2 * Copyright (c) 2003,2004 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
34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.66 2004/07/24 20:21:35 dillon Exp $
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
50 #include <sys/rtprio.h>
51 #include <sys/queue.h>
52 #include <sys/thread2.h>
53 #include <sys/sysctl.h>
54 #include <sys/kthread.h>
55 #include <machine/cpu.h>
60 #include <vm/vm_param.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_pager.h>
66 #include <vm/vm_extern.h>
67 #include <vm/vm_zone.h>
69 #include <machine/stdarg.h>
70 #include <machine/ipl.h>
71 #include <machine/smp.h>
73 #define THREAD_STACK (UPAGES * PAGE_SIZE)
77 #include <sys/stdint.h>
78 #include <libcaps/thread.h>
79 #include <sys/thread.h>
80 #include <sys/msgport.h>
81 #include <sys/errno.h>
82 #include <libcaps/globaldata.h>
83 #include <machine/cpufunc.h>
84 #include <sys/thread2.h>
85 #include <sys/msgport2.h>
89 #include <machine/lock.h>
93 static int untimely_switch = 0;
95 static int panic_on_cscount = 0;
97 static __int64_t switch_count = 0;
98 static __int64_t preempt_hit = 0;
99 static __int64_t preempt_miss = 0;
100 static __int64_t preempt_weird = 0;
104 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
106 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
109 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
116 * These helper procedures handle the runq, they can only be called from
117 * within a critical section.
119 * WARNING! Prior to SMP being brought up it is possible to enqueue and
120 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
121 * instead of 'mycpu' when referencing the globaldata structure. Once
122 * SMP live enqueuing and dequeueing only occurs on the current cpu.
126 _lwkt_dequeue(thread_t td)
128 if (td->td_flags & TDF_RUNQ) {
129 int nq = td->td_pri & TDPRI_MASK;
130 struct globaldata *gd = td->td_gd;
132 td->td_flags &= ~TDF_RUNQ;
133 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
134 /* runqmask is passively cleaned up by the switcher */
140 _lwkt_enqueue(thread_t td)
142 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
143 int nq = td->td_pri & TDPRI_MASK;
144 struct globaldata *gd = td->td_gd;
146 td->td_flags |= TDF_RUNQ;
147 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
148 gd->gd_runqmask |= 1 << nq;
153 * Schedule a thread to run. As the current thread we can always safely
154 * schedule ourselves, and a shortcut procedure is provided for that
157 * (non-blocking, self contained on a per cpu basis)
160 lwkt_schedule_self(thread_t td)
162 crit_enter_quick(td);
163 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
164 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
167 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
168 panic("SCHED SELF PANIC");
174 * Deschedule a thread.
176 * (non-blocking, self contained on a per cpu basis)
179 lwkt_deschedule_self(thread_t td)
181 crit_enter_quick(td);
182 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
190 * LWKTs operate on a per-cpu basis
192 * WARNING! Called from early boot, 'mycpu' may not work yet.
195 lwkt_gdinit(struct globaldata *gd)
199 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
200 TAILQ_INIT(&gd->gd_tdrunq[i]);
202 TAILQ_INIT(&gd->gd_tdallq);
208 * Initialize a thread wait structure prior to first use.
210 * NOTE! called from low level boot code, we cannot do anything fancy!
213 lwkt_wait_init(lwkt_wait_t w)
215 lwkt_token_init(&w->wa_token);
216 TAILQ_INIT(&w->wa_waitq);
222 * Create a new thread. The thread must be associated with a process context
223 * or LWKT start address before it can be scheduled. If the target cpu is
224 * -1 the thread will be created on the current cpu.
226 * If you intend to create a thread without a process context this function
227 * does everything except load the startup and switcher function.
