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,
134 &fairq_enable, 0, "Turn on fairq priority accumulators");
135 static int lwkt_spin_loops = 10;
136 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
137 &lwkt_spin_loops, 0, "");
138 static int lwkt_spin_delay = 1;
139 SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
140 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
141 static int lwkt_spin_method = 1;
142 SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
143 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
144 static int preempt_enable = 1;
145 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
146 &preempt_enable, 0, "Enable preemption");
148 static __cachealign int lwkt_cseq_rindex;
149 static __cachealign int lwkt_cseq_windex;
152 * These helper procedures handle the runq, they can only be called from
153 * within a critical section.
155 * WARNING! Prior to SMP being brought up it is possible to enqueue and
156 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
157 * instead of 'mycpu' when referencing the globaldata structure. Once
158 * SMP live enqueuing and dequeueing only occurs on the current cpu.
162 _lwkt_dequeue(thread_t td)
164 if (td->td_flags & TDF_RUNQ) {
165 struct globaldata *gd = td->td_gd;
167 td->td_flags &= ~TDF_RUNQ;
168 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
169 gd->gd_fairq_total_pri -= td->td_pri;
170 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
171 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
178 * NOTE: There are a limited number of lwkt threads runnable since user
179 * processes only schedule one at a time per cpu.
183 _lwkt_enqueue(thread_t td)
187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
188 struct globaldata *gd = td->td_gd;
190 td->td_flags |= TDF_RUNQ;
191 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
193 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
194 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
196 while (xtd && xtd->td_pri > td->td_pri)
197 xtd = TAILQ_NEXT(xtd, td_threadq);
199 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
201 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
203 gd->gd_fairq_total_pri += td->td_pri;
208 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
210 struct thread *td = (struct thread *)obj;
212 td->td_kstack = NULL;
213 td->td_kstack_size = 0;
214 td->td_flags = TDF_ALLOCATED_THREAD;
219 _lwkt_thread_dtor(void *obj, void *privdata)
221 struct thread *td = (struct thread *)obj;
223 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
224 ("_lwkt_thread_dtor: not allocated from objcache"));
225 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
226 td->td_kstack_size > 0,
227 ("_lwkt_thread_dtor: corrupted stack"));
228 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
232 * Initialize the lwkt s/system.
237 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
238 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
239 NULL, CACHE_NTHREADS/2,
240 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
244 * Schedule a thread to run. As the current thread we can always safely
245 * schedule ourselves, and a shortcut procedure is provided for that
248 * (non-blocking, self contained on a per cpu basis)
251 lwkt_schedule_self(thread_t td)
253 crit_enter_quick(td);
254 KASSERT(td != &td->td_gd->gd_idlethread,
255 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
256 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
262 * Deschedule a thread.
264 * (non-blocking, self contained on a per cpu basis)
267 lwkt_deschedule_self(thread_t td)
269 crit_enter_quick(td);
275 * LWKTs operate on a per-cpu basis
277 * WARNING! Called from early boot, 'mycpu' may not work yet.
280 lwkt_gdinit(struct globaldata *gd)
282 TAILQ_INIT(&gd->gd_tdrunq);
283 TAILQ_INIT(&gd->gd_tdallq);
287 * Create a new thread. The thread must be associated with a process context
288 * or LWKT start address before it can be scheduled. If the target cpu is
289 * -1 the thread will be created on the current cpu.
291 * If you intend to create a thread without a process context this function
292 * does everything except load the startup and switcher function.
295 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
297 globaldata_t gd = mycpu;
301 * If static thread storage is not supplied allocate a thread. Reuse
302 * a cached free thread if possible. gd_freetd is used to keep an exiting
303 * thread intact through the exit.
307 if ((td = gd->gd_freetd) != NULL) {
308 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
310 gd->gd_freetd = NULL;
312 td = objcache_get(thread_cache, M_WAITOK);
313 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
317 KASSERT((td->td_flags &
318 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
319 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
320 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
324 * Try to reuse cached stack.
326 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
327 if (flags & TDF_ALLOCATED_STACK) {
328 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
333 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
334 flags |= TDF_ALLOCATED_STACK;
337 lwkt_init_thread(td, stack, stksize, flags, gd);
339 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
344 * Initialize a preexisting thread structure. This function is used by
345 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
347 * All threads start out in a critical section at a priority of
348 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
349 * appropriate. This function may send an IPI message when the
350 * requested cpu is not the current cpu and consequently gd_tdallq may
351 * not be initialized synchronously from the point of view of the originating
354 * NOTE! we have to be careful in regards to creating threads for other cpus
355 * if SMP has not yet been activated.
