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 lwkt_spin_fatal = 0; /* disabled */
145 SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
146 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
147 static int preempt_enable = 1;
148 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
149 &preempt_enable, 0, "Enable preemption");
150 static int lwkt_cache_threads = 32;
151 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
152 &lwkt_cache_threads, 0, "thread+kstack cache");
154 static __cachealign int lwkt_cseq_rindex;
155 static __cachealign int lwkt_cseq_windex;
158 * These helper procedures handle the runq, they can only be called from
159 * within a critical section.
161 * WARNING! Prior to SMP being brought up it is possible to enqueue and
162 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
163 * instead of 'mycpu' when referencing the globaldata structure. Once
164 * SMP live enqueuing and dequeueing only occurs on the current cpu.
168 _lwkt_dequeue(thread_t td)
170 if (td->td_flags & TDF_RUNQ) {
171 struct globaldata *gd = td->td_gd;
173 td->td_flags &= ~TDF_RUNQ;
174 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
175 gd->gd_fairq_total_pri -= td->td_pri;
176 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
177 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
184 * NOTE: There are a limited number of lwkt threads runnable since user
185 * processes only schedule one at a time per cpu.
189 _lwkt_enqueue(thread_t td)
193 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
194 struct globaldata *gd = td->td_gd;
196 td->td_flags |= TDF_RUNQ;
197 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
199 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
200 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
202 while (xtd && xtd->td_pri > td->td_pri)
203 xtd = TAILQ_NEXT(xtd, td_threadq);
205 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
209 gd->gd_fairq_total_pri += td->td_pri;
214 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
216 struct thread *td = (struct thread *)obj;
218 td->td_kstack = NULL;
219 td->td_kstack_size = 0;
220 td->td_flags = TDF_ALLOCATED_THREAD;
225 _lwkt_thread_dtor(void *obj, void *privdata)
227 struct thread *td = (struct thread *)obj;
229 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
230 ("_lwkt_thread_dtor: not allocated from objcache"));
231 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
232 td->td_kstack_size > 0,
233 ("_lwkt_thread_dtor: corrupted stack"));
234 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
238 * Initialize the lwkt s/system.
240 * Nominally cache up to 32 thread + kstack structures.
245 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
246 thread_cache = objcache_create_mbacked(
247 M_THREAD, sizeof(struct thread),
248 NULL, lwkt_cache_threads,
249 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
253 * Schedule a thread to run. As the current thread we can always safely
254 * schedule ourselves, and a shortcut procedure is provided for that
257 * (non-blocking, self contained on a per cpu basis)
260 lwkt_schedule_self(thread_t td)
262 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
263 crit_enter_quick(td);
264 KASSERT(td != &td->td_gd->gd_idlethread,
265 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
266 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
272 * Deschedule a thread.
274 * (non-blocking, self contained on a per cpu basis)
277 lwkt_deschedule_self(thread_t td)
279 crit_enter_quick(td);
285 * LWKTs operate on a per-cpu basis
287 * WARNING! Called from early boot, 'mycpu' may not work yet.
290 lwkt_gdinit(struct globaldata *gd)
292 TAILQ_INIT(&gd->gd_tdrunq);
293 TAILQ_INIT(&gd->gd_tdallq);
297 * Create a new thread. The thread must be associated with a process context
298 * or LWKT start address before it can be scheduled. If the target cpu is
299 * -1 the thread will be created on the current cpu.
301 * If you intend to create a thread without a process context this function
302 * does everything except load the startup and switcher function.
305 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
307 globaldata_t gd = mycpu;
311 * If static thread storage is not supplied allocate a thread. Reuse
312 * a cached free thread if possible. gd_freetd is used to keep an exiting
313 * thread intact through the exit.
317 if ((td = gd->gd_freetd) != NULL) {
318 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
320 gd->gd_freetd = NULL;
322 td = objcache_get(thread_cache, M_WAITOK);
323 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
327 KASSERT((td->td_flags &
328 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
329 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
330 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
334 * Try to reuse cached stack.
336 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
337 if (flags & TDF_ALLOCATED_STACK) {
338 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
343 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
344 flags |= TDF_ALLOCATED_STACK;
347 lwkt_init_thread(td, stack, stksize, flags, gd);
349 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
354 * Initialize a preexisting thread structure. This function is used by
355 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
357 * All threads start out in a critical section at a priority of
358 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
359 * appropriate. This function may send an IPI message when the
360 * requested cpu is not the current cpu and consequently gd_tdallq may
361 * not be initialized synchronously from the point of view of the originating
364 * NOTE! we have to be careful in regards to creating threads for other cpus
365 * if SMP has not yet been activated.
