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 * Note that the td_switch() function cannot do anything that requires
476 * the MP lock since the MP lock will have already been setup for
477 * the target thread (not the current thread). It's nice to have a scheduler
478 * that does not need the MP lock to work because it allows us to do some
479 * really cool high-performance MP lock optimizations.
481 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
482 * is not called by the current thread in the preemption case, only when
483 * the preempting thread blocks (in order to return to the original thread).
488 globaldata_t gd = mycpu;
489 thread_t td = gd->gd_curthread;
497 * Switching from within a 'fast' (non thread switched) interrupt or IPI
498 * is illegal. However, we may have to do it anyway if we hit a fatal
499 * kernel trap or we have paniced.
501 * If this case occurs save and restore the interrupt nesting level.
503 if (gd->gd_intr_nesting_level) {
507 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
508 panic("lwkt_switch: Attempt to switch from a "
509 "a fast interrupt, ipi, or hard code section, "
513 savegdnest = gd->gd_intr_nesting_level;
514 savegdtrap = gd->gd_trap_nesting_level;
515 gd->gd_intr_nesting_level = 0;
516 gd->gd_trap_nesting_level = 0;
517 if ((td->td_flags & TDF_PANICWARN) == 0) {
518 td->td_flags |= TDF_PANICWARN;
519 kprintf("Warning: thread switch from interrupt, IPI, "
520 "or hard code section.\n"
521 "thread %p (%s)\n", td, td->td_comm);
525 gd->gd_intr_nesting_level = savegdnest;
526 gd->gd_trap_nesting_level = savegdtrap;
532 * Passive release (used to transition from user to kernel mode
533 * when we block or switch rather then when we enter the kernel).
534 * This function is NOT called if we are switching into a preemption
535 * or returning from a preemption. Typically this causes us to lose
536 * our current process designation (if we have one) and become a true
537 * LWKT thread, and may also hand the current process designation to
538 * another process and schedule thread.
544 if (TD_TOKS_HELD(td))
545 lwkt_relalltokens(td);
548 * We had better not be holding any spin locks, but don't get into an
549 * endless panic loop.
551 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
552 ("lwkt_switch: still holding %d exclusive spinlocks!",
553 gd->gd_spinlocks_wr));
558 if (td->td_cscount) {
559 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
561 if (panic_on_cscount)
562 panic("switching while mastering cpusync");
568 * If we had preempted another thread on this cpu, resume the preempted
569 * thread. This occurs transparently, whether the preempted thread
570 * was scheduled or not (it may have been preempted after descheduling
573 * We have to setup the MP lock for the original thread after backing
574 * out the adjustment that was made to curthread when the original
577 if ((ntd = td->td_preempted) != NULL) {
578 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
579 ntd->td_flags |= TDF_PREEMPT_DONE;
582 * The interrupt may have woken a thread up, we need to properly
583 * set the reschedule flag if the originally interrupted thread is
584 * at a lower priority.
586 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
587 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
590 /* YYY release mp lock on switchback if original doesn't need it */
591 goto havethread_preempted;
595 * Implement round-robin fairq with priority insertion. The priority
596 * insertion is handled by _lwkt_enqueue()
598 * We have to adjust the MP lock for the target thread. If we
599 * need the MP lock and cannot obtain it we try to locate a
600 * thread that does not need the MP lock. If we cannot, we spin
603 * A similar issue exists for the tokens held by the target thread.
604 * If we cannot obtain ownership of the tokens we cannot immediately
605 * schedule the thread.
608 clear_lwkt_resched();
610 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
613 * Hotpath if we can get all necessary resources.