360 lwkt_init_thread_remote(void *arg)
365 * Protected by critical section held by IPI dispatch
367 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
373 * lwkt core thread structural initialization.
375 * NOTE: All threads are initialized as mpsafe threads.
378 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
379 struct globaldata *gd)
381 globaldata_t mygd = mycpu;
383 bzero(td, sizeof(struct thread));
384 td->td_kstack = stack;
385 td->td_kstack_size = stksize;
386 td->td_flags = flags;
388 td->td_pri = TDPRI_KERN_DAEMON;
389 td->td_critcount = 1;
390 td->td_toks_stop = &td->td_toks_base;
391 if (lwkt_use_spin_port)
392 lwkt_initport_spin(&td->td_msgport);
394 lwkt_initport_thread(&td->td_msgport, td);
395 pmap_init_thread(td);
398 * Normally initializing a thread for a remote cpu requires sending an
399 * IPI. However, the idlethread is setup before the other cpus are
400 * activated so we have to treat it as a special case. XXX manipulation
401 * of gd_tdallq requires the BGL.
403 if (gd == mygd || td == &gd->gd_idlethread) {
405 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
408 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
412 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
416 dsched_new_thread(td);
420 lwkt_set_comm(thread_t td, const char *ctl, ...)
425 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
427 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
431 lwkt_hold(thread_t td)
433 atomic_add_int(&td->td_refs, 1);
437 lwkt_rele(thread_t td)
439 KKASSERT(td->td_refs > 0);
440 atomic_add_int(&td->td_refs, -1);
444 lwkt_wait_free(thread_t td)
447 tsleep(td, 0, "tdreap", hz);
451 lwkt_free_thread(thread_t td)
453 KKASSERT(td->td_refs == 0);
454 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
455 if (td->td_flags & TDF_ALLOCATED_THREAD) {
456 objcache_put(thread_cache, td);
457 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
458 /* client-allocated struct with internally allocated stack */
459 KASSERT(td->td_kstack && td->td_kstack_size > 0,
460 ("lwkt_free_thread: corrupted stack"));
461 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
462 td->td_kstack = NULL;
463 td->td_kstack_size = 0;
465 KTR_LOG(ctxsw_deadtd, td);
470 * Switch to the next runnable lwkt. If no LWKTs are runnable then
471 * switch to the idlethread. Switching must occur within a critical
472 * section to avoid races with the scheduling queue.
474 * We always have full control over our cpu's run queue. Other cpus
475 * that wish to manipulate our queue must use the cpu_*msg() calls to
476 * talk to our cpu, so a critical section is all that is needed and
477 * the result is very, very fast thread switching.
479 * The LWKT scheduler uses a fixed priority model and round-robins at
480 * each priority level. User process scheduling is a totally
481 * different beast and LWKT priorities should not be confused with
482 * user process priorities.
484 * Note that the td_switch() function cannot do anything that requires
485 * the MP lock since the MP lock will have already been setup for
486 * the target thread (not the current thread). It's nice to have a scheduler
487 * that does not need the MP lock to work because it allows us to do some
488 * really cool high-performance MP lock optimizations.
490 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
491 * is not called by the current thread in the preemption case, only when
492 * the preempting thread blocks (in order to return to the original thread).
497 globaldata_t gd = mycpu;
498 thread_t td = gd->gd_curthread;
501 int spinning = lwkt_spin_loops; /* loops before HLTing */
507 * Switching from within a 'fast' (non thread switched) interrupt or IPI
508 * is illegal. However, we may have to do it anyway if we hit a fatal
509 * kernel trap or we have paniced.
511 * If this case occurs save and restore the interrupt nesting level.