370 lwkt_init_thread_remote(void *arg)
375 * Protected by critical section held by IPI dispatch
377 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
383 * lwkt core thread structural initialization.
385 * NOTE: All threads are initialized as mpsafe threads.
388 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
389 struct globaldata *gd)
391 globaldata_t mygd = mycpu;
393 bzero(td, sizeof(struct thread));
394 td->td_kstack = stack;
395 td->td_kstack_size = stksize;
396 td->td_flags = flags;
398 td->td_pri = TDPRI_KERN_DAEMON;
399 td->td_critcount = 1;
400 td->td_toks_stop = &td->td_toks_base;
401 if (lwkt_use_spin_port)
402 lwkt_initport_spin(&td->td_msgport);
404 lwkt_initport_thread(&td->td_msgport, td);
405 pmap_init_thread(td);
408 * Normally initializing a thread for a remote cpu requires sending an
409 * IPI. However, the idlethread is setup before the other cpus are
410 * activated so we have to treat it as a special case. XXX manipulation
411 * of gd_tdallq requires the BGL.
413 if (gd == mygd || td == &gd->gd_idlethread) {
415 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
418 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
422 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
426 dsched_new_thread(td);
430 lwkt_set_comm(thread_t td, const char *ctl, ...)
435 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
437 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
441 lwkt_hold(thread_t td)
443 atomic_add_int(&td->td_refs, 1);
447 lwkt_rele(thread_t td)
449 KKASSERT(td->td_refs > 0);
450 atomic_add_int(&td->td_refs, -1);
454 lwkt_wait_free(thread_t td)
457 tsleep(td, 0, "tdreap", hz);
461 lwkt_free_thread(thread_t td)
463 KKASSERT(td->td_refs == 0);
464 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
465 if (td->td_flags & TDF_ALLOCATED_THREAD) {
466 objcache_put(thread_cache, td);
467 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
468 /* client-allocated struct with internally allocated stack */
469 KASSERT(td->td_kstack && td->td_kstack_size > 0,
470 ("lwkt_free_thread: corrupted stack"));
471 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
472 td->td_kstack = NULL;
473 td->td_kstack_size = 0;
475 KTR_LOG(ctxsw_deadtd, td);
480 * Switch to the next runnable lwkt. If no LWKTs are runnable then
481 * switch to the idlethread. Switching must occur within a critical
482 * section to avoid races with the scheduling queue.
484 * We always have full control over our cpu's run queue. Other cpus
485 * that wish to manipulate our queue must use the cpu_*msg() calls to
486 * talk to our cpu, so a critical section is all that is needed and
487 * the result is very, very fast thread switching.
489 * The LWKT scheduler uses a fixed priority model and round-robins at
490 * each priority level. User process scheduling is a totally
491 * different beast and LWKT priorities should not be confused with
492 * user process priorities.
494 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
495 * is not called by the current thread in the preemption case, only when
496 * the preempting thread blocks (in order to return to the original thread).
498 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
499 * migration and tsleep deschedule the current lwkt thread and call
500 * lwkt_switch(). In particular, the target cpu of the migration fully
501 * expects the thread to become non-runnable and can deadlock against
502 * cpusync operations if we run any IPIs prior to switching the thread out.
504 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
505 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
510 globaldata_t gd = mycpu;
511 thread_t td = gd->gd_curthread;
514 int spinning = lwkt_spin_loops; /* loops before HLTing */
521 * Switching from within a 'fast' (non thread switched) interrupt or IPI
522 * is illegal. However, we may have to do it anyway if we hit a fatal
523 * kernel trap or we have paniced.
525 * If this case occurs save and restore the interrupt nesting level.
527 if (gd->gd_intr_nesting_level) {
531 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
532 panic("lwkt_switch: Attempt to switch from a "
533 "a fast interrupt, ipi, or hard code section, "
537 savegdnest = gd->gd_intr_nesting_level;
538 savegdtrap = gd->gd_trap_nesting_level;
539 gd->gd_intr_nesting_level = 0;
540 gd->gd_trap_nesting_level = 0;
541 if ((td->td_flags & TDF_PANICWARN) == 0) {
542 td->td_flags |= TDF_PANICWARN;
543 kprintf("Warning: thread switch from interrupt, IPI, "
544 "or hard code section.\n"
545 "thread %p (%s)\n", td, td->td_comm);
549 gd->gd_intr_nesting_level = savegdnest;
550 gd->gd_trap_nesting_level = savegdtrap;
556 * Passive release (used to transition from user to kernel mode
557 * when we block or switch rather then when we enter the kernel).
