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 1.62 2004/06/03 13:09:07 joerg 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|TDF_MIGRATING)) == 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);
307 * Normally initializing a thread for a remote cpu requires sending an
308 * IPI. However, the idlethread is setup before the other cpus are
309 * activated so we have to treat it as a special case. XXX manipulation
310 * of gd_tdallq requires the BGL.
312 if (gd == mygd || td == &gd->gd_idlethread) {
314 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
317 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
321 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
329 lwkt_set_comm(thread_t td, const char *ctl, ...)
334 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
339 lwkt_hold(thread_t td)
345 lwkt_rele(thread_t td)
347 KKASSERT(td->td_refs > 0);
354 lwkt_wait_free(thread_t td)
357 tsleep(td, 0, "tdreap", hz);
363 lwkt_free_thread(thread_t td)
365 struct globaldata *gd = mycpu;
367 KASSERT((td->td_flags & TDF_RUNNING) == 0,
368 ("lwkt_free_thread: did not exit! %p", td));
371 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
372 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
373 (td->td_flags & TDF_ALLOCATED_THREAD)
375 ++gd->gd_tdfreecount;
376 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
380 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
382 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
384 libcaps_free_stack(td->td_kstack, THREAD_STACK);
387 td->td_kstack = NULL;
389 if (td->td_flags & TDF_ALLOCATED_THREAD) {
391 zfree(thread_zone, td);
401 * Switch to the next runnable lwkt. If no LWKTs are runnable then
402 * switch to the idlethread. Switching must occur within a critical
403 * section to avoid races with the scheduling queue.
405 * We always have full control over our cpu's run queue. Other cpus
406 * that wish to manipulate our queue must use the cpu_*msg() calls to
407 * talk to our cpu, so a critical section is all that is needed and
408 * the result is very, very fast thread switching.
410 * The LWKT scheduler uses a fixed priority model and round-robins at
411 * each priority level. User process scheduling is a totally
412 * different beast and LWKT priorities should not be confused with
413 * user process priorities.
415 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
416 * cleans it up. Note that the td_switch() function cannot do anything that
417 * requires the MP lock since the MP lock will have already been setup for
418 * the target thread (not the current thread). It's nice to have a scheduler
419 * that does not need the MP lock to work because it allows us to do some
420 * really cool high-performance MP lock optimizations.
426 globaldata_t gd = mycpu;
427 thread_t td = gd->gd_curthread;
434 * Switching from within a 'fast' (non thread switched) interrupt is
437 if (gd->gd_intr_nesting_level && panicstr == NULL) {
438 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
442 * Passive release (used to transition from user to kernel mode
443 * when we block or switch rather then when we enter the kernel).
444 * This function is NOT called if we are switching into a preemption
445 * or returning from a preemption. Typically this causes us to lose
446 * our current process designation (if we have one) and become a true
447 * LWKT thread, and may also hand the current process designation to
448 * another process and schedule thread.
458 * td_mpcount cannot be used to determine if we currently hold the
459 * MP lock because get_mplock() will increment it prior to attempting
460 * to get the lock, and switch out if it can't. Our ownership of
461 * the actual lock will remain stable while we are in a critical section
462 * (but, of course, another cpu may own or release the lock so the
463 * actual value of mp_lock is not stable).
465 mpheld = MP_LOCK_HELD();
467 if (td->td_cscount) {
468 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
470 if (panic_on_cscount)
471 panic("switching while mastering cpusync");
475 if ((ntd = td->td_preempted) != NULL) {
477 * We had preempted another thread on this cpu, resume the preempted
478 * thread. This occurs transparently, whether the preempted thread
479 * was scheduled or not (it may have been preempted after descheduling
482 * We have to setup the MP lock for the original thread after backing
483 * out the adjustment that was made to curthread when the original
486 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
488 if (ntd->td_mpcount && mpheld == 0) {
489 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
490 td, ntd, td->td_mpcount, ntd->td_mpcount);
492 if (ntd->td_mpcount) {
493 td->td_mpcount -= ntd->td_mpcount;
494 KKASSERT(td->td_mpcount >= 0);
497 ntd->td_flags |= TDF_PREEMPT_DONE;
498 /* YYY release mp lock on switchback if original doesn't need it */
501 * Priority queue / round-robin at each priority. Note that user
502 * processes run at a fixed, low priority and the user process
503 * scheduler deals with interactions between user processes
504 * by scheduling and descheduling them from the LWKT queue as
507 * We have to adjust the MP lock for the target thread. If we
508 * need the MP lock and cannot obtain it we try to locate a
509 * thread that does not need the MP lock. If we cannot, we spin
512 * A similar issue exists for the tokens held by the target thread.
