2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.59 2004/04/10 20:55:23 dillon Exp $
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
42 #include <sys/rtprio.h>
43 #include <sys/queue.h>
44 #include <sys/thread2.h>
45 #include <sys/sysctl.h>
46 #include <sys/kthread.h>
47 #include <machine/cpu.h>
52 #include <vm/vm_param.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_zone.h>
61 #include <machine/stdarg.h>
62 #include <machine/ipl.h>
63 #include <machine/smp.h>
65 #define THREAD_STACK (UPAGES * PAGE_SIZE)
69 #include <sys/stdint.h>
70 #include <libcaps/thread.h>
71 #include <sys/thread.h>
72 #include <sys/msgport.h>
73 #include <sys/errno.h>
74 #include <libcaps/globaldata.h>
75 #include <sys/thread2.h>
76 #include <sys/msgport2.h>
80 #include <machine/cpufunc.h>
81 #include <machine/lock.h>
85 static int untimely_switch = 0;
87 static int panic_on_cscount = 0;
89 static __int64_t switch_count = 0;
90 static __int64_t preempt_hit = 0;
91 static __int64_t preempt_miss = 0;
92 static __int64_t preempt_weird = 0;
96 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
98 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
100 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
101 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
102 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
103 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
108 * These helper procedures handle the runq, they can only be called from
109 * within a critical section.
111 * WARNING! Prior to SMP being brought up it is possible to enqueue and
112 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
113 * instead of 'mycpu' when referencing the globaldata structure. Once
114 * SMP live enqueuing and dequeueing only occurs on the current cpu.
118 _lwkt_dequeue(thread_t td)
120 if (td->td_flags & TDF_RUNQ) {
121 int nq = td->td_pri & TDPRI_MASK;
122 struct globaldata *gd = td->td_gd;
124 td->td_flags &= ~TDF_RUNQ;
125 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
126 /* runqmask is passively cleaned up by the switcher */
132 _lwkt_enqueue(thread_t td)
134 if ((td->td_flags & TDF_RUNQ) == 0) {
135 int nq = td->td_pri & TDPRI_MASK;
136 struct globaldata *gd = td->td_gd;
138 td->td_flags |= TDF_RUNQ;
139 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
140 gd->gd_runqmask |= 1 << nq;
145 * Schedule a thread to run. As the current thread we can always safely
146 * schedule ourselves, and a shortcut procedure is provided for that
149 * (non-blocking, self contained on a per cpu basis)
152 lwkt_schedule_self(thread_t td)
154 crit_enter_quick(td);
155 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
156 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
159 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
160 panic("SCHED SELF PANIC");
166 * Deschedule a thread.
168 * (non-blocking, self contained on a per cpu basis)
171 lwkt_deschedule_self(thread_t td)
173 crit_enter_quick(td);
174 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
182 * LWKTs operate on a per-cpu basis
184 * WARNING! Called from early boot, 'mycpu' may not work yet.
187 lwkt_gdinit(struct globaldata *gd)
191 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
192 TAILQ_INIT(&gd->gd_tdrunq[i]);
194 TAILQ_INIT(&gd->gd_tdallq);
200 * Initialize a thread wait structure prior to first use.
202 * NOTE! called from low level boot code, we cannot do anything fancy!
205 lwkt_wait_init(lwkt_wait_t w)
207 lwkt_token_init(&w->wa_token);
208 TAILQ_INIT(&w->wa_waitq);
214 * Create a new thread. The thread must be associated with a process context
215 * or LWKT start address before it can be scheduled. If the target cpu is
216 * -1 the thread will be created on the current cpu.
218 * If you intend to create a thread without a process context this function
219 * does everything except load the startup and switcher function.