558 * This function is NOT called if we are switching into a preemption
559 * or returning from a preemption. Typically this causes us to lose
560 * our current process designation (if we have one) and become a true
561 * LWKT thread, and may also hand the current process designation to
562 * another process and schedule thread.
568 if (TD_TOKS_HELD(td))
569 lwkt_relalltokens(td);
572 * We had better not be holding any spin locks, but don't get into an
573 * endless panic loop.
575 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
576 ("lwkt_switch: still holding %d exclusive spinlocks!",
577 gd->gd_spinlocks_wr));
582 if (td->td_cscount) {
583 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
585 if (panic_on_cscount)
586 panic("switching while mastering cpusync");
592 * If we had preempted another thread on this cpu, resume the preempted
593 * thread. This occurs transparently, whether the preempted thread
594 * was scheduled or not (it may have been preempted after descheduling
597 * We have to setup the MP lock for the original thread after backing
598 * out the adjustment that was made to curthread when the original
601 if ((ntd = td->td_preempted) != NULL) {
602 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
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 * If we cannot obtain ownership of the tokens we cannot immediately
623 * schedule the target thread.
625 * Reminder: Again, we cannot afford to run any IPIs in this path if
626 * the current thread has been descheduled.
630 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
631 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
635 reqflags = gd->gd_reqflags;
636 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
637 (reqflags & ~RQF_AST_LWKT_RESCHED) |
644 * Hotpath - pull the head of the run queue and attempt to schedule
645 * it. Fairq exhaustion moves the task to the end of the list. If
646 * no threads are runnable we switch to the idle thread.
649 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
653 * Runq is empty, switch to idle and clear RQF_WAKEUP
654 * to allow it to halt.
656 ntd = &gd->gd_idlethread;
658 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
659 ASSERT_NO_TOKENS_HELD(ntd);
661 cpu_time.cp_msg[0] = 0;
662 cpu_time.cp_stallpc = 0;
663 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
667 if (ntd->td_fairq_accum >= 0)
670 /*splz_check(); cannot do this here, see above */
671 lwkt_fairq_accumulate(gd, ntd);
672 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
673 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
677 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
678 * unwanted decontention IPIs.
680 * NOTE: For UP there is no mplock and lwkt_getalltokens()
683 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
687 * Coldpath (SMP only since tokens always succeed on UP)
689 * We had some contention on the thread we wanted to schedule.
690 * What we do now is try to find a thread that we can schedule
691 * in its stead until decontention reschedules on our cpu.
693 * The coldpath scan does NOT rearrange threads in the run list
694 * and it also ignores the accumulator.
696 * We do not immediately schedule a user priority thread, instead
697 * we record it in xtd and continue looking for kernel threads.
698 * A cpu can only have one user priority thread (normally) so just
699 * record the first one.
701 * NOTE: This scan will also include threads whos fairq's were
702 * accumulated in the first loop.
704 ++token_contention_count;
706 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
708 * Try to switch to this thread. If the thread is running at
709 * user priority we clear WAKEUP to allow decontention IPIs
710 * (since this thread is simply running until the one we wanted
711 * decontends), and we make sure that LWKT_RESCHED is not set.
713 * Otherwise for kernel threads we leave WAKEUP set to avoid
714 * unnecessary decontention IPIs.
716 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
723 * Do not let the fairq get too negative. Even though we are
724 * ignoring it atm once the scheduler decontends a very negative
725 * thread will get moved to the end of the queue.
727 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
728 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
729 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
734 * Well fubar, this thread is contended as well, loop
740 * We exhausted the run list but we may have recorded a user
741 * thread to try. We have three choices based on
742 * lwkt.decontention_method.
744 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
745 * IPIs (to interrupt the user process) and test
746 * RQF_AST_LWKT_RESCHED at the same time.
748 * This results in significant decontention IPI traffic but may
749 * be more responsive.
751 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
752 * An automatic LWKT reschedule will occur on the next hardclock
755 * This results in no decontention IPI traffic but may be less
756 * responsive. This is the default.
758 * (2) Refuse to schedule the user process at this time.