513 if (gd->gd_intr_nesting_level) {
517 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
518 panic("lwkt_switch: Attempt to switch from a "
519 "a fast interrupt, ipi, or hard code section, "
523 savegdnest = gd->gd_intr_nesting_level;
524 savegdtrap = gd->gd_trap_nesting_level;
525 gd->gd_intr_nesting_level = 0;
526 gd->gd_trap_nesting_level = 0;
527 if ((td->td_flags & TDF_PANICWARN) == 0) {
528 td->td_flags |= TDF_PANICWARN;
529 kprintf("Warning: thread switch from interrupt, IPI, "
530 "or hard code section.\n"
531 "thread %p (%s)\n", td, td->td_comm);
535 gd->gd_intr_nesting_level = savegdnest;
536 gd->gd_trap_nesting_level = savegdtrap;
542 * Passive release (used to transition from user to kernel mode
543 * when we block or switch rather then when we enter the kernel).
544 * This function is NOT called if we are switching into a preemption
545 * or returning from a preemption. Typically this causes us to lose
546 * our current process designation (if we have one) and become a true
547 * LWKT thread, and may also hand the current process designation to
548 * another process and schedule thread.
554 if (TD_TOKS_HELD(td))
555 lwkt_relalltokens(td);
558 * We had better not be holding any spin locks, but don't get into an
559 * endless panic loop.
561 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
562 ("lwkt_switch: still holding %d exclusive spinlocks!",
563 gd->gd_spinlocks_wr));
568 if (td->td_cscount) {
569 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
571 if (panic_on_cscount)
572 panic("switching while mastering cpusync");
578 * If we had preempted another thread on this cpu, resume the preempted
579 * thread. This occurs transparently, whether the preempted thread
580 * was scheduled or not (it may have been preempted after descheduling
583 * We have to setup the MP lock for the original thread after backing
584 * out the adjustment that was made to curthread when the original
587 if ((ntd = td->td_preempted) != NULL) {
588 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
589 ntd->td_flags |= TDF_PREEMPT_DONE;
592 * The interrupt may have woken a thread up, we need to properly
593 * set the reschedule flag if the originally interrupted thread is
594 * at a lower priority.
596 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
597 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
600 /* YYY release mp lock on switchback if original doesn't need it */
601 goto havethread_preempted;
605 * Implement round-robin fairq with priority insertion. The priority
606 * insertion is handled by _lwkt_enqueue()
608 * We have to adjust the MP lock for the target thread. If we
609 * need the MP lock and cannot obtain it we try to locate a
610 * thread that does not need the MP lock. If we cannot, we spin
613 * A similar issue exists for the tokens held by the target thread.
614 * If we cannot obtain ownership of the tokens we cannot immediately
615 * schedule the thread.
619 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
620 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
624 reqflags = gd->gd_reqflags;
625 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
626 (reqflags & ~RQF_AST_LWKT_RESCHED) |
633 * Hotpath - pull the head of the run queue and attempt to schedule
634 * it. Fairq exhaustion moves the task to the end of the list. If
635 * no threads are runnable we switch to the idle thread.
638 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
642 * Runq is empty, switch to idle and clear RQF_WAKEUP
643 * to allow it to halt.
645 ntd = &gd->gd_idlethread;
647 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
648 ASSERT_NO_TOKENS_HELD(ntd);
650 cpu_time.cp_msg[0] = 0;
651 cpu_time.cp_stallpc = 0;
652 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
656 if (ntd->td_fairq_accum >= 0)
660 lwkt_fairq_accumulate(gd, ntd);
661 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
662 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
666 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
667 * unwanted decontention IPIs.
669 * NOTE: For UP there is no mplock and lwkt_getalltokens()
672 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
676 * Coldpath (SMP only since tokens always succeed on UP)
678 * We had some contention on the thread we wanted to schedule.
679 * What we do now is try to find a thread that we can schedule
680 * in its stead until decontention reschedules on our cpu.
682 * The coldpath scan does NOT rearrange threads in the run list
683 * and it also ignores the accumulator.
685 * We do not immediately schedule a user priority thread, instead
686 * we record it in xtd and continue looking for kernel threads.
687 * A cpu can only have one user priority thread (normally) so just
688 * record the first one.
690 * NOTE: This scan will also include threads whos fairq's were
691 * accumulated in the first loop.
693 ++token_contention_count;
695 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
697 * Try to switch to this thread. If the thread is running at
698 * user priority we clear WAKEUP to allow decontention IPIs
699 * (since this thread is simply running until the one we wanted
700 * decontends), and we make sure that LWKT_RESCHED is not set.
702 * Otherwise for kernel threads we leave WAKEUP set to avoid
703 * unnecessary decontention IPIs.