222 lwkt_alloc_thread(struct thread *td, int cpu)
226 globaldata_t gd = mycpu;
230 if (gd->gd_tdfreecount > 0) {
231 --gd->gd_tdfreecount;
232 td = TAILQ_FIRST(&gd->gd_tdfreeq);
233 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
234 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
235 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
237 stack = td->td_kstack;
238 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
242 td = zalloc(thread_zone);
244 td = malloc(sizeof(struct thread));
246 td->td_kstack = NULL;
247 flags |= TDF_ALLOCATED_THREAD;
250 if ((stack = td->td_kstack) == NULL) {
252 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
254 stack = libcaps_alloc_stack(THREAD_STACK);
256 flags |= TDF_ALLOCATED_STACK;
259 lwkt_init_thread(td, stack, flags, mycpu);
261 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
268 * Initialize a preexisting thread structure. This function is used by
269 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
271 * All threads start out in a critical section at a priority of
272 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
273 * appropriate. This function may send an IPI message when the
274 * requested cpu is not the current cpu and consequently gd_tdallq may
275 * not be initialized synchronously from the point of view of the originating
278 * NOTE! we have to be careful in regards to creating threads for other cpus
279 * if SMP has not yet been activated.
284 lwkt_init_thread_remote(void *arg)
288 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
294 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
296 globaldata_t mygd = mycpu;
298 bzero(td, sizeof(struct thread));
299 td->td_kstack = stack;
300 td->td_flags |= flags;
302 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
303 lwkt_initport(&td->td_msgport, td);
304 pmap_init_thread(td);
308 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
311 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
315 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
323 lwkt_set_comm(thread_t td, const char *ctl, ...)
328 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
333 lwkt_hold(thread_t td)
339 lwkt_rele(thread_t td)
341 KKASSERT(td->td_refs > 0);
348 lwkt_wait_free(thread_t td)
351 tsleep(td, 0, "tdreap", hz);
357 lwkt_free_thread(thread_t td)
359 struct globaldata *gd = mycpu;
361 KASSERT((td->td_flags & TDF_RUNNING) == 0,
362 ("lwkt_free_thread: did not exit! %p", td));
365 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
366 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
367 (td->td_flags & TDF_ALLOCATED_THREAD)
369 ++gd->gd_tdfreecount;
370 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
374 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
376 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
378 libcaps_free_stack(td->td_kstack, THREAD_STACK);
381 td->td_kstack = NULL;
383 if (td->td_flags & TDF_ALLOCATED_THREAD) {
385 zfree(thread_zone, td);
395 * Switch to the next runnable lwkt. If no LWKTs are runnable then
396 * switch to the idlethread. Switching must occur within a critical
397 * section to avoid races with the scheduling queue.
399 * We always have full control over our cpu's run queue. Other cpus
400 * that wish to manipulate our queue must use the cpu_*msg() calls to
401 * talk to our cpu, so a critical section is all that is needed and
402 * the result is very, very fast thread switching.
404 * The LWKT scheduler uses a fixed priority model and round-robins at
405 * each priority level. User process scheduling is a totally
406 * different beast and LWKT priorities should not be confused with
407 * user process priorities.
409 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
410 * cleans it up. Note that the td_switch() function cannot do anything that
411 * requires the MP lock since the MP lock will have already been setup for
412 * the target thread (not the current thread). It's nice to have a scheduler
413 * that does not need the MP lock to work because it allows us to do some
414 * really cool high-performance MP lock optimizations.
420 globaldata_t gd = mycpu;
421 thread_t td = gd->gd_curthread;
428 * Switching from within a 'fast' (non thread switched) interrupt is
431 if (gd->gd_intr_nesting_level && panicstr == NULL) {
432 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
436 * Passive release (used to transition from user to kernel mode
437 * when we block or switch rather then when we enter the kernel).
438 * This function is NOT called if we are switching into a preemption
439 * or returning from a preemption. Typically this causes us to lose
440 * our current process designation (if we have one) and become a true
441 * LWKT thread, and may also hand the current process designation to
442 * another process and schedule thread.
452 * td_mpcount cannot be used to determine if we currently hold the
453 * MP lock because get_mplock() will increment it prior to attempting
454 * to get the lock, and switch out if it can't. Our ownership of
455 * the actual lock will remain stable while we are in a critical section
456 * (but, of course, another cpu may own or release the lock so the
457 * actual value of mp_lock is not stable).