615 * If nothing is runnable switch to the idle thread
618 ntd = &gd->gd_idlethread;
619 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
620 ntd->td_flags |= TDF_IDLE_NOHLT;
622 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
623 ASSERT_NO_TOKENS_HELD(ntd);
624 clr_cpu_contention_mask(gd);
626 cpu_time.cp_msg[0] = 0;
627 cpu_time.cp_stallpc = 0;
634 * NOTE: For UP there is no mplock and lwkt_getalltokens()
637 if (ntd->td_fairq_accum >= 0 &&
638 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
641 clr_cpu_contention_mask(gd);
647 if (ntd->td_fairq_accum >= 0)
648 set_cpu_contention_mask(gd);
652 * Coldpath - unable to schedule ntd, continue looking for threads
653 * to schedule. This is only allowed of the (presumably) kernel
654 * thread exhausted its fair share. A kernel thread stuck on
655 * resources does not currently allow a user thread to get in
659 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
660 (ntd->td_fairq_accum < 0));
668 * If the fair-share scheduler ran out ntd gets moved to the
669 * end and its accumulator will be bumped, if it didn't we
670 * maintain the same queue position.
672 * nlast keeps track of the last element prior to any moves.
674 if (ntd->td_fairq_accum < 0) {
675 lwkt_fairq_accumulate(gd, ntd);
681 xtd = TAILQ_NEXT(ntd, td_threadq);
682 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
683 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
686 * Set terminal element (nlast)
695 ntd = TAILQ_NEXT(ntd, td_threadq);
699 * If we exhausted the run list switch to the idle thread.
700 * Since one or more threads had resource acquisition issues
701 * we do not allow the idle thread to halt.
703 * NOTE: nlast can be NULL.
707 ntd = &gd->gd_idlethread;
708 ntd->td_flags |= TDF_IDLE_NOHLT;
710 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
711 ASSERT_NO_TOKENS_HELD(ntd);
712 /* contention case, do not clear contention mask */
716 * If fairq accumulations occured we do not schedule the
717 * idle thread. This will cause us to try again from
721 break; /* try again from the top, almost */
726 * Try to switch to this thread.
728 * NOTE: For UP there is no mplock and lwkt_getalltokens()
731 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
732 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
733 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
736 clr_cpu_contention_mask(gd);
742 * Thread was runnable but we were unable to get the required
743 * resources (tokens and/or mplock).
746 if (ntd->td_fairq_accum >= 0)
747 set_cpu_contention_mask(gd);
748 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
754 * All threads exhausted but we can loop due to a negative
757 * While we are looping in the scheduler be sure to service
758 * any interrupts which were made pending due to our critical
759 * section, otherwise we could livelock (e.g.) IPIs.
765 * We must always decrement td_fairq_accum on non-idle threads just
766 * in case a thread never gets a tick due to being in a continuous
767 * critical section. The page-zeroing code does that.
769 * If the thread we came up with is a higher or equal priority verses
770 * the thread at the head of the queue we move our thread to the
771 * front. This way we can always check the front of the queue.
774 ++gd->gd_cnt.v_swtch;
775 --ntd->td_fairq_accum;
776 ntd->td_wmesg = NULL;
777 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
778 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
779 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
780 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
782 havethread_preempted:
785 * If the new target does not need the MP lock and we are holding it,
786 * release the MP lock. If the new target requires the MP lock we have
787 * already acquired it for the target.
790 KASSERT(ntd->td_critcount,
791 ("priority problem in lwkt_switch %d %d",
792 td->td_critcount, ntd->td_critcount));
796 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
799 /* NOTE: current cpu may have changed after switch */
804 * Request that the target thread preempt the current thread. Preemption
805 * only works under a specific set of conditions:
807 * - We are not preempting ourselves
808 * - The target thread is owned by the current cpu
809 * - We are not currently being preempted
810 * - The target is not currently being preempted
811 * - We are not holding any spin locks
812 * - The target thread is not holding any tokens
813 * - We are able to satisfy the target's MP lock requirements (if any).
815 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
816 * this is called via lwkt_schedule() through the td_preemptable callback.
817 * critcount is the managed critical priority that we should ignore in order
818 * to determine whether preemption is possible (aka usually just the crit
819 * priority of lwkt_schedule() itself).