513 * If we cannot obtain ownership of the tokens we cannot immediately
514 * schedule the thread.
518 * We are switching threads. If there are any pending requests for
519 * tokens we can satisfy all of them here.
522 if (gd->gd_tokreqbase)
523 lwkt_drain_token_requests();
527 if (gd->gd_runqmask) {
528 int nq = bsrl(gd->gd_runqmask);
529 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
530 gd->gd_runqmask &= ~(1 << nq);
535 * If the target needs the MP lock and we couldn't get it,
536 * or if the target is holding tokens and we could not
537 * gain ownership of the tokens, continue looking for a
538 * thread to schedule and spin instead of HLT if we can't.
540 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
541 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
543 u_int32_t rqmask = gd->gd_runqmask;
545 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
546 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
548 mpheld = MP_LOCK_HELD();
549 if (ntd->td_toks && !lwkt_chktokens(ntd))
555 rqmask &= ~(1 << nq);
559 ntd = &gd->gd_idlethread;
560 ntd->td_flags |= TDF_IDLE_NOHLT;
562 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
563 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
566 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
567 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
570 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
571 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
575 * We have nothing to run but only let the idle loop halt
576 * the cpu if there are no pending interrupts.
578 ntd = &gd->gd_idlethread;
579 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
580 ntd->td_flags |= TDF_IDLE_NOHLT;
583 KASSERT(ntd->td_pri >= TDPRI_CRIT,
584 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
587 * Do the actual switch. If the new target does not need the MP lock
588 * and we are holding it, release the MP lock. If the new target requires
589 * the MP lock we have already acquired it for the target.
592 if (ntd->td_mpcount == 0 ) {
596 ASSERT_MP_LOCK_HELD();
601 /* NOTE: current cpu may have changed after switch */
606 * Request that the target thread preempt the current thread. Preemption
607 * only works under a specific set of conditions:
609 * - We are not preempting ourselves
610 * - The target thread is owned by the current cpu
611 * - We are not currently being preempted
612 * - The target is not currently being preempted
613 * - We are able to satisfy the target's MP lock requirements (if any).
615 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
616 * this is called via lwkt_schedule() through the td_preemptable callback.
617 * critpri is the managed critical priority that we should ignore in order
618 * to determine whether preemption is possible (aka usually just the crit
619 * priority of lwkt_schedule() itself).
621 * XXX at the moment we run the target thread in a critical section during
622 * the preemption in order to prevent the target from taking interrupts
623 * that *WE* can't. Preemption is strictly limited to interrupt threads
624 * and interrupt-like threads, outside of a critical section, and the
625 * preempted source thread will be resumed the instant the target blocks
626 * whether or not the source is scheduled (i.e. preemption is supposed to
627 * be as transparent as possible).
629 * The target thread inherits our MP count (added to its own) for the
630 * duration of the preemption in order to preserve the atomicy of the
631 * MP lock during the preemption. Therefore, any preempting targets must be
632 * careful in regards to MP assertions. Note that the MP count may be
633 * out of sync with the physical mp_lock, but we do not have to preserve
634 * the original ownership of the lock if it was out of synch (that is, we
635 * can leave it synchronized on return).
638 lwkt_preempt(thread_t ntd, int critpri)
640 struct globaldata *gd = mycpu;
648 * The caller has put us in a critical section. We can only preempt
649 * if the caller of the caller was not in a critical section (basically
650 * a local interrupt), as determined by the 'critpri' parameter.
652 * YYY The target thread must be in a critical section (else it must
653 * inherit our critical section? I dunno yet).
655 * Any tokens held by the target may not be held by thread(s) being
656 * preempted. We take the easy way out and do not preempt if
657 * the target is holding tokens.
659 * Set need_lwkt_resched() unconditionally for now YYY.