459 mpheld = MP_LOCK_HELD();
461 if (td->td_cscount) {
462 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
464 if (panic_on_cscount)
465 panic("switching while mastering cpusync");
469 if ((ntd = td->td_preempted) != NULL) {
471 * We had preempted another thread on this cpu, resume the preempted
472 * thread. This occurs transparently, whether the preempted thread
473 * was scheduled or not (it may have been preempted after descheduling
476 * We have to setup the MP lock for the original thread after backing
477 * out the adjustment that was made to curthread when the original
480 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
482 if (ntd->td_mpcount && mpheld == 0) {
483 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
484 td, ntd, td->td_mpcount, ntd->td_mpcount);
486 if (ntd->td_mpcount) {
487 td->td_mpcount -= ntd->td_mpcount;
488 KKASSERT(td->td_mpcount >= 0);
491 ntd->td_flags |= TDF_PREEMPT_DONE;
492 /* YYY release mp lock on switchback if original doesn't need it */
495 * Priority queue / round-robin at each priority. Note that user
496 * processes run at a fixed, low priority and the user process
497 * scheduler deals with interactions between user processes
498 * by scheduling and descheduling them from the LWKT queue as
501 * We have to adjust the MP lock for the target thread. If we
502 * need the MP lock and cannot obtain it we try to locate a
503 * thread that does not need the MP lock. If we cannot, we spin
506 * A similar issue exists for the tokens held by the target thread.
507 * If we cannot obtain ownership of the tokens we cannot immediately
508 * schedule the thread.
512 * We are switching threads. If there are any pending requests for
513 * tokens we can satisfy all of them here.
516 if (gd->gd_tokreqbase)
517 lwkt_drain_token_requests();
521 if (gd->gd_runqmask) {
522 int nq = bsrl(gd->gd_runqmask);
523 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
524 gd->gd_runqmask &= ~(1 << nq);
529 * If the target needs the MP lock and we couldn't get it,
530 * or if the target is holding tokens and we could not
531 * gain ownership of the tokens, continue looking for a
532 * thread to schedule and spin instead of HLT if we can't.
534 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
535 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
537 u_int32_t rqmask = gd->gd_runqmask;
539 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
540 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
542 mpheld = MP_LOCK_HELD();
543 if (ntd->td_toks && !lwkt_chktokens(ntd))
549 rqmask &= ~(1 << nq);
553 ntd = &gd->gd_idlethread;
554 ntd->td_flags |= TDF_IDLE_NOHLT;
556 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
557 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
560 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
561 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
564 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
565 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
569 * We have nothing to run but only let the idle loop halt
570 * the cpu if there are no pending interrupts.
572 ntd = &gd->gd_idlethread;
573 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
574 ntd->td_flags |= TDF_IDLE_NOHLT;
577 KASSERT(ntd->td_pri >= TDPRI_CRIT,
578 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
581 * Do the actual switch. If the new target does not need the MP lock
582 * and we are holding it, release the MP lock. If the new target requires
583 * the MP lock we have already acquired it for the target.
586 if (ntd->td_mpcount == 0 ) {
590 ASSERT_MP_LOCK_HELD();
595 /* NOTE: current cpu may have changed after switch */
600 * Request that the target thread preempt the current thread. Preemption
601 * only works under a specific set of conditions:
603 * - We are not preempting ourselves
604 * - The target thread is owned by the current cpu
605 * - We are not currently being preempted
606 * - The target is not currently being preempted
607 * - We are able to satisfy the target's MP lock requirements (if any).
609 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
610 * this is called via lwkt_schedule() through the td_preemptable callback.
611 * critpri is the managed critical priority that we should ignore in order
612 * to determine whether preemption is possible (aka usually just the crit
613 * priority of lwkt_schedule() itself).
615 * XXX at the moment we run the target thread in a critical section during
616 * the preemption in order to prevent the target from taking interrupts
617 * that *WE* can't. Preemption is strictly limited to interrupt threads
618 * and interrupt-like threads, outside of a critical section, and the
619 * preempted source thread will be resumed the instant the target blocks
620 * whether or not the source is scheduled (i.e. preemption is supposed to
621 * be as transparent as possible).
623 * The target thread inherits our MP count (added to its own) for the
624 * duration of the preemption in order to preserve the atomicy of the
625 * MP lock during the preemption. Therefore, any preempting targets must be
626 * careful in regards to MP assertions. Note that the MP count may be
627 * out of sync with the physical mp_lock, but we do not have to preserve
628 * the original ownership of the lock if it was out of synch (that is, we
629 * can leave it synchronized on return).