821 * XXX at the moment we run the target thread in a critical section during
822 * the preemption in order to prevent the target from taking interrupts
823 * that *WE* can't. Preemption is strictly limited to interrupt threads
824 * and interrupt-like threads, outside of a critical section, and the
825 * preempted source thread will be resumed the instant the target blocks
826 * whether or not the source is scheduled (i.e. preemption is supposed to
827 * be as transparent as possible).
830 lwkt_preempt(thread_t ntd, int critcount)
832 struct globaldata *gd = mycpu;
834 int save_gd_intr_nesting_level;
837 * The caller has put us in a critical section. We can only preempt
838 * if the caller of the caller was not in a critical section (basically
839 * a local interrupt), as determined by the 'critcount' parameter. We
840 * also can't preempt if the caller is holding any spinlocks (even if
841 * he isn't in a critical section). This also handles the tokens test.
843 * YYY The target thread must be in a critical section (else it must
844 * inherit our critical section? I dunno yet).
846 * Set need_lwkt_resched() unconditionally for now YYY.
848 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
850 if (preempt_enable == 0) {
855 td = gd->gd_curthread;
856 if (ntd->td_pri <= td->td_pri) {
860 if (td->td_critcount > critcount) {
866 if (ntd->td_gd != gd) {
873 * We don't have to check spinlocks here as they will also bump
876 * Do not try to preempt if the target thread is holding any tokens.
877 * We could try to acquire the tokens but this case is so rare there
878 * is no need to support it.
880 KKASSERT(gd->gd_spinlocks_wr == 0);
882 if (TD_TOKS_HELD(ntd)) {
887 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
892 if (ntd->td_preempted) {
899 * Since we are able to preempt the current thread, there is no need to
900 * call need_lwkt_resched().
902 * We must temporarily clear gd_intr_nesting_level around the switch
903 * since switchouts from the target thread are allowed (they will just
904 * return to our thread), and since the target thread has its own stack.
907 ntd->td_preempted = td;
908 td->td_flags |= TDF_PREEMPT_LOCK;
909 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
910 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
911 gd->gd_intr_nesting_level = 0;
913 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
915 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
916 ntd->td_preempted = NULL;
917 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
921 * Conditionally call splz() if gd_reqflags indicates work is pending.
922 * This will work inside a critical section but not inside a hard code
925 * (self contained on a per cpu basis)
930 globaldata_t gd = mycpu;
931 thread_t td = gd->gd_curthread;
933 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
934 gd->gd_intr_nesting_level == 0 &&
935 td->td_nest_count < 2)
942 * This version is integrated into crit_exit, reqflags has already
943 * been tested but td_critcount has not.
945 * We only want to execute the splz() on the 1->0 transition of
946 * critcount and not in a hard code section or if too deeply nested.
949 lwkt_maybe_splz(thread_t td)
951 globaldata_t gd = td->td_gd;
953 if (td->td_critcount == 0 &&
954 gd->gd_intr_nesting_level == 0 &&
955 td->td_nest_count < 2)
962 * This function is used to negotiate a passive release of the current
963 * process/lwp designation with the user scheduler, allowing the user
964 * scheduler to schedule another user thread. The related kernel thread
965 * (curthread) continues running in the released state.
968 lwkt_passive_release(struct thread *td)
970 struct lwp *lp = td->td_lwp;
972 td->td_release = NULL;
973 lwkt_setpri_self(TDPRI_KERN_USER);
974 lp->lwp_proc->p_usched->release_curproc(lp);
979 * This implements a normal yield. This routine is virtually a nop if
980 * there is nothing to yield to but it will always run any pending interrupts
981 * if called from a critical section.
983 * This yield is designed for kernel threads without a user context.
985 * (self contained on a per cpu basis)
990 globaldata_t gd = mycpu;
991 thread_t td = gd->gd_curthread;
994 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
996 if (td->td_fairq_accum < 0) {
997 lwkt_schedule_self(curthread);
1000 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1001 if (xtd && xtd->td_pri > td->td_pri) {
1002 lwkt_schedule_self(curthread);
1009 * This yield is designed for kernel threads with a user context.