230 lwkt_alloc_thread(struct thread *td, int cpu)
234 globaldata_t gd = mycpu;
238 if (gd->gd_tdfreecount > 0) {
239 --gd->gd_tdfreecount;
240 td = TAILQ_FIRST(&gd->gd_tdfreeq);
241 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
242 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
243 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
245 stack = td->td_kstack;
246 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
250 td = zalloc(thread_zone);
252 td = malloc(sizeof(struct thread));
254 td->td_kstack = NULL;
255 flags |= TDF_ALLOCATED_THREAD;
258 if ((stack = td->td_kstack) == NULL) {
260 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
262 stack = libcaps_alloc_stack(THREAD_STACK);
264 flags |= TDF_ALLOCATED_STACK;
267 lwkt_init_thread(td, stack, flags, mycpu);
269 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
276 * Initialize a preexisting thread structure. This function is used by
277 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
279 * All threads start out in a critical section at a priority of
280 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
281 * appropriate. This function may send an IPI message when the
282 * requested cpu is not the current cpu and consequently gd_tdallq may
283 * not be initialized synchronously from the point of view of the originating
286 * NOTE! we have to be careful in regards to creating threads for other cpus
287 * if SMP has not yet been activated.
292 lwkt_init_thread_remote(void *arg)
296 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
302 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
304 globaldata_t mygd = mycpu;
306 bzero(td, sizeof(struct thread));
307 td->td_kstack = stack;
308 td->td_flags |= flags;
310 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
311 lwkt_initport(&td->td_msgport, td);
312 pmap_init_thread(td);
315 * Normally initializing a thread for a remote cpu requires sending an
316 * IPI. However, the idlethread is setup before the other cpus are
317 * activated so we have to treat it as a special case. XXX manipulation
318 * of gd_tdallq requires the BGL.
320 if (gd == mygd || td == &gd->gd_idlethread) {
322 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
325 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
329 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
337 lwkt_set_comm(thread_t td, const char *ctl, ...)
342 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
347 lwkt_hold(thread_t td)
353 lwkt_rele(thread_t td)
355 KKASSERT(td->td_refs > 0);
362 lwkt_wait_free(thread_t td)
365 tsleep(td, 0, "tdreap", hz);
371 lwkt_free_thread(thread_t td)
373 struct globaldata *gd = mycpu;
375 KASSERT((td->td_flags & TDF_RUNNING) == 0,
376 ("lwkt_free_thread: did not exit! %p", td));
379 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
380 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
381 (td->td_flags & TDF_ALLOCATED_THREAD)
383 ++gd->gd_tdfreecount;
384 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
388 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
390 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
392 libcaps_free_stack(td->td_kstack, THREAD_STACK);
395 td->td_kstack = NULL;
397 if (td->td_flags & TDF_ALLOCATED_THREAD) {
399 zfree(thread_zone, td);
409 * Switch to the next runnable lwkt. If no LWKTs are runnable then
410 * switch to the idlethread. Switching must occur within a critical
411 * section to avoid races with the scheduling queue.
413 * We always have full control over our cpu's run queue. Other cpus
414 * that wish to manipulate our queue must use the cpu_*msg() calls to
415 * talk to our cpu, so a critical section is all that is needed and
416 * the result is very, very fast thread switching.
418 * The LWKT scheduler uses a fixed priority model and round-robins at
419 * each priority level. User process scheduling is a totally
420 * different beast and LWKT priorities should not be confused with
421 * user process priorities.
423 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
424 * cleans it up. Note that the td_switch() function cannot do anything that
425 * requires the MP lock since the MP lock will have already been setup for
426 * the target thread (not the current thread). It's nice to have a scheduler
427 * that does not need the MP lock to work because it allows us to do some
428 * really cool high-performance MP lock optimizations.
434 globaldata_t gd = mycpu;
435 thread_t td = gd->gd_curthread;
442 * Switching from within a 'fast' (non thread switched) interrupt is
445 if (gd->gd_intr_nesting_level && panicstr == NULL) {
446 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
450 * Passive release (used to transition from user to kernel mode
451 * when we block or switch rather then when we enter the kernel).
452 * This function is NOT called if we are switching into a preemption
453 * or returning from a preemption. Typically this causes us to lose
454 * our current process designation (if we have one) and become a true
455 * LWKT thread, and may also hand the current process designation to
456 * another process and schedule thread.