632 lwkt_preempt(thread_t ntd, int critpri)
634 struct globaldata *gd = mycpu;
642 * The caller has put us in a critical section. We can only preempt
643 * if the caller of the caller was not in a critical section (basically
644 * a local interrupt), as determined by the 'critpri' parameter.
646 * YYY The target thread must be in a critical section (else it must
647 * inherit our critical section? I dunno yet).
649 * Any tokens held by the target may not be held by thread(s) being
650 * preempted. We take the easy way out and do not preempt if
651 * the target is holding tokens.
653 * Set need_lwkt_resched() unconditionally for now YYY.
655 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
657 td = gd->gd_curthread;
659 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
663 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
668 if (ntd->td_gd != gd) {
674 * Take the easy way out and do not preempt if the target is holding
675 * one or more tokens. We could test whether the thread(s) being
676 * preempted interlock against the target thread's tokens and whether
677 * we can get all the target thread's tokens, but this situation
678 * should not occur very often so its easier to simply not preempt.
680 if (ntd->td_toks != NULL) {
684 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
688 if (ntd->td_preempted) {
694 * note: an interrupt might have occured just as we were transitioning
695 * to or from the MP lock. In this case td_mpcount will be pre-disposed
696 * (non-zero) but not actually synchronized with the actual state of the
697 * lock. We can use it to imply an MP lock requirement for the
698 * preemption but we cannot use it to test whether we hold the MP lock
701 savecnt = td->td_mpcount;
702 mpheld = MP_LOCK_HELD();
703 ntd->td_mpcount += td->td_mpcount;
704 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
705 ntd->td_mpcount -= td->td_mpcount;
712 ntd->td_preempted = td;
713 td->td_flags |= TDF_PREEMPT_LOCK;
715 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
717 KKASSERT(savecnt == td->td_mpcount);
718 mpheld = MP_LOCK_HELD();
719 if (mpheld && td->td_mpcount == 0)
721 else if (mpheld == 0 && td->td_mpcount)
722 panic("lwkt_preempt(): MP lock was not held through");
724 ntd->td_preempted = NULL;
725 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
729 * Yield our thread while higher priority threads are pending. This is
730 * typically called when we leave a critical section but it can be safely
731 * called while we are in a critical section.
733 * This function will not generally yield to equal priority threads but it
734 * can occur as a side effect. Note that lwkt_switch() is called from
735 * inside the critical section to prevent its own crit_exit() from reentering
736 * lwkt_yield_quick().
738 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
739 * came along but was blocked and made pending.
741 * (self contained on a per cpu basis)
744 lwkt_yield_quick(void)
746 globaldata_t gd = mycpu;
747 thread_t td = gd->gd_curthread;
750 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
751 * it with a non-zero cpl then we might not wind up calling splz after
752 * a task switch when the critical section is exited even though the
753 * new task could accept the interrupt.
755 * XXX from crit_exit() only called after last crit section is released.
756 * If called directly will run splz() even if in a critical section.
758 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
759 * except for this special case, we MUST call splz() here to handle any
760 * pending ints, particularly after we switch, or we might accidently
761 * halt the cpu with interrupts pending.
763 if (gd->gd_reqflags && td->td_nest_count < 2)
767 * YYY enabling will cause wakeup() to task-switch, which really
768 * confused the old 4.x code. This is a good way to simulate
769 * preemption and MP without actually doing preemption or MP, because a
770 * lot of code assumes that wakeup() does not block.
772 if (untimely_switch && td->td_nest_count == 0 &&
773 gd->gd_intr_nesting_level == 0
775 crit_enter_quick(td);
777 * YYY temporary hacks until we disassociate the userland scheduler
778 * from the LWKT scheduler.
780 if (td->td_flags & TDF_RUNQ) {
781 lwkt_switch(); /* will not reenter yield function */
783 lwkt_schedule_self(td); /* make sure we are scheduled */
784 lwkt_switch(); /* will not reenter yield function */
785 lwkt_deschedule_self(td); /* make sure we are descheduled */
787 crit_exit_noyield(td);
792 * This implements a normal yield which, unlike _quick, will yield to equal
793 * priority threads as well. Note that gd_reqflags tests will be handled by
794 * the crit_exit() call in lwkt_switch().