705 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
712 * Do not let the fairq get too negative. Even though we are
713 * ignoring it atm once the scheduler decontends a very negative
714 * thread will get moved to the end of the queue.
716 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
717 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
718 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
723 * Well fubar, this thread is contended as well, loop
729 * We exhausted the run list but we may have recorded a user
730 * thread to try. We have three choices based on
731 * lwkt.decontention_method.
733 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
734 * IPIs (to interrupt the user process) and test
735 * RQF_AST_LWKT_RESCHED at the same time.
737 * This results in significant decontention IPI traffic but may
738 * be more responsive.
740 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
741 * An automatic LWKT reschedule will occur on the next hardclock
744 * This results in no decontention IPI traffic but may be less
745 * responsive. This is the default.
747 * (2) Refuse to schedule the user process at this time.
749 * This is highly experimental and should not be used under
750 * normal circumstances. This can cause a user process to
751 * get starved out in situations where kernel threads are
752 * fighting each other for tokens.
757 switch(lwkt_spin_method) {
760 reqflags = gd->gd_reqflags;
761 if (atomic_cmpset_int(&gd->gd_reqflags,
763 reqflags & ~RQF_WAKEUP)) {
769 reqflags = gd->gd_reqflags;
775 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
776 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
778 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
779 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
785 * Make sure RQF_WAKEUP is set if we failed to schedule the
786 * user thread to prevent the idle thread from halting.
788 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
792 * We exhausted the run list, meaning that all runnable threads
796 ntd = &gd->gd_idlethread;
798 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
799 ASSERT_NO_TOKENS_HELD(ntd);
800 /* contention case, do not clear contention mask */
804 * Ok, we might want to spin a few times as some tokens are held for
805 * very short periods of time and IPI overhead is 1uS or worse
806 * (meaning it is usually better to spin). Regardless we have to
807 * call splz_check() to be sure to service any interrupts blocked
808 * by our critical section, otherwise we could livelock e.g. IPIs.
810 * The IPI mechanic is really a last resort. In nearly all other
811 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
813 * When we decide not to spin we clear RQF_WAKEUP and switch to
814 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
815 * to halt and decontended tokens will issue an IPI to us. The
816 * idle thread will check for pending reschedules already set
817 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
821 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
827 * When spinning a delay is required both to avoid livelocks from
828 * token order reversals (a thread may be trying to acquire multiple
829 * tokens), and also to reduce cpu cache management traffic.
831 * In order to scale to a large number of CPUs we use a time slot
832 * resequencer to force contending cpus into non-contending
833 * time-slots. The scheduler may still contend with the lock holder
834 * but will not (generally) contend with all the other cpus trying
835 * trying to get the same token.
837 * The resequencer uses a FIFO counter mechanic. The owner of the
838 * rindex at the head of the FIFO is allowed to pull itself off
839 * the FIFO and fetchadd is used to enter into the FIFO. This bit
840 * of code is VERY cache friendly and forces all spinning schedulers
841 * into their own time slots.
843 * This code has been tested to 48-cpus and caps the cache
844 * contention load at ~1uS intervals regardless of the number of
845 * cpus. Scaling beyond 64 cpus might require additional smarts
846 * (such as separate FIFOs for specific token cases).
848 * WARNING! We can't call splz_check() or anything else here as
849 * it could cause a deadlock.
851 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
852 while ((oseq = lwkt_cseq_rindex) != cseq) {
854 #if !defined(_KERNEL_VIRTUAL)
855 if (cpu_mi_feature & CPU_MI_MONITOR) {
856 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
864 cseq = lwkt_spin_delay; /* don't trust the system operator */
871 atomic_add_int(&lwkt_cseq_rindex, 1);
873 /* highest level for(;;) loop */
878 * We must always decrement td_fairq_accum on non-idle threads just
879 * in case a thread never gets a tick due to being in a continuous
880 * critical section. The page-zeroing code does this, for example.
882 * If the thread we came up with is a higher or equal priority verses
883 * the thread at the head of the queue we move our thread to the
884 * front. This way we can always check the front of the queue.
886 * Clear gd_idle_repeat when doing a normal switch to a non-idle
889 ++gd->gd_cnt.v_swtch;
890 --ntd->td_fairq_accum;
891 ntd->td_wmesg = NULL;
892 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
893 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
894 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
895 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
897 gd->gd_idle_repeat = 0;
899 havethread_preempted:
901 * If the new target does not need the MP lock and we are holding it,
902 * release the MP lock. If the new target requires the MP lock we have
903 * already acquired it for the target.