661 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
663 td = gd->gd_curthread;
665 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
669 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
674 if (ntd->td_gd != gd) {
680 * Take the easy way out and do not preempt if the target is holding
681 * one or more tokens. We could test whether the thread(s) being
682 * preempted interlock against the target thread's tokens and whether
683 * we can get all the target thread's tokens, but this situation
684 * should not occur very often so its easier to simply not preempt.
686 if (ntd->td_toks != NULL) {
690 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
694 if (ntd->td_preempted) {
700 * note: an interrupt might have occured just as we were transitioning
701 * to or from the MP lock. In this case td_mpcount will be pre-disposed
702 * (non-zero) but not actually synchronized with the actual state of the
703 * lock. We can use it to imply an MP lock requirement for the
704 * preemption but we cannot use it to test whether we hold the MP lock
707 savecnt = td->td_mpcount;
708 mpheld = MP_LOCK_HELD();
709 ntd->td_mpcount += td->td_mpcount;
710 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
711 ntd->td_mpcount -= td->td_mpcount;
718 ntd->td_preempted = td;
719 td->td_flags |= TDF_PREEMPT_LOCK;
721 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
723 KKASSERT(savecnt == td->td_mpcount);
724 mpheld = MP_LOCK_HELD();
725 if (mpheld && td->td_mpcount == 0)
727 else if (mpheld == 0 && td->td_mpcount)
728 panic("lwkt_preempt(): MP lock was not held through");
730 ntd->td_preempted = NULL;
731 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
735 * Yield our thread while higher priority threads are pending. This is
736 * typically called when we leave a critical section but it can be safely
737 * called while we are in a critical section.
739 * This function will not generally yield to equal priority threads but it
740 * can occur as a side effect. Note that lwkt_switch() is called from
741 * inside the critical section to prevent its own crit_exit() from reentering
742 * lwkt_yield_quick().
744 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
745 * came along but was blocked and made pending.
747 * (self contained on a per cpu basis)
750 lwkt_yield_quick(void)
752 globaldata_t gd = mycpu;
753 thread_t td = gd->gd_curthread;
756 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
757 * it with a non-zero cpl then we might not wind up calling splz after
758 * a task switch when the critical section is exited even though the
759 * new task could accept the interrupt.
761 * XXX from crit_exit() only called after last crit section is released.
762 * If called directly will run splz() even if in a critical section.
764 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
765 * except for this special case, we MUST call splz() here to handle any
766 * pending ints, particularly after we switch, or we might accidently
767 * halt the cpu with interrupts pending.
769 if (gd->gd_reqflags && td->td_nest_count < 2)
773 * YYY enabling will cause wakeup() to task-switch, which really
774 * confused the old 4.x code. This is a good way to simulate
775 * preemption and MP without actually doing preemption or MP, because a
776 * lot of code assumes that wakeup() does not block.
778 if (untimely_switch && td->td_nest_count == 0 &&
779 gd->gd_intr_nesting_level == 0
781 crit_enter_quick(td);
783 * YYY temporary hacks until we disassociate the userland scheduler
784 * from the LWKT scheduler.
786 if (td->td_flags & TDF_RUNQ) {
787 lwkt_switch(); /* will not reenter yield function */
789 lwkt_schedule_self(td); /* make sure we are scheduled */
790 lwkt_switch(); /* will not reenter yield function */
791 lwkt_deschedule_self(td); /* make sure we are descheduled */
793 crit_exit_noyield(td);
798 * This implements a normal yield which, unlike _quick, will yield to equal
799 * priority threads as well. Note that gd_reqflags tests will be handled by
800 * the crit_exit() call in lwkt_switch().
802 * (self contained on a per cpu basis)
807 lwkt_schedule_self(curthread);
812 * Generic schedule. Possibly schedule threads belonging to other cpus and
813 * deal with threads that might be blocked on a wait queue.
815 * We have a little helper inline function which does additional work after
816 * the thread has been enqueued, including dealing with preemption and
817 * setting need_lwkt_resched() (which prevents the kernel from returning
818 * to userland until it has processed higher priority threads).