796 * (self contained on a per cpu basis)
801 lwkt_schedule_self(curthread);
806 * Generic schedule. Possibly schedule threads belonging to other cpus and
807 * deal with threads that might be blocked on a wait queue.
809 * We have a little helper inline function which does additional work after
810 * the thread has been enqueued, including dealing with preemption and
811 * setting need_lwkt_resched() (which prevents the kernel from returning
812 * to userland until it has processed higher priority threads).
816 _lwkt_schedule_post(thread_t ntd, int cpri)
818 if (ntd->td_preemptable) {
819 ntd->td_preemptable(ntd, cpri); /* YYY +token */
821 if ((ntd->td_flags & TDF_NORESCHED) == 0) {
822 if ((ntd->td_pri & TDPRI_MASK) >= TDPRI_KERN_USER)
829 lwkt_schedule(thread_t td)
831 globaldata_t mygd = mycpu;
834 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
835 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
836 && td->td_proc->p_stat == SSLEEP
838 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
840 curthread->td_proc ? curthread->td_proc->p_pid : -1,
841 curthread->td_proc ? curthread->td_proc->p_stat : -1,
843 td->td_proc ? curthread->td_proc->p_pid : -1,
844 td->td_proc ? curthread->td_proc->p_stat : -1
846 panic("SCHED PANIC");
850 if (td == mygd->gd_curthread) {
856 * If the thread is on a wait list we have to send our scheduling
857 * request to the owner of the wait structure. Otherwise we send
858 * the scheduling request to the cpu owning the thread. Races
859 * are ok, the target will forward the message as necessary (the
860 * message may chase the thread around before it finally gets
863 * (remember, wait structures use stable storage)
865 * NOTE: tokens no longer enter a critical section, so we only need
866 * to account for the crit_enter() above when calling
867 * _lwkt_schedule_post().
869 if ((w = td->td_wait) != NULL) {
872 if (lwkt_trytoken(&wref, &w->wa_token)) {
873 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
877 if (td->td_gd == mycpu) {
879 _lwkt_schedule_post(td, TDPRI_CRIT);
881 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
885 _lwkt_schedule_post(td, TDPRI_CRIT);
887 lwkt_reltoken(&wref);
889 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
893 * If the wait structure is NULL and we own the thread, there
894 * is no race (since we are in a critical section). If we
895 * do not own the thread there might be a race but the
896 * target cpu will deal with it.
899 if (td->td_gd == mygd) {
901 _lwkt_schedule_post(td, TDPRI_CRIT);
903 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
907 _lwkt_schedule_post(td, TDPRI_CRIT);
915 * Managed acquisition. This code assumes that the MP lock is held for
916 * the tdallq operation and that the thread has been descheduled from its
917 * original cpu. We also have to wait for the thread to be entirely switched
918 * out on its original cpu (this is usually fast enough that we never loop)
919 * since the LWKT system does not have to hold the MP lock while switching
920 * and the target may have released it before switching.
923 lwkt_acquire(thread_t td)
930 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
931 while (td->td_flags & TDF_RUNNING) /* XXX spin */
935 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
937 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
943 * Generic deschedule. Descheduling threads other then your own should be
944 * done only in carefully controlled circumstances. Descheduling is
947 * This function may block if the cpu has run out of messages.
950 lwkt_deschedule(thread_t td)
953 if (td == curthread) {
956 if (td->td_gd == mycpu) {
959 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
966 * Set the target thread's priority. This routine does not automatically
967 * switch to a higher priority thread, LWKT threads are not designed for
968 * continuous priority changes. Yield if you want to switch.
970 * We have to retain the critical section count which uses the high bits
971 * of the td_pri field. The specified priority may also indicate zero or
972 * more critical sections by adding TDPRI_CRIT*N.
974 * Note that we requeue the thread whether it winds up on a different runq
975 * or not. uio_yield() depends on this and the routine is not normally
976 * called with the same priority otherwise.