907 KASSERT(ntd->td_critcount,
908 ("priority problem in lwkt_switch %d %d",
909 td->td_critcount, ntd->td_critcount));
913 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
916 /* NOTE: current cpu may have changed after switch */
921 * Request that the target thread preempt the current thread. Preemption
922 * only works under a specific set of conditions:
924 * - We are not preempting ourselves
925 * - The target thread is owned by the current cpu
926 * - We are not currently being preempted
927 * - The target is not currently being preempted
928 * - We are not holding any spin locks
929 * - The target thread is not holding any tokens
930 * - We are able to satisfy the target's MP lock requirements (if any).
932 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
933 * this is called via lwkt_schedule() through the td_preemptable callback.
934 * critcount is the managed critical priority that we should ignore in order
935 * to determine whether preemption is possible (aka usually just the crit
936 * priority of lwkt_schedule() itself).
938 * XXX at the moment we run the target thread in a critical section during
939 * the preemption in order to prevent the target from taking interrupts
940 * that *WE* can't. Preemption is strictly limited to interrupt threads
941 * and interrupt-like threads, outside of a critical section, and the
942 * preempted source thread will be resumed the instant the target blocks
943 * whether or not the source is scheduled (i.e. preemption is supposed to
944 * be as transparent as possible).
947 lwkt_preempt(thread_t ntd, int critcount)
949 struct globaldata *gd = mycpu;
951 int save_gd_intr_nesting_level;
954 * The caller has put us in a critical section. We can only preempt
955 * if the caller of the caller was not in a critical section (basically
956 * a local interrupt), as determined by the 'critcount' parameter. We
957 * also can't preempt if the caller is holding any spinlocks (even if
958 * he isn't in a critical section). This also handles the tokens test.
960 * YYY The target thread must be in a critical section (else it must
961 * inherit our critical section? I dunno yet).
963 * Set need_lwkt_resched() unconditionally for now YYY.
965 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
967 if (preempt_enable == 0) {
972 td = gd->gd_curthread;
973 if (ntd->td_pri <= td->td_pri) {
977 if (td->td_critcount > critcount) {
983 if (ntd->td_gd != gd) {
990 * We don't have to check spinlocks here as they will also bump
993 * Do not try to preempt if the target thread is holding any tokens.
994 * We could try to acquire the tokens but this case is so rare there
995 * is no need to support it.
997 KKASSERT(gd->gd_spinlocks_wr == 0);
999 if (TD_TOKS_HELD(ntd)) {
1001 need_lwkt_resched();
1004 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1006 need_lwkt_resched();
1009 if (ntd->td_preempted) {
1011 need_lwkt_resched();
1016 * Since we are able to preempt the current thread, there is no need to
1017 * call need_lwkt_resched().
1019 * We must temporarily clear gd_intr_nesting_level around the switch
1020 * since switchouts from the target thread are allowed (they will just
1021 * return to our thread), and since the target thread has its own stack.
1024 ntd->td_preempted = td;
1025 td->td_flags |= TDF_PREEMPT_LOCK;
1026 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1027 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1028 gd->gd_intr_nesting_level = 0;
1030 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1032 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1033 ntd->td_preempted = NULL;
1034 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1038 * Conditionally call splz() if gd_reqflags indicates work is pending.
1039 * This will work inside a critical section but not inside a hard code
1042 * (self contained on a per cpu basis)
1047 globaldata_t gd = mycpu;
1048 thread_t td = gd->gd_curthread;
1050 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1051 gd->gd_intr_nesting_level == 0 &&
1052 td->td_nest_count < 2)
1059 * This version is integrated into crit_exit, reqflags has already
1060 * been tested but td_critcount has not.
1062 * We only want to execute the splz() on the 1->0 transition of
1063 * critcount and not in a hard code section or if too deeply nested.
1066 lwkt_maybe_splz(thread_t td)
1068 globaldata_t gd = td->td_gd;
1070 if (td->td_critcount == 0 &&
1071 gd->gd_intr_nesting_level == 0 &&
1072 td->td_nest_count < 2)
1079 * This function is used to negotiate a passive release of the current
1080 * process/lwp designation with the user scheduler, allowing the user
1081 * scheduler to schedule another user thread. The related kernel thread
1082 * (curthread) continues running in the released state.