760 * This is highly experimental and should not be used under
761 * normal circumstances. This can cause a user process to
762 * get starved out in situations where kernel threads are
763 * fighting each other for tokens.
768 switch(lwkt_spin_method) {
771 reqflags = gd->gd_reqflags;
772 if (atomic_cmpset_int(&gd->gd_reqflags,
774 reqflags & ~RQF_WAKEUP)) {
780 reqflags = gd->gd_reqflags;
786 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
787 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
789 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
790 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
796 * Make sure RQF_WAKEUP is set if we failed to schedule the
797 * user thread to prevent the idle thread from halting.
799 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
803 * We exhausted the run list, meaning that all runnable threads
807 ntd = &gd->gd_idlethread;
809 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
810 ASSERT_NO_TOKENS_HELD(ntd);
811 /* contention case, do not clear contention mask */
815 * Ok, we might want to spin a few times as some tokens are held for
816 * very short periods of time and IPI overhead is 1uS or worse
817 * (meaning it is usually better to spin). Regardless we have to
818 * call splz_check() to be sure to service any interrupts blocked
819 * by our critical section, otherwise we could livelock e.g. IPIs.
821 * The IPI mechanic is really a last resort. In nearly all other
822 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
824 * When we decide not to spin we clear RQF_WAKEUP and switch to
825 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
826 * to halt and decontended tokens will issue an IPI to us. The
827 * idle thread will check for pending reschedules already set
828 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
831 * Also, if TDF_RUNQ is not set the current thread is trying to
832 * deschedule, possibly in an atomic fashion. We cannot afford to
835 if (spinning <= 0 || (td->td_flags & TDF_RUNQ) == 0) {
836 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
842 * When spinning a delay is required both to avoid livelocks from
843 * token order reversals (a thread may be trying to acquire multiple
844 * tokens), and also to reduce cpu cache management traffic.
846 * In order to scale to a large number of CPUs we use a time slot
847 * resequencer to force contending cpus into non-contending
848 * time-slots. The scheduler may still contend with the lock holder
849 * but will not (generally) contend with all the other cpus trying
850 * trying to get the same token.
852 * The resequencer uses a FIFO counter mechanic. The owner of the
853 * rindex at the head of the FIFO is allowed to pull itself off
854 * the FIFO and fetchadd is used to enter into the FIFO. This bit
855 * of code is VERY cache friendly and forces all spinning schedulers
856 * into their own time slots.
858 * This code has been tested to 48-cpus and caps the cache
859 * contention load at ~1uS intervals regardless of the number of
860 * cpus. Scaling beyond 64 cpus might require additional smarts
861 * (such as separate FIFOs for specific token cases).
863 * WARNING! We can't call splz_check() or anything else here as
864 * it could cause a deadlock.
866 #if defined(INVARIANTS) && defined(__amd64__)
867 if ((read_rflags() & PSL_I) == 0) {
869 panic("lwkt_switch() called with interrupts disabled");
872 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
873 fatal_count = lwkt_spin_fatal;
874 while ((oseq = lwkt_cseq_rindex) != cseq) {
876 #if !defined(_KERNEL_VIRTUAL)
877 if (cpu_mi_feature & CPU_MI_MONITOR) {
878 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
885 if (fatal_count && --fatal_count == 0)
886 panic("lwkt_switch: fatal spin wait");
888 cseq = lwkt_spin_delay; /* don't trust the system operator */
895 atomic_add_int(&lwkt_cseq_rindex, 1);
896 splz_check(); /* ok, we already checked that td is still scheduled */
897 /* highest level for(;;) loop */
902 * We must always decrement td_fairq_accum on non-idle threads just
903 * in case a thread never gets a tick due to being in a continuous
904 * critical section. The page-zeroing code does this, for example.
906 * If the thread we came up with is a higher or equal priority verses
907 * the thread at the head of the queue we move our thread to the
908 * front. This way we can always check the front of the queue.
910 * Clear gd_idle_repeat when doing a normal switch to a non-idle
913 ++gd->gd_cnt.v_swtch;
914 --ntd->td_fairq_accum;
915 ntd->td_wmesg = NULL;
916 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
917 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
918 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
919 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
921 gd->gd_idle_repeat = 0;
923 havethread_preempted:
925 * If the new target does not need the MP lock and we are holding it,
926 * release the MP lock. If the new target requires the MP lock we have
927 * already acquired it for the target.
