
1317 crit_exit_noyield(td);
1318 _lwkt_schedule(ntd, 1);
1319 crit_enter_quick(td);
1321 _lwkt_schedule(ntd, 1);
1326 * Thread migration using a 'Pull' method. The thread may or may not be
1327 * the current thread. It MUST be descheduled and in a stable state.
1328 * lwkt_giveaway() must be called on the cpu owning the thread.
1330 * At any point after lwkt_giveaway() is called, the target cpu may
1331 * 'pull' the thread by calling lwkt_acquire().
1333 * We have to make sure the thread is not sitting on a per-cpu tsleep
1334 * queue or it will blow up when it moves to another cpu.
1336 * MPSAFE - must be called under very specific conditions.
1339 lwkt_giveaway(thread_t td)
1341 globaldata_t gd = mycpu;
1344 if (td->td_flags & TDF_TSLEEPQ)
1346 KKASSERT(td->td_gd == gd);
1347 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1348 td->td_flags |= TDF_MIGRATING;
1353 lwkt_acquire(thread_t td)
1358 KKASSERT(td->td_flags & TDF_MIGRATING);
1363 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1364 crit_enter_gd(mygd);
1365 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1367 lwkt_process_ipiq();
1373 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1374 td->td_flags &= ~TDF_MIGRATING;
1377 crit_enter_gd(mygd);
1378 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1379 td->td_flags &= ~TDF_MIGRATING;
1387 * Generic deschedule. Descheduling threads other then your own should be
1388 * done only in carefully controlled circumstances. Descheduling is
1391 * This function may block if the cpu has run out of messages.
1394 lwkt_deschedule(thread_t td)
1398 if (td == curthread) {
1401 if (td->td_gd == mycpu) {
1404 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1414 * Set the target thread's priority. This routine does not automatically
1415 * switch to a higher priority thread, LWKT threads are not designed for
1416 * continuous priority changes. Yield if you want to switch.
1419 lwkt_setpri(thread_t td, int pri)
1421 KKASSERT(td->td_gd == mycpu);
1422 if (td->td_pri != pri) {
1425 if (td->td_flags & TDF_RUNQ) {
1437 * Set the initial priority for a thread prior to it being scheduled for
1438 * the first time. The thread MUST NOT be scheduled before or during
1439 * this call. The thread may be assigned to a cpu other then the current
1442 * Typically used after a thread has been created with TDF_STOPPREQ,
1443 * and before the thread is initially scheduled.
1446 lwkt_setpri_initial(thread_t td, int pri)
1449 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1454 lwkt_setpri_self(int pri)
1456 thread_t td = curthread;
1458 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1460 if (td->td_flags & TDF_RUNQ) {
1471 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1473 * Example: two competing threads, same priority N. decrement by (2*N)
1474 * increment by N*8, each thread will get 4 ticks.
1477 lwkt_fairq_schedulerclock(thread_t td)
1481 if (td != &td->td_gd->gd_idlethread) {
1482 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1483 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1484 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1485 if (td->td_fairq_accum < 0)
1486 need_lwkt_resched();
1487 td->td_fairq_lticks = ticks;
1489 td = td->td_preempted;
1495 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1497 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1498 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1499 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1503 * Migrate the current thread to the specified cpu.
1505 * This is accomplished by descheduling ourselves from the current cpu,
1506 * moving our thread to the tdallq of the target cpu, IPI messaging the
1507 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1508 * races while the thread is being migrated.
1510 * We must be sure to remove ourselves from the current cpu's tsleepq
1511 * before potentially moving to another queue. The thread can be on
1512 * a tsleepq due to a left-over tsleep_interlock().