466 * td_mpcount cannot be used to determine if we currently hold the
467 * MP lock because get_mplock() will increment it prior to attempting
468 * to get the lock, and switch out if it can't. Our ownership of
469 * the actual lock will remain stable while we are in a critical section
470 * (but, of course, another cpu may own or release the lock so the
471 * actual value of mp_lock is not stable).
473 mpheld = MP_LOCK_HELD();
475 if (td->td_cscount) {
476 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
478 if (panic_on_cscount)
479 panic("switching while mastering cpusync");
483 if ((ntd = td->td_preempted) != NULL) {
485 * We had preempted another thread on this cpu, resume the preempted
486 * thread. This occurs transparently, whether the preempted thread
487 * was scheduled or not (it may have been preempted after descheduling
490 * We have to setup the MP lock for the original thread after backing
491 * out the adjustment that was made to curthread when the original
494 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
496 if (ntd->td_mpcount && mpheld == 0) {
497 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
498 td, ntd, td->td_mpcount, ntd->td_mpcount);
500 if (ntd->td_mpcount) {
501 td->td_mpcount -= ntd->td_mpcount;
502 KKASSERT(td->td_mpcount >= 0);
505 ntd->td_flags |= TDF_PREEMPT_DONE;
508 * XXX. The interrupt may have woken a thread up, we need to properly
509 * set the reschedule flag if the originally interrupted thread is at
512 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
514 /* YYY release mp lock on switchback if original doesn't need it */
517 * Priority queue / round-robin at each priority. Note that user
518 * processes run at a fixed, low priority and the user process
519 * scheduler deals with interactions between user processes
520 * by scheduling and descheduling them from the LWKT queue as
523 * We have to adjust the MP lock for the target thread. If we
524 * need the MP lock and cannot obtain it we try to locate a
525 * thread that does not need the MP lock. If we cannot, we spin
528 * A similar issue exists for the tokens held by the target thread.
529 * If we cannot obtain ownership of the tokens we cannot immediately
530 * schedule the thread.
534 * We are switching threads. If there are any pending requests for
535 * tokens we can satisfy all of them here.
538 if (gd->gd_tokreqbase)
539 lwkt_drain_token_requests();
543 * If an LWKT reschedule was requested, well that is what we are
544 * doing now so clear it.
546 clear_lwkt_resched();
548 if (gd->gd_runqmask) {
549 int nq = bsrl(gd->gd_runqmask);
550 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
551 gd->gd_runqmask &= ~(1 << nq);
556 * If the target needs the MP lock and we couldn't get it,
557 * or if the target is holding tokens and we could not
558 * gain ownership of the tokens, continue looking for a
559 * thread to schedule and spin instead of HLT if we can't.
561 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
562 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
564 u_int32_t rqmask = gd->gd_runqmask;
566 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
567 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
569 mpheld = MP_LOCK_HELD();
570 if (ntd->td_toks && !lwkt_chktokens(ntd))
576 rqmask &= ~(1 << nq);
580 ntd = &gd->gd_idlethread;
581 ntd->td_flags |= TDF_IDLE_NOHLT;
583 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
584 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
587 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
588 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
591 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
592 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
596 * We have nothing to run but only let the idle loop halt
597 * the cpu if there are no pending interrupts.
599 ntd = &gd->gd_idlethread;
600 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
601 ntd->td_flags |= TDF_IDLE_NOHLT;
604 KASSERT(ntd->td_pri >= TDPRI_CRIT,
605 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
608 * Do the actual switch. If the new target does not need the MP lock
609 * and we are holding it, release the MP lock. If the new target requires
610 * the MP lock we have already acquired it for the target.
613 if (ntd->td_mpcount == 0 ) {
617 ASSERT_MP_LOCK_HELD();
622 /* NOTE: current cpu may have changed after switch */
627 * Request that the target thread preempt the current thread. Preemption
628 * only works under a specific set of conditions:
630 * - We are not preempting ourselves
631 * - The target thread is owned by the current cpu
632 * - We are not currently being preempted
633 * - The target is not currently being preempted
634 * - We are able to satisfy the target's MP lock requirements (if any).