1085 lwkt_passive_release(struct thread *td)
1087 struct lwp *lp = td->td_lwp;
1089 td->td_release = NULL;
1090 lwkt_setpri_self(TDPRI_KERN_USER);
1091 lp->lwp_proc->p_usched->release_curproc(lp);
1096 * This implements a normal yield. This routine is virtually a nop if
1097 * there is nothing to yield to but it will always run any pending interrupts
1098 * if called from a critical section.
1100 * This yield is designed for kernel threads without a user context.
1102 * (self contained on a per cpu basis)
1107 globaldata_t gd = mycpu;
1108 thread_t td = gd->gd_curthread;
1111 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1113 if (td->td_fairq_accum < 0) {
1114 lwkt_schedule_self(curthread);
1117 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1118 if (xtd && xtd->td_pri > td->td_pri) {
1119 lwkt_schedule_self(curthread);
1126 * This yield is designed for kernel threads with a user context.
1128 * The kernel acting on behalf of the user is potentially cpu-bound,
1129 * this function will efficiently allow other threads to run and also
1130 * switch to other processes by releasing.
1132 * The lwkt_user_yield() function is designed to have very low overhead
1133 * if no yield is determined to be needed.
1136 lwkt_user_yield(void)
1138 globaldata_t gd = mycpu;
1139 thread_t td = gd->gd_curthread;
1142 * Always run any pending interrupts in case we are in a critical
1145 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1149 * Switch (which forces a release) if another kernel thread needs
1150 * the cpu, if userland wants us to resched, or if our kernel
1151 * quantum has run out.
1153 if (lwkt_resched_wanted() ||
1154 user_resched_wanted() ||
1155 td->td_fairq_accum < 0)
1162 * Reacquire the current process if we are released.
1164 * XXX not implemented atm. The kernel may be holding locks and such,
1165 * so we want the thread to continue to receive cpu.
1167 if (td->td_release == NULL && lp) {
1168 lp->lwp_proc->p_usched->acquire_curproc(lp);
1169 td->td_release = lwkt_passive_release;
1170 lwkt_setpri_self(TDPRI_USER_NORM);
1176 * Generic schedule. Possibly schedule threads belonging to other cpus and
1177 * deal with threads that might be blocked on a wait queue.
1179 * We have a little helper inline function which does additional work after
1180 * the thread has been enqueued, including dealing with preemption and
1181 * setting need_lwkt_resched() (which prevents the kernel from returning
1182 * to userland until it has processed higher priority threads).
1184 * It is possible for this routine to be called after a failed _enqueue
1185 * (due to the target thread migrating, sleeping, or otherwise blocked).
1186 * We have to check that the thread is actually on the run queue!
1188 * reschedok is an optimized constant propagated from lwkt_schedule() or
1189 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1190 * reschedule to be requested if the target thread has a higher priority.
1191 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1192 * be 0, prevented undesired reschedules.
1196 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1200 if (ntd->td_flags & TDF_RUNQ) {
1201 if (ntd->td_preemptable && reschedok) {
1202 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1203 } else if (reschedok) {
1205 if (ntd->td_pri > otd->td_pri)
1206 need_lwkt_resched();
1210 * Give the thread a little fair share scheduler bump if it
1211 * has been asleep for a while. This is primarily to avoid
1212 * a degenerate case for interrupt threads where accumulator
1213 * crosses into negative territory unnecessarily.
1215 if (ntd->td_fairq_lticks != ticks) {
1216 ntd->td_fairq_lticks = ticks;
1217 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1218 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1219 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1226 _lwkt_schedule(thread_t td, int reschedok)
1228 globaldata_t mygd = mycpu;
1230 KASSERT(td != &td->td_gd->gd_idlethread,
1231 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1232 crit_enter_gd(mygd);
1233 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1234 if (td == mygd->gd_curthread) {
1238 * If we own the thread, there is no race (since we are in a
1239 * critical section). If we do not own the thread there might
1240 * be a race but the target cpu will deal with it.