822 _lwkt_schedule_post(thread_t ntd, int cpri)
824 if (ntd->td_preemptable) {
825 ntd->td_preemptable(ntd, cpri); /* YYY +token */
827 if ((ntd->td_flags & TDF_NORESCHED) == 0) {
828 if ((ntd->td_pri & TDPRI_MASK) >= TDPRI_KERN_USER)
835 lwkt_schedule(thread_t td)
837 globaldata_t mygd = mycpu;
840 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
841 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
842 && td->td_proc->p_stat == SSLEEP
844 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
846 curthread->td_proc ? curthread->td_proc->p_pid : -1,
847 curthread->td_proc ? curthread->td_proc->p_stat : -1,
849 td->td_proc ? curthread->td_proc->p_pid : -1,
850 td->td_proc ? curthread->td_proc->p_stat : -1
852 panic("SCHED PANIC");
856 if (td == mygd->gd_curthread) {
862 * If the thread is on a wait list we have to send our scheduling
863 * request to the owner of the wait structure. Otherwise we send
864 * the scheduling request to the cpu owning the thread. Races
865 * are ok, the target will forward the message as necessary (the
866 * message may chase the thread around before it finally gets
869 * (remember, wait structures use stable storage)
871 * NOTE: tokens no longer enter a critical section, so we only need
872 * to account for the crit_enter() above when calling
873 * _lwkt_schedule_post().
875 if ((w = td->td_wait) != NULL) {
878 if (lwkt_trytoken(&wref, &w->wa_token)) {
879 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
883 if (td->td_gd == mycpu) {
885 _lwkt_schedule_post(td, TDPRI_CRIT);
887 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
891 _lwkt_schedule_post(td, TDPRI_CRIT);
893 lwkt_reltoken(&wref);
895 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
899 * If the wait structure is NULL and we own the thread, there
900 * is no race (since we are in a critical section). If we
901 * do not own the thread there might be a race but the
902 * target cpu will deal with it.
905 if (td->td_gd == mygd) {
907 _lwkt_schedule_post(td, TDPRI_CRIT);
909 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
913 _lwkt_schedule_post(td, TDPRI_CRIT);
921 * Managed acquisition. This code assumes that the MP lock is held for
922 * the tdallq operation and that the thread has been descheduled from its
923 * original cpu. We also have to wait for the thread to be entirely switched
924 * out on its original cpu (this is usually fast enough that we never loop)
925 * since the LWKT system does not have to hold the MP lock while switching
926 * and the target may have released it before switching.
929 lwkt_acquire(thread_t td)
936 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
937 while (td->td_flags & TDF_RUNNING) /* XXX spin */
941 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
943 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
949 * Generic deschedule. Descheduling threads other then your own should be
950 * done only in carefully controlled circumstances. Descheduling is
953 * This function may block if the cpu has run out of messages.
956 lwkt_deschedule(thread_t td)
959 if (td == curthread) {
962 if (td->td_gd == mycpu) {
965 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
972 * Set the target thread's priority. This routine does not automatically
973 * switch to a higher priority thread, LWKT threads are not designed for
974 * continuous priority changes. Yield if you want to switch.
976 * We have to retain the critical section count which uses the high bits
977 * of the td_pri field. The specified priority may also indicate zero or
978 * more critical sections by adding TDPRI_CRIT*N.
980 * Note that we requeue the thread whether it winds up on a different runq
981 * or not. uio_yield() depends on this and the routine is not normally
982 * called with the same priority otherwise.
985 lwkt_setpri(thread_t td, int pri)
988 KKASSERT(td->td_gd == mycpu);
990 if (td->td_flags & TDF_RUNQ) {
992 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
995 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1001 lwkt_setpri_self(int pri)
1003 thread_t td = curthread;
1005 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1007 if (td->td_flags & TDF_RUNQ) {
1009 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1012 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1018 * Migrate the current thread to the specified cpu. The BGL must be held
1019 * (for the gd_tdallq manipulation XXX). This is accomplished by
1020 * descheduling ourselves from the current cpu, moving our thread to the
1021 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1022 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1025 static void lwkt_setcpu_remote(void *arg);
1029 lwkt_setcpu_self(globaldata_t rgd)
1032 thread_t td = curthread;
1034 if (td->td_gd != rgd) {
1035 crit_enter_quick(td);
1036 td->td_flags |= TDF_MIGRATING;
1037 lwkt_deschedule_self(td);
1038 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1039 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1040 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1042 /* we are now on the target cpu */
1043 crit_exit_quick(td);
1049 * Remote IPI for cpu migration (called while in a critical section so we
1050 * do not have to enter another one). The thread has already been moved to
1051 * our cpu's allq, but we must wait for the thread to be completely switched
1052 * out on the originating cpu before we schedule it on ours or the stack
1053 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1054 * change to main memory.