931 KASSERT(ntd->td_critcount,
932 ("priority problem in lwkt_switch %d %d",
933 td->td_critcount, ntd->td_critcount));
937 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
940 /* NOTE: current cpu may have changed after switch */
945 * Request that the target thread preempt the current thread. Preemption
946 * only works under a specific set of conditions:
948 * - We are not preempting ourselves
949 * - The target thread is owned by the current cpu
950 * - We are not currently being preempted
951 * - The target is not currently being preempted
952 * - We are not holding any spin locks
953 * - The target thread is not holding any tokens
954 * - We are able to satisfy the target's MP lock requirements (if any).
956 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
957 * this is called via lwkt_schedule() through the td_preemptable callback.
958 * critcount is the managed critical priority that we should ignore in order
959 * to determine whether preemption is possible (aka usually just the crit
960 * priority of lwkt_schedule() itself).
962 * XXX at the moment we run the target thread in a critical section during
963 * the preemption in order to prevent the target from taking interrupts
964 * that *WE* can't. Preemption is strictly limited to interrupt threads
965 * and interrupt-like threads, outside of a critical section, and the
966 * preempted source thread will be resumed the instant the target blocks
967 * whether or not the source is scheduled (i.e. preemption is supposed to
968 * be as transparent as possible).
971 lwkt_preempt(thread_t ntd, int critcount)
973 struct globaldata *gd = mycpu;
975 int save_gd_intr_nesting_level;
978 * The caller has put us in a critical section. We can only preempt
979 * if the caller of the caller was not in a critical section (basically
980 * a local interrupt), as determined by the 'critcount' parameter. We
981 * also can't preempt if the caller is holding any spinlocks (even if
982 * he isn't in a critical section). This also handles the tokens test.
984 * YYY The target thread must be in a critical section (else it must
985 * inherit our critical section? I dunno yet).
987 * Set need_lwkt_resched() unconditionally for now YYY.
989 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
991 if (preempt_enable == 0) {
996 td = gd->gd_curthread;
997 if (ntd->td_pri <= td->td_pri) {
1001 if (td->td_critcount > critcount) {
1003 need_lwkt_resched();
1007 if (ntd->td_gd != gd) {
1009 need_lwkt_resched();
1014 * We don't have to check spinlocks here as they will also bump
1017 * Do not try to preempt if the target thread is holding any tokens.
1018 * We could try to acquire the tokens but this case is so rare there
1019 * is no need to support it.
1021 KKASSERT(gd->gd_spinlocks_wr == 0);
1023 if (TD_TOKS_HELD(ntd)) {
1025 need_lwkt_resched();
1028 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1030 need_lwkt_resched();
1033 if (ntd->td_preempted) {
1035 need_lwkt_resched();
1040 * Since we are able to preempt the current thread, there is no need to
1041 * call need_lwkt_resched().
1043 * We must temporarily clear gd_intr_nesting_level around the switch
1044 * since switchouts from the target thread are allowed (they will just
1045 * return to our thread), and since the target thread has its own stack.
1048 ntd->td_preempted = td;
1049 td->td_flags |= TDF_PREEMPT_LOCK;
1050 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1051 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1052 gd->gd_intr_nesting_level = 0;
1054 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1056 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1057 ntd->td_preempted = NULL;
1058 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1062 * Conditionally call splz() if gd_reqflags indicates work is pending.
1063 * This will work inside a critical section but not inside a hard code
1066 * (self contained on a per cpu basis)
1071 globaldata_t gd = mycpu;
1072 thread_t td = gd->gd_curthread;
1074 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1075 gd->gd_intr_nesting_level == 0 &&
1076 td->td_nest_count < 2)
1083 * This version is integrated into crit_exit, reqflags has already
1084 * been tested but td_critcount has not.
1086 * We only want to execute the splz() on the 1->0 transition of
1087 * critcount and not in a hard code section or if too deeply nested.
1090 lwkt_maybe_splz(thread_t td)
1092 globaldata_t gd = td->td_gd;
1094 if (td->td_critcount == 0 &&
1095 gd->gd_intr_nesting_level == 0 &&
1096 td->td_nest_count < 2)
1103 * This function is used to negotiate a passive release of the current
1104 * process/lwp designation with the user scheduler, allowing the user
1105 * scheduler to schedule another user thread. The related kernel thread
1106 * (curthread) continues running in the released state.