1011 * The kernel acting on behalf of the user is potentially cpu-bound,
1012 * this function will efficiently allow other threads to run and also
1013 * switch to other processes by releasing.
1015 * The lwkt_user_yield() function is designed to have very low overhead
1016 * if no yield is determined to be needed.
1019 lwkt_user_yield(void)
1021 globaldata_t gd = mycpu;
1022 thread_t td = gd->gd_curthread;
1025 * Always run any pending interrupts in case we are in a critical
1028 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1032 * Switch (which forces a release) if another kernel thread needs
1033 * the cpu, if userland wants us to resched, or if our kernel
1034 * quantum has run out.
1036 if (lwkt_resched_wanted() ||
1037 user_resched_wanted() ||
1038 td->td_fairq_accum < 0)
1045 * Reacquire the current process if we are released.
1047 * XXX not implemented atm. The kernel may be holding locks and such,
1048 * so we want the thread to continue to receive cpu.
1050 if (td->td_release == NULL && lp) {
1051 lp->lwp_proc->p_usched->acquire_curproc(lp);
1052 td->td_release = lwkt_passive_release;
1053 lwkt_setpri_self(TDPRI_USER_NORM);
1059 * Generic schedule. Possibly schedule threads belonging to other cpus and
1060 * deal with threads that might be blocked on a wait queue.
1062 * We have a little helper inline function which does additional work after
1063 * the thread has been enqueued, including dealing with preemption and
1064 * setting need_lwkt_resched() (which prevents the kernel from returning
1065 * to userland until it has processed higher priority threads).
1067 * It is possible for this routine to be called after a failed _enqueue
1068 * (due to the target thread migrating, sleeping, or otherwise blocked).
1069 * We have to check that the thread is actually on the run queue!
1071 * reschedok is an optimized constant propagated from lwkt_schedule() or
1072 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1073 * reschedule to be requested if the target thread has a higher priority.
1074 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1075 * be 0, prevented undesired reschedules.
1079 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1083 if (ntd->td_flags & TDF_RUNQ) {
1084 if (ntd->td_preemptable && reschedok) {
1085 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1086 } else if (reschedok) {
1088 if (ntd->td_pri > otd->td_pri)
1089 need_lwkt_resched();
1093 * Give the thread a little fair share scheduler bump if it
1094 * has been asleep for a while. This is primarily to avoid
1095 * a degenerate case for interrupt threads where accumulator
1096 * crosses into negative territory unnecessarily.
1098 if (ntd->td_fairq_lticks != ticks) {
1099 ntd->td_fairq_lticks = ticks;
1100 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1101 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1102 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1109 _lwkt_schedule(thread_t td, int reschedok)
1111 globaldata_t mygd = mycpu;
1113 KASSERT(td != &td->td_gd->gd_idlethread,
1114 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1115 crit_enter_gd(mygd);
1116 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1117 if (td == mygd->gd_curthread) {
1121 * If we own the thread, there is no race (since we are in a
1122 * critical section). If we do not own the thread there might
1123 * be a race but the target cpu will deal with it.
1126 if (td->td_gd == mygd) {
1128 _lwkt_schedule_post(mygd, td, 1, reschedok);
1130 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1134 _lwkt_schedule_post(mygd, td, 1, reschedok);
1141 lwkt_schedule(thread_t td)
1143 _lwkt_schedule(td, 1);
1147 lwkt_schedule_noresched(thread_t td)
1149 _lwkt_schedule(td, 0);
1155 * When scheduled remotely if frame != NULL the IPIQ is being
1156 * run via doreti or an interrupt then preemption can be allowed.
1158 * To allow preemption we have to drop the critical section so only
1159 * one is present in _lwkt_schedule_post.
1162 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1164 thread_t td = curthread;
1167 if (frame && ntd->td_preemptable) {
1168 crit_exit_noyield(td);
1169 _lwkt_schedule(ntd, 1);
1170 crit_enter_quick(td);
1172 _lwkt_schedule(ntd, 1);
1177 * Thread migration using a 'Pull' method. The thread may or may not be
1178 * the current thread. It MUST be descheduled and in a stable state.