1243 if (td->td_gd == mygd) {
1245 _lwkt_schedule_post(mygd, td, 1, reschedok);
1247 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1251 _lwkt_schedule_post(mygd, td, 1, reschedok);
1258 lwkt_schedule(thread_t td)
1260 _lwkt_schedule(td, 1);
1264 lwkt_schedule_noresched(thread_t td)
1266 _lwkt_schedule(td, 0);
1272 * When scheduled remotely if frame != NULL the IPIQ is being
1273 * run via doreti or an interrupt then preemption can be allowed.
1275 * To allow preemption we have to drop the critical section so only
1276 * one is present in _lwkt_schedule_post.
1279 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1281 thread_t td = curthread;
1284 if (frame && ntd->td_preemptable) {
1285 crit_exit_noyield(td);
1286 _lwkt_schedule(ntd, 1);
1287 crit_enter_quick(td);
1289 _lwkt_schedule(ntd, 1);
1294 * Thread migration using a 'Pull' method. The thread may or may not be
1295 * the current thread. It MUST be descheduled and in a stable state.
1296 * lwkt_giveaway() must be called on the cpu owning the thread.
1298 * At any point after lwkt_giveaway() is called, the target cpu may
1299 * 'pull' the thread by calling lwkt_acquire().
1301 * We have to make sure the thread is not sitting on a per-cpu tsleep
1302 * queue or it will blow up when it moves to another cpu.
1304 * MPSAFE - must be called under very specific conditions.
1307 lwkt_giveaway(thread_t td)
1309 globaldata_t gd = mycpu;
1312 if (td->td_flags & TDF_TSLEEPQ)
1314 KKASSERT(td->td_gd == gd);
1315 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1316 td->td_flags |= TDF_MIGRATING;
1321 lwkt_acquire(thread_t td)
1326 KKASSERT(td->td_flags & TDF_MIGRATING);
1331 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1332 crit_enter_gd(mygd);
1333 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1335 lwkt_process_ipiq();
1341 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1342 td->td_flags &= ~TDF_MIGRATING;
1345 crit_enter_gd(mygd);
1346 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1347 td->td_flags &= ~TDF_MIGRATING;
1355 * Generic deschedule. Descheduling threads other then your own should be
1356 * done only in carefully controlled circumstances. Descheduling is
1359 * This function may block if the cpu has run out of messages.
1362 lwkt_deschedule(thread_t td)
1366 if (td == curthread) {
1369 if (td->td_gd == mycpu) {
1372 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1382 * Set the target thread's priority. This routine does not automatically
1383 * switch to a higher priority thread, LWKT threads are not designed for
1384 * continuous priority changes. Yield if you want to switch.
1387 lwkt_setpri(thread_t td, int pri)
1389 KKASSERT(td->td_gd == mycpu);
1390 if (td->td_pri != pri) {
1393 if (td->td_flags & TDF_RUNQ) {
1405 * Set the initial priority for a thread prior to it being scheduled for
1406 * the first time. The thread MUST NOT be scheduled before or during
1407 * this call. The thread may be assigned to a cpu other then the current
1410 * Typically used after a thread has been created with TDF_STOPPREQ,
1411 * and before the thread is initially scheduled.
1414 lwkt_setpri_initial(thread_t td, int pri)
1417 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1422 lwkt_setpri_self(int pri)
1424 thread_t td = curthread;
1426 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1428 if (td->td_flags & TDF_RUNQ) {
1439 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1441 * Example: two competing threads, same priority N. decrement by (2*N)
1442 * increment by N*8, each thread will get 4 ticks.
1445 lwkt_fairq_schedulerclock(thread_t td)
1452 if (td != &gd->gd_idlethread) {
1453 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1454 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1455 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
1456 if (td->td_fairq_accum < 0)
1457 need_lwkt_resched();
1458 td->td_fairq_lticks = ticks;
1460 td = td->td_preempted;
1466 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1468 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1469 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1470 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1474 * Migrate the current thread to the specified cpu.
1476 * This is accomplished by descheduling ourselves from the current cpu,
1477 * moving our thread to the tdallq of the target cpu, IPI messaging the
1478 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1479 * races while the thread is being migrated.
1481 * We must be sure to remove ourselves from the current cpu's tsleepq
1482 * before potentially moving to another queue. The thread can be on
1483 * a tsleepq due to a left-over tsleep_interlock().