1515 static void lwkt_setcpu_remote(void *arg);
1519 lwkt_setcpu_self(globaldata_t rgd)
1522 thread_t td = curthread;
1524 if (td->td_gd != rgd) {
1525 crit_enter_quick(td);
1526 if (td->td_flags & TDF_TSLEEPQ)
1528 td->td_flags |= TDF_MIGRATING;
1529 lwkt_deschedule_self(td);
1530 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1531 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1533 /* we are now on the target cpu */
1534 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1535 crit_exit_quick(td);
1541 lwkt_migratecpu(int cpuid)
1546 rgd = globaldata_find(cpuid);
1547 lwkt_setcpu_self(rgd);
1552 * Remote IPI for cpu migration (called while in a critical section so we
1553 * do not have to enter another one). The thread has already been moved to
1554 * our cpu's allq, but we must wait for the thread to be completely switched
1555 * out on the originating cpu before we schedule it on ours or the stack
1556 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1557 * change to main memory.
1559 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1560 * against wakeups. It is best if this interface is used only when there
1561 * are no pending events that might try to schedule the thread.
1565 lwkt_setcpu_remote(void *arg)
1568 globaldata_t gd = mycpu;
1570 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1572 lwkt_process_ipiq();
1579 td->td_flags &= ~TDF_MIGRATING;
1580 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1586 lwkt_preempted_proc(void)
1588 thread_t td = curthread;
1589 while (td->td_preempted)
1590 td = td->td_preempted;
1595 * Create a kernel process/thread/whatever. It shares it's address space
1596 * with proc0 - ie: kernel only.
1598 * NOTE! By default new threads are created with the MP lock held. A
1599 * thread which does not require the MP lock should release it by calling
1600 * rel_mplock() at the start of the new thread.
1603 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1604 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1609 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1613 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1616 * Set up arg0 for 'ps' etc
1618 __va_start(ap, fmt);
1619 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1623 * Schedule the thread to run
1625 if ((td->td_flags & TDF_STOPREQ) == 0)
1628 td->td_flags &= ~TDF_STOPREQ;
1633 * Destroy an LWKT thread. Warning! This function is not called when
1634 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1635 * uses a different reaping mechanism.
1640 thread_t td = curthread;
1645 * Do any cleanup that might block here
1647 if (td->td_flags & TDF_VERBOSE)
1648 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1651 dsched_exit_thread(td);
1654 * Get us into a critical section to interlock gd_freetd and loop
1655 * until we can get it freed.
1657 * We have to cache the current td in gd_freetd because objcache_put()ing
1658 * it would rip it out from under us while our thread is still active.
1661 crit_enter_quick(td);
1662 while ((std = gd->gd_freetd) != NULL) {
1663 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1664 gd->gd_freetd = NULL;
1665 objcache_put(thread_cache, std);
1669 * Remove thread resources from kernel lists and deschedule us for
1670 * the last time. We cannot block after this point or we may end
1671 * up with a stale td on the tsleepq.
1673 if (td->td_flags & TDF_TSLEEPQ)
1675 lwkt_deschedule_self(td);
1676 lwkt_remove_tdallq(td);
1681 KKASSERT(gd->gd_freetd == NULL);
1682 if (td->td_flags & TDF_ALLOCATED_THREAD)
1688 lwkt_remove_tdallq(thread_t td)
1690 KKASSERT(td->td_gd == mycpu);
1691 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1695 * Code reduction and branch prediction improvements. Call/return
1696 * overhead on modern cpus often degenerates into 0 cycles due to
1697 * the cpu's branch prediction hardware and return pc cache. We
1698 * can take advantage of this by not inlining medium-complexity
1699 * functions and we can also reduce the branch prediction impact
1700 * by collapsing perfectly predictable branches into a single
1701 * procedure instead of duplicating it.
1703 * Is any of this noticeable? Probably not, so I'll take the
1704 * smaller code size.
1707 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1709 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1715 thread_t td = curthread;
1716 int lcrit = td->td_critcount;
1718 td->td_critcount = 0;
1719 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1726 * Called from debugger/panic on cpus which have been stopped. We must still
1727 * process the IPIQ while stopped, even if we were stopped while in a critical
1730 * If we are dumping also try to process any pending interrupts. This may
1731 * or may not work depending on the state of the cpu at the point it was
1735 lwkt_smp_stopped(void)
1737 globaldata_t gd = mycpu;
1741 lwkt_process_ipiq();
1744 lwkt_process_ipiq();