1179 * lwkt_giveaway() must be called on the cpu owning the thread.
1181 * At any point after lwkt_giveaway() is called, the target cpu may
1182 * 'pull' the thread by calling lwkt_acquire().
1184 * We have to make sure the thread is not sitting on a per-cpu tsleep
1185 * queue or it will blow up when it moves to another cpu.
1187 * MPSAFE - must be called under very specific conditions.
1190 lwkt_giveaway(thread_t td)
1192 globaldata_t gd = mycpu;
1195 if (td->td_flags & TDF_TSLEEPQ)
1197 KKASSERT(td->td_gd == gd);
1198 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1199 td->td_flags |= TDF_MIGRATING;
1204 lwkt_acquire(thread_t td)
1209 KKASSERT(td->td_flags & TDF_MIGRATING);
1214 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1215 crit_enter_gd(mygd);
1216 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1218 lwkt_process_ipiq();
1224 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1225 td->td_flags &= ~TDF_MIGRATING;
1228 crit_enter_gd(mygd);
1229 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1230 td->td_flags &= ~TDF_MIGRATING;
1238 * Generic deschedule. Descheduling threads other then your own should be
1239 * done only in carefully controlled circumstances. Descheduling is
1242 * This function may block if the cpu has run out of messages.
1245 lwkt_deschedule(thread_t td)
1249 if (td == curthread) {
1252 if (td->td_gd == mycpu) {
1255 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1265 * Set the target thread's priority. This routine does not automatically
1266 * switch to a higher priority thread, LWKT threads are not designed for
1267 * continuous priority changes. Yield if you want to switch.
1270 lwkt_setpri(thread_t td, int pri)
1272 KKASSERT(td->td_gd == mycpu);
1273 if (td->td_pri != pri) {
1276 if (td->td_flags & TDF_RUNQ) {
1288 * Set the initial priority for a thread prior to it being scheduled for
1289 * the first time. The thread MUST NOT be scheduled before or during
1290 * this call. The thread may be assigned to a cpu other then the current
1293 * Typically used after a thread has been created with TDF_STOPPREQ,
1294 * and before the thread is initially scheduled.
1297 lwkt_setpri_initial(thread_t td, int pri)
1300 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1305 lwkt_setpri_self(int pri)
1307 thread_t td = curthread;
1309 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1311 if (td->td_flags & TDF_RUNQ) {
1322 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1324 * Example: two competing threads, same priority N. decrement by (2*N)
1325 * increment by N*8, each thread will get 4 ticks.
1328 lwkt_fairq_schedulerclock(thread_t td)
1332 if (td != &td->td_gd->gd_idlethread) {
1333 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1334 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1335 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1336 if (td->td_fairq_accum < 0)
1337 need_lwkt_resched();
1338 td->td_fairq_lticks = ticks;
1340 td = td->td_preempted;
1346 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1348 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1349 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1350 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1354 * Migrate the current thread to the specified cpu.
1356 * This is accomplished by descheduling ourselves from the current cpu,
1357 * moving our thread to the tdallq of the target cpu, IPI messaging the
1358 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1359 * races while the thread is being migrated.
1361 * We must be sure to remove ourselves from the current cpu's tsleepq
1362 * before potentially moving to another queue. The thread can be on
1363 * a tsleepq due to a left-over tsleep_interlock().
1366 static void lwkt_setcpu_remote(void *arg);
1370 lwkt_setcpu_self(globaldata_t rgd)
1373 thread_t td = curthread;
1375 if (td->td_gd != rgd) {
1376 crit_enter_quick(td);
1377 if (td->td_flags & TDF_TSLEEPQ)
1379 td->td_flags |= TDF_MIGRATING;
1380 lwkt_deschedule_self(td);
1381 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1382 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1384 /* we are now on the target cpu */
1385 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1386 crit_exit_quick(td);
1392 lwkt_migratecpu(int cpuid)
1397 rgd = globaldata_find(cpuid);
1398 lwkt_setcpu_self(rgd);
1403 * Remote IPI for cpu migration (called while in a critical section so we
1404 * do not have to enter another one). The thread has already been moved to
1405 * our cpu's allq, but we must wait for the thread to be completely switched
1406 * out on the originating cpu before we schedule it on ours or the stack
1407 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1408 * change to main memory.