1486 static void lwkt_setcpu_remote(void *arg);
1490 lwkt_setcpu_self(globaldata_t rgd)
1493 thread_t td = curthread;
1495 if (td->td_gd != rgd) {
1496 crit_enter_quick(td);
1497 if (td->td_flags & TDF_TSLEEPQ)
1499 td->td_flags |= TDF_MIGRATING;
1500 lwkt_deschedule_self(td);
1501 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1502 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1504 /* we are now on the target cpu */
1505 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1506 crit_exit_quick(td);
1512 lwkt_migratecpu(int cpuid)
1517 rgd = globaldata_find(cpuid);
1518 lwkt_setcpu_self(rgd);
1523 * Remote IPI for cpu migration (called while in a critical section so we
1524 * do not have to enter another one). The thread has already been moved to
1525 * our cpu's allq, but we must wait for the thread to be completely switched
1526 * out on the originating cpu before we schedule it on ours or the stack
1527 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1528 * change to main memory.
1530 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1531 * against wakeups. It is best if this interface is used only when there
1532 * are no pending events that might try to schedule the thread.
1536 lwkt_setcpu_remote(void *arg)
1539 globaldata_t gd = mycpu;
1541 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1543 lwkt_process_ipiq();
1550 td->td_flags &= ~TDF_MIGRATING;
1551 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1557 lwkt_preempted_proc(void)
1559 thread_t td = curthread;
1560 while (td->td_preempted)
1561 td = td->td_preempted;
1566 * Create a kernel process/thread/whatever. It shares it's address space
1567 * with proc0 - ie: kernel only.
1569 * NOTE! By default new threads are created with the MP lock held. A
1570 * thread which does not require the MP lock should release it by calling
1571 * rel_mplock() at the start of the new thread.
1574 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1575 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1580 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1584 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1587 * Set up arg0 for 'ps' etc
1589 __va_start(ap, fmt);
1590 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1594 * Schedule the thread to run
1596 if ((td->td_flags & TDF_STOPREQ) == 0)
1599 td->td_flags &= ~TDF_STOPREQ;
1604 * Destroy an LWKT thread. Warning! This function is not called when
1605 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1606 * uses a different reaping mechanism.
1611 thread_t td = curthread;
1616 * Do any cleanup that might block here
1618 if (td->td_flags & TDF_VERBOSE)
1619 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1622 dsched_exit_thread(td);
1625 * Get us into a critical section to interlock gd_freetd and loop
1626 * until we can get it freed.
1628 * We have to cache the current td in gd_freetd because objcache_put()ing
1629 * it would rip it out from under us while our thread is still active.
1632 crit_enter_quick(td);
1633 while ((std = gd->gd_freetd) != NULL) {
1634 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1635 gd->gd_freetd = NULL;
1636 objcache_put(thread_cache, std);
1640 * Remove thread resources from kernel lists and deschedule us for
1641 * the last time. We cannot block after this point or we may end
1642 * up with a stale td on the tsleepq.
1644 if (td->td_flags & TDF_TSLEEPQ)
1646 lwkt_deschedule_self(td);
1647 lwkt_remove_tdallq(td);
1648 KKASSERT(td->td_refs == 0);
1653 KKASSERT(gd->gd_freetd == NULL);
1654 if (td->td_flags & TDF_ALLOCATED_THREAD)
1660 lwkt_remove_tdallq(thread_t td)
1662 KKASSERT(td->td_gd == mycpu);
1663 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1667 * Code reduction and branch prediction improvements. Call/return
1668 * overhead on modern cpus often degenerates into 0 cycles due to
1669 * the cpu's branch prediction hardware and return pc cache. We
1670 * can take advantage of this by not inlining medium-complexity
1671 * functions and we can also reduce the branch prediction impact
1672 * by collapsing perfectly predictable branches into a single
1673 * procedure instead of duplicating it.
1675 * Is any of this noticeable? Probably not, so I'll take the
1676 * smaller code size.
1679 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1681 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1687 thread_t td = curthread;
1688 int lcrit = td->td_critcount;
1690 td->td_critcount = 0;
1691 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1698 * Called from debugger/panic on cpus which have been stopped. We must still
1699 * process the IPIQ while stopped, even if we were stopped while in a critical
1702 * If we are dumping also try to process any pending interrupts. This may
1703 * or may not work depending on the state of the cpu at the point it was
1707 lwkt_smp_stopped(void)
1709 globaldata_t gd = mycpu;
1713 lwkt_process_ipiq();
1716 lwkt_process_ipiq();