1109 lwkt_passive_release(struct thread *td)
1111 struct lwp *lp = td->td_lwp;
1113 td->td_release = NULL;
1114 lwkt_setpri_self(TDPRI_KERN_USER);
1115 lp->lwp_proc->p_usched->release_curproc(lp);
1120 * This implements a normal yield. This routine is virtually a nop if
1121 * there is nothing to yield to but it will always run any pending interrupts
1122 * if called from a critical section.
1124 * This yield is designed for kernel threads without a user context.
1126 * (self contained on a per cpu basis)
1131 globaldata_t gd = mycpu;
1132 thread_t td = gd->gd_curthread;
1135 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1137 if (td->td_fairq_accum < 0) {
1138 lwkt_schedule_self(curthread);
1141 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1142 if (xtd && xtd->td_pri > td->td_pri) {
1143 lwkt_schedule_self(curthread);
1150 * This yield is designed for kernel threads with a user context.
1152 * The kernel acting on behalf of the user is potentially cpu-bound,
1153 * this function will efficiently allow other threads to run and also
1154 * switch to other processes by releasing.
1156 * The lwkt_user_yield() function is designed to have very low overhead
1157 * if no yield is determined to be needed.
1160 lwkt_user_yield(void)
1162 globaldata_t gd = mycpu;
1163 thread_t td = gd->gd_curthread;
1166 * Always run any pending interrupts in case we are in a critical
1169 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1173 * Switch (which forces a release) if another kernel thread needs
1174 * the cpu, if userland wants us to resched, or if our kernel
1175 * quantum has run out.
1177 if (lwkt_resched_wanted() ||
1178 user_resched_wanted() ||
1179 td->td_fairq_accum < 0)
1186 * Reacquire the current process if we are released.
1188 * XXX not implemented atm. The kernel may be holding locks and such,
1189 * so we want the thread to continue to receive cpu.
1191 if (td->td_release == NULL && lp) {
1192 lp->lwp_proc->p_usched->acquire_curproc(lp);
1193 td->td_release = lwkt_passive_release;
1194 lwkt_setpri_self(TDPRI_USER_NORM);
1200 * Generic schedule. Possibly schedule threads belonging to other cpus and
1201 * deal with threads that might be blocked on a wait queue.
1203 * We have a little helper inline function which does additional work after
1204 * the thread has been enqueued, including dealing with preemption and
1205 * setting need_lwkt_resched() (which prevents the kernel from returning
1206 * to userland until it has processed higher priority threads).
1208 * It is possible for this routine to be called after a failed _enqueue
1209 * (due to the target thread migrating, sleeping, or otherwise blocked).
1210 * We have to check that the thread is actually on the run queue!
1212 * reschedok is an optimized constant propagated from lwkt_schedule() or
1213 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1214 * reschedule to be requested if the target thread has a higher priority.
1215 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1216 * be 0, prevented undesired reschedules.
1220 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1224 if (ntd->td_flags & TDF_RUNQ) {
1225 if (ntd->td_preemptable && reschedok) {
1226 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1227 } else if (reschedok) {
1229 if (ntd->td_pri > otd->td_pri)
1230 need_lwkt_resched();
1234 * Give the thread a little fair share scheduler bump if it
1235 * has been asleep for a while. This is primarily to avoid
1236 * a degenerate case for interrupt threads where accumulator
1237 * crosses into negative territory unnecessarily.
1239 if (ntd->td_fairq_lticks != ticks) {
1240 ntd->td_fairq_lticks = ticks;
1241 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1242 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1243 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1250 _lwkt_schedule(thread_t td, int reschedok)
1252 globaldata_t mygd = mycpu;
1254 KASSERT(td != &td->td_gd->gd_idlethread,
1255 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1256 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1257 crit_enter_gd(mygd);
1258 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1259 if (td == mygd->gd_curthread) {
1263 * If we own the thread, there is no race (since we are in a
1264 * critical section). If we do not own the thread there might
1265 * be a race but the target cpu will deal with it.
1268 if (td->td_gd == mygd) {
1270 _lwkt_schedule_post(mygd, td, 1, reschedok);
1272 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1276 _lwkt_schedule_post(mygd, td, 1, reschedok);
1283 lwkt_schedule(thread_t td)
1285 _lwkt_schedule(td, 1);
1289 lwkt_schedule_noresched(thread_t td)
1291 _lwkt_schedule(td, 0);
1297 * When scheduled remotely if frame != NULL the IPIQ is being
1298 * run via doreti or an interrupt then preemption can be allowed.