1410 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1411 * against wakeups. It is best if this interface is used only when there
1412 * are no pending events that might try to schedule the thread.
1416 lwkt_setcpu_remote(void *arg)
1419 globaldata_t gd = mycpu;
1421 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1423 lwkt_process_ipiq();
1430 td->td_flags &= ~TDF_MIGRATING;
1431 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1437 lwkt_preempted_proc(void)
1439 thread_t td = curthread;
1440 while (td->td_preempted)
1441 td = td->td_preempted;
1446 * Create a kernel process/thread/whatever. It shares it's address space
1447 * with proc0 - ie: kernel only.
1449 * NOTE! By default new threads are created with the MP lock held. A
1450 * thread which does not require the MP lock should release it by calling
1451 * rel_mplock() at the start of the new thread.
1454 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1455 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1460 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1464 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1467 * Set up arg0 for 'ps' etc
1469 __va_start(ap, fmt);
1470 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1474 * Schedule the thread to run
1476 if ((td->td_flags & TDF_STOPREQ) == 0)
1479 td->td_flags &= ~TDF_STOPREQ;
1484 * Destroy an LWKT thread. Warning! This function is not called when
1485 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1486 * uses a different reaping mechanism.
1491 thread_t td = curthread;
1496 * Do any cleanup that might block here
1498 if (td->td_flags & TDF_VERBOSE)
1499 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1502 dsched_exit_thread(td);
1505 * Get us into a critical section to interlock gd_freetd and loop
1506 * until we can get it freed.
1508 * We have to cache the current td in gd_freetd because objcache_put()ing
1509 * it would rip it out from under us while our thread is still active.
1512 crit_enter_quick(td);
1513 while ((std = gd->gd_freetd) != NULL) {
1514 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1515 gd->gd_freetd = NULL;
1516 objcache_put(thread_cache, std);
1520 * Remove thread resources from kernel lists and deschedule us for
1521 * the last time. We cannot block after this point or we may end
1522 * up with a stale td on the tsleepq.
1524 if (td->td_flags & TDF_TSLEEPQ)
1526 lwkt_deschedule_self(td);
1527 lwkt_remove_tdallq(td);
1532 KKASSERT(gd->gd_freetd == NULL);
1533 if (td->td_flags & TDF_ALLOCATED_THREAD)
1539 lwkt_remove_tdallq(thread_t td)
1541 KKASSERT(td->td_gd == mycpu);
1542 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1546 * Code reduction and branch prediction improvements. Call/return
1547 * overhead on modern cpus often degenerates into 0 cycles due to
1548 * the cpu's branch prediction hardware and return pc cache. We
1549 * can take advantage of this by not inlining medium-complexity
1550 * functions and we can also reduce the branch prediction impact
1551 * by collapsing perfectly predictable branches into a single
1552 * procedure instead of duplicating it.
1554 * Is any of this noticeable? Probably not, so I'll take the
1555 * smaller code size.
1558 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1560 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1566 thread_t td = curthread;
1567 int lcrit = td->td_critcount;
1569 td->td_critcount = 0;
1570 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1577 * Called from debugger/panic on cpus which have been stopped. We must still
1578 * process the IPIQ while stopped, even if we were stopped while in a critical
1581 * If we are dumping also try to process any pending interrupts. This may
1582 * or may not work depending on the state of the cpu at the point it was
1586 lwkt_smp_stopped(void)
1588 globaldata_t gd = mycpu;
1592 lwkt_process_ipiq();
1595 lwkt_process_ipiq();