1056 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1057 * against wakeups. It is best if this interface is used only when there
1058 * are no pending events that might try to schedule the thread.
1062 lwkt_setcpu_remote(void *arg)
1065 globaldata_t gd = mycpu;
1067 while (td->td_flags & TDF_RUNNING)
1071 td->td_flags &= ~TDF_MIGRATING;
1077 lwkt_preempted_proc(void)
1079 thread_t td = curthread;
1080 while (td->td_preempted)
1081 td = td->td_preempted;
1082 return(td->td_proc);
1086 * Block on the specified wait queue until signaled. A generation number
1087 * must be supplied to interlock the wait queue. The function will
1088 * return immediately if the generation number does not match the wait
1089 * structure's generation number.
1092 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1094 thread_t td = curthread;
1097 lwkt_gettoken(&ilock, &w->wa_token);
1099 if (w->wa_gen == *gen) {
1101 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1104 td->td_wmesg = wmesg;
1107 if (td->td_wmesg != NULL) {
1114 lwkt_reltoken(&ilock);
1118 * Signal a wait queue. We gain ownership of the wait queue in order to
1119 * signal it. Once a thread is removed from the wait queue we have to
1120 * deal with the cpu owning the thread.
1122 * Note: alternatively we could message the target cpu owning the wait
1123 * queue. YYY implement as sysctl.
1126 lwkt_signal(lwkt_wait_t w, int count)
1131 lwkt_gettoken(&ilock, &w->wa_token);
1135 count = w->wa_count;
1136 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1139 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1141 td->td_wmesg = NULL;
1142 if (td->td_gd == mycpu) {
1145 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1149 lwkt_reltoken(&ilock);
1153 * Create a kernel process/thread/whatever. It shares it's address space
1154 * with proc0 - ie: kernel only.
1156 * NOTE! By default new threads are created with the MP lock held. A
1157 * thread which does not require the MP lock should release it by calling
1158 * rel_mplock() at the start of the new thread.
1161 lwkt_create(void (*func)(void *), void *arg,
1162 struct thread **tdp, thread_t template, int tdflags, int cpu,
1163 const char *fmt, ...)
1168 td = lwkt_alloc_thread(template, cpu);
1171 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1172 td->td_flags |= TDF_VERBOSE | tdflags;
1178 * Set up arg0 for 'ps' etc
1180 __va_start(ap, fmt);
1181 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1185 * Schedule the thread to run
1187 if ((td->td_flags & TDF_STOPREQ) == 0)
1190 td->td_flags &= ~TDF_STOPREQ;
1195 * kthread_* is specific to the kernel and is not needed by userland.
1200 * Destroy an LWKT thread. Warning! This function is not called when
1201 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1202 * uses a different reaping mechanism.
1207 thread_t td = curthread;
1209 if (td->td_flags & TDF_VERBOSE)
1210 printf("kthread %p %s has exited\n", td, td->td_comm);
1212 crit_enter_quick(td);
1213 lwkt_deschedule_self(td);
1214 ++mycpu->gd_tdfreecount;
1215 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1220 * Create a kernel process/thread/whatever. It shares it's address space
1221 * with proc0 - ie: kernel only. 5.x compatible.
1223 * NOTE! By default kthreads are created with the MP lock held. A
1224 * thread which does not require the MP lock should release it by calling
1225 * rel_mplock() at the start of the new thread.
1228 kthread_create(void (*func)(void *), void *arg,
1229 struct thread **tdp, const char *fmt, ...)
1234 td = lwkt_alloc_thread(NULL, -1);
1237 cpu_set_thread_handler(td, kthread_exit, func, arg);
1238 td->td_flags |= TDF_VERBOSE;
1244 * Set up arg0 for 'ps' etc
1246 __va_start(ap, fmt);
1247 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1251 * Schedule the thread to run
1258 * Destroy an LWKT thread. Warning! This function is not called when
1259 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1260 * uses a different reaping mechanism.
1262 * XXX duplicates lwkt_exit()
1270 #endif /* _KERNEL */
1275 thread_t td = curthread;
1276 int lpri = td->td_pri;
1279 panic("td_pri is/would-go negative! %p %d", td, lpri);