1300 * To allow preemption we have to drop the critical section so only
1301 * one is present in _lwkt_schedule_post.
1304 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1306 thread_t td = curthread;
1309 if (frame && ntd->td_preemptable) {
1310 crit_exit_noyield(td);
1311 _lwkt_schedule(ntd, 1);
1312 crit_enter_quick(td);
1314 _lwkt_schedule(ntd, 1);
1319 * Thread migration using a 'Pull' method. The thread may or may not be
1320 * the current thread. It MUST be descheduled and in a stable state.
1321 * lwkt_giveaway() must be called on the cpu owning the thread.
1323 * At any point after lwkt_giveaway() is called, the target cpu may
1324 * 'pull' the thread by calling lwkt_acquire().
1326 * We have to make sure the thread is not sitting on a per-cpu tsleep
1327 * queue or it will blow up when it moves to another cpu.
1329 * MPSAFE - must be called under very specific conditions.
1332 lwkt_giveaway(thread_t td)
1334 globaldata_t gd = mycpu;
1337 if (td->td_flags & TDF_TSLEEPQ)
1339 KKASSERT(td->td_gd == gd);
1340 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1341 td->td_flags |= TDF_MIGRATING;
1346 lwkt_acquire(thread_t td)
1351 KKASSERT(td->td_flags & TDF_MIGRATING);
1356 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1357 crit_enter_gd(mygd);
1358 DEBUG_PUSH_INFO("lwkt_acquire");
1359 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1361 lwkt_process_ipiq();
1368 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1369 td->td_flags &= ~TDF_MIGRATING;
1372 crit_enter_gd(mygd);
1373 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1374 td->td_flags &= ~TDF_MIGRATING;
1382 * Generic deschedule. Descheduling threads other then your own should be
1383 * done only in carefully controlled circumstances. Descheduling is
1386 * This function may block if the cpu has run out of messages.
1389 lwkt_deschedule(thread_t td)
1393 if (td == curthread) {
1396 if (td->td_gd == mycpu) {
1399 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1409 * Set the target thread's priority. This routine does not automatically
1410 * switch to a higher priority thread, LWKT threads are not designed for
1411 * continuous priority changes. Yield if you want to switch.
1414 lwkt_setpri(thread_t td, int pri)
1416 KKASSERT(td->td_gd == mycpu);
1417 if (td->td_pri != pri) {
1420 if (td->td_flags & TDF_RUNQ) {
1432 * Set the initial priority for a thread prior to it being scheduled for
1433 * the first time. The thread MUST NOT be scheduled before or during
1434 * this call. The thread may be assigned to a cpu other then the current
1437 * Typically used after a thread has been created with TDF_STOPPREQ,
1438 * and before the thread is initially scheduled.
1441 lwkt_setpri_initial(thread_t td, int pri)
1444 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1449 lwkt_setpri_self(int pri)
1451 thread_t td = curthread;
1453 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1455 if (td->td_flags & TDF_RUNQ) {
1466 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1468 * Example: two competing threads, same priority N. decrement by (2*N)
1469 * increment by N*8, each thread will get 4 ticks.
1472 lwkt_fairq_schedulerclock(thread_t td)
1479 if (td != &gd->gd_idlethread) {
1480 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1481 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1482 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
1483 if (td->td_fairq_accum < 0)
1484 need_lwkt_resched();
1485 td->td_fairq_lticks = ticks;
1487 td = td->td_preempted;
1493 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1495 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1496 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1497 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1501 * Migrate the current thread to the specified cpu.
1503 * This is accomplished by descheduling ourselves from the current cpu,
1504 * moving our thread to the tdallq of the target cpu, IPI messaging the
1505 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1506 * races while the thread is being migrated.
1508 * We must be sure to remove ourselves from the current cpu's tsleepq
1509 * before potentially moving to another queue. The thread can be on
1510 * a tsleepq due to a left-over tsleep_interlock().
1512 * We also have to make sure that the switch code doesn't allow an IPI
1513 * processing operation to leak in between our send and our switch, or
1514 * any other potential livelock such that might occur when we release the
1515 * current process designation, so do that first.