979 lwkt_setpri(thread_t td, int pri)
982 KKASSERT(td->td_gd == mycpu);
984 if (td->td_flags & TDF_RUNQ) {
986 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
989 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
995 lwkt_setpri_self(int pri)
997 thread_t td = curthread;
999 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1001 if (td->td_flags & TDF_RUNQ) {
1003 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1006 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1012 lwkt_preempted_proc(void)
1014 thread_t td = curthread;
1015 while (td->td_preempted)
1016 td = td->td_preempted;
1017 return(td->td_proc);
1021 * Block on the specified wait queue until signaled. A generation number
1022 * must be supplied to interlock the wait queue. The function will
1023 * return immediately if the generation number does not match the wait
1024 * structure's generation number.
1027 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1029 thread_t td = curthread;
1032 lwkt_gettoken(&ilock, &w->wa_token);
1034 if (w->wa_gen == *gen) {
1036 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1039 td->td_wmesg = wmesg;
1042 if (td->td_wmesg != NULL) {
1049 lwkt_reltoken(&ilock);
1053 * Signal a wait queue. We gain ownership of the wait queue in order to
1054 * signal it. Once a thread is removed from the wait queue we have to
1055 * deal with the cpu owning the thread.
1057 * Note: alternatively we could message the target cpu owning the wait
1058 * queue. YYY implement as sysctl.
1061 lwkt_signal(lwkt_wait_t w, int count)
1066 lwkt_gettoken(&ilock, &w->wa_token);
1070 count = w->wa_count;
1071 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1074 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1076 td->td_wmesg = NULL;
1077 if (td->td_gd == mycpu) {
1080 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1084 lwkt_reltoken(&ilock);
1088 * Create a kernel process/thread/whatever. It shares it's address space
1089 * with proc0 - ie: kernel only.
1091 * NOTE! By default new threads are created with the MP lock held. A
1092 * thread which does not require the MP lock should release it by calling
1093 * rel_mplock() at the start of the new thread.
1096 lwkt_create(void (*func)(void *), void *arg,
1097 struct thread **tdp, thread_t template, int tdflags, int cpu,
1098 const char *fmt, ...)
1103 td = lwkt_alloc_thread(template, cpu);
1106 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1107 td->td_flags |= TDF_VERBOSE | tdflags;
1113 * Set up arg0 for 'ps' etc
1115 __va_start(ap, fmt);
1116 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1120 * Schedule the thread to run
1122 if ((td->td_flags & TDF_STOPREQ) == 0)
1125 td->td_flags &= ~TDF_STOPREQ;
1130 * kthread_* is specific to the kernel and is not needed by userland.
1135 * Destroy an LWKT thread. Warning! This function is not called when
1136 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1137 * uses a different reaping mechanism.
1142 thread_t td = curthread;
1144 if (td->td_flags & TDF_VERBOSE)
1145 printf("kthread %p %s has exited\n", td, td->td_comm);
1147 crit_enter_quick(td);
1148 lwkt_deschedule_self(td);
1149 ++mycpu->gd_tdfreecount;
1150 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1155 * Create a kernel process/thread/whatever. It shares it's address space
1156 * with proc0 - ie: kernel only. 5.x compatible.
1158 * NOTE! By default kthreads are created with the MP lock held. A
1159 * thread which does not require the MP lock should release it by calling
1160 * rel_mplock() at the start of the new thread.
1163 kthread_create(void (*func)(void *), void *arg,
1164 struct thread **tdp, const char *fmt, ...)
1169 td = lwkt_alloc_thread(NULL, -1);
1172 cpu_set_thread_handler(td, kthread_exit, func, arg);
1173 td->td_flags |= TDF_VERBOSE;
1179 * Set up arg0 for 'ps' etc
1181 __va_start(ap, fmt);
1182 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1186 * Schedule the thread to run
1193 * Destroy an LWKT thread. Warning! This function is not called when
1194 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1195 * uses a different reaping mechanism.
1197 * XXX duplicates lwkt_exit()
1205 #endif /* _KERNEL */
1210 thread_t td = curthread;
1211 int lpri = td->td_pri;
1214 panic("td_pri is/would-go negative! %p %d", td, lpri);