1518 static void lwkt_setcpu_remote(void *arg);
1522 lwkt_setcpu_self(globaldata_t rgd)
1525 thread_t td = curthread;
1527 if (td->td_gd != rgd) {
1528 crit_enter_quick(td);
1531 if (td->td_flags & TDF_TSLEEPQ)
1533 td->td_flags |= TDF_MIGRATING;
1534 lwkt_deschedule_self(td);
1535 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1536 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1538 /* we are now on the target cpu */
1539 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1540 crit_exit_quick(td);
1546 lwkt_migratecpu(int cpuid)
1551 rgd = globaldata_find(cpuid);
1552 lwkt_setcpu_self(rgd);
1557 * Remote IPI for cpu migration (called while in a critical section so we
1558 * do not have to enter another one). The thread has already been moved to
1559 * our cpu's allq, but we must wait for the thread to be completely switched
1560 * out on the originating cpu before we schedule it on ours or the stack
1561 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1562 * change to main memory.
1564 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1565 * against wakeups. It is best if this interface is used only when there
1566 * are no pending events that might try to schedule the thread.
1570 lwkt_setcpu_remote(void *arg)
1573 globaldata_t gd = mycpu;
1574 int retry = 10000000;
1576 DEBUG_PUSH_INFO("lwkt_setcpu_remote");
1577 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1579 lwkt_process_ipiq();
1584 kprintf("lwkt_setcpu_remote: td->td_flags %08x\n",
1592 td->td_flags &= ~TDF_MIGRATING;
1593 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1599 lwkt_preempted_proc(void)
1601 thread_t td = curthread;
1602 while (td->td_preempted)
1603 td = td->td_preempted;
1608 * Create a kernel process/thread/whatever. It shares it's address space
1609 * with proc0 - ie: kernel only.
1611 * NOTE! By default new threads are created with the MP lock held. A
1612 * thread which does not require the MP lock should release it by calling
1613 * rel_mplock() at the start of the new thread.
1616 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1617 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1622 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1626 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1629 * Set up arg0 for 'ps' etc
1631 __va_start(ap, fmt);
1632 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1636 * Schedule the thread to run
1638 if ((td->td_flags & TDF_STOPREQ) == 0)
1641 td->td_flags &= ~TDF_STOPREQ;
1646 * Destroy an LWKT thread. Warning! This function is not called when
1647 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1648 * uses a different reaping mechanism.
1653 thread_t td = curthread;
1658 * Do any cleanup that might block here
1660 if (td->td_flags & TDF_VERBOSE)
1661 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1664 dsched_exit_thread(td);
1667 * Get us into a critical section to interlock gd_freetd and loop
1668 * until we can get it freed.
1670 * We have to cache the current td in gd_freetd because objcache_put()ing
1671 * it would rip it out from under us while our thread is still active.
1674 crit_enter_quick(td);
1675 while ((std = gd->gd_freetd) != NULL) {
1676 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1677 gd->gd_freetd = NULL;
1678 objcache_put(thread_cache, std);
1682 * Remove thread resources from kernel lists and deschedule us for
1683 * the last time. We cannot block after this point or we may end
1684 * up with a stale td on the tsleepq.
1686 if (td->td_flags & TDF_TSLEEPQ)
1688 lwkt_deschedule_self(td);
1689 lwkt_remove_tdallq(td);
1690 KKASSERT(td->td_refs == 0);
1695 KKASSERT(gd->gd_freetd == NULL);
1696 if (td->td_flags & TDF_ALLOCATED_THREAD)
1702 lwkt_remove_tdallq(thread_t td)
1704 KKASSERT(td->td_gd == mycpu);
1705 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1709 * Code reduction and branch prediction improvements. Call/return
1710 * overhead on modern cpus often degenerates into 0 cycles due to
1711 * the cpu's branch prediction hardware and return pc cache. We
1712 * can take advantage of this by not inlining medium-complexity
1713 * functions and we can also reduce the branch prediction impact
1714 * by collapsing perfectly predictable branches into a single
1715 * procedure instead of duplicating it.
1717 * Is any of this noticeable? Probably not, so I'll take the
1718 * smaller code size.
1721 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1723 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1729 thread_t td = curthread;
1730 int lcrit = td->td_critcount;
1732 td->td_critcount = 0;
1733 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1740 * Called from debugger/panic on cpus which have been stopped. We must still
1741 * process the IPIQ while stopped, even if we were stopped while in a critical
1744 * If we are dumping also try to process any pending interrupts. This may
1745 * or may not work depending on the state of the cpu at the point it was
1749 lwkt_smp_stopped(void)
1751 globaldata_t gd = mycpu;
1755 lwkt_process_ipiq();
1758 lwkt_process_ipiq();