636 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
637 * this is called via lwkt_schedule() through the td_preemptable callback.
638 * critpri is the managed critical priority that we should ignore in order
639 * to determine whether preemption is possible (aka usually just the crit
640 * priority of lwkt_schedule() itself).
642 * XXX at the moment we run the target thread in a critical section during
643 * the preemption in order to prevent the target from taking interrupts
644 * that *WE* can't. Preemption is strictly limited to interrupt threads
645 * and interrupt-like threads, outside of a critical section, and the
646 * preempted source thread will be resumed the instant the target blocks
647 * whether or not the source is scheduled (i.e. preemption is supposed to
648 * be as transparent as possible).
650 * The target thread inherits our MP count (added to its own) for the
651 * duration of the preemption in order to preserve the atomicy of the
652 * MP lock during the preemption. Therefore, any preempting targets must be
653 * careful in regards to MP assertions. Note that the MP count may be
654 * out of sync with the physical mp_lock, but we do not have to preserve
655 * the original ownership of the lock if it was out of synch (that is, we
656 * can leave it synchronized on return).
659 lwkt_preempt(thread_t ntd, int critpri)
661 struct globaldata *gd = mycpu;
669 * The caller has put us in a critical section. We can only preempt
670 * if the caller of the caller was not in a critical section (basically
671 * a local interrupt), as determined by the 'critpri' parameter.
673 * YYY The target thread must be in a critical section (else it must
674 * inherit our critical section? I dunno yet).
676 * Any tokens held by the target may not be held by thread(s) being
677 * preempted. We take the easy way out and do not preempt if
678 * the target is holding tokens.
680 * Set need_lwkt_resched() unconditionally for now YYY.
682 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
684 td = gd->gd_curthread;
685 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
689 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
695 if (ntd->td_gd != gd) {
702 * Take the easy way out and do not preempt if the target is holding
703 * one or more tokens. We could test whether the thread(s) being
704 * preempted interlock against the target thread's tokens and whether
705 * we can get all the target thread's tokens, but this situation
706 * should not occur very often so its easier to simply not preempt.
708 if (ntd->td_toks != NULL) {
713 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
718 if (ntd->td_preempted) {
725 * note: an interrupt might have occured just as we were transitioning
726 * to or from the MP lock. In this case td_mpcount will be pre-disposed
727 * (non-zero) but not actually synchronized with the actual state of the
728 * lock. We can use it to imply an MP lock requirement for the
729 * preemption but we cannot use it to test whether we hold the MP lock
732 savecnt = td->td_mpcount;
733 mpheld = MP_LOCK_HELD();
734 ntd->td_mpcount += td->td_mpcount;
735 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
736 ntd->td_mpcount -= td->td_mpcount;
744 * Since we are able to preempt the current thread, there is no need to
745 * call need_lwkt_resched().
748 ntd->td_preempted = td;
749 td->td_flags |= TDF_PREEMPT_LOCK;
751 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
753 KKASSERT(savecnt == td->td_mpcount);
754 mpheld = MP_LOCK_HELD();
755 if (mpheld && td->td_mpcount == 0)
757 else if (mpheld == 0 && td->td_mpcount)
758 panic("lwkt_preempt(): MP lock was not held through");
760 ntd->td_preempted = NULL;
761 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
765 * Yield our thread while higher priority threads are pending. This is
766 * typically called when we leave a critical section but it can be safely
767 * called while we are in a critical section.
769 * This function will not generally yield to equal priority threads but it
770 * can occur as a side effect. Note that lwkt_switch() is called from
771 * inside the critical section to prevent its own crit_exit() from reentering
772 * lwkt_yield_quick().
774 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
775 * came along but was blocked and made pending.
777 * (self contained on a per cpu basis)
780 lwkt_yield_quick(void)
782 globaldata_t gd = mycpu;
783 thread_t td = gd->gd_curthread;
786 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
787 * it with a non-zero cpl then we might not wind up calling splz after
788 * a task switch when the critical section is exited even though the
789 * new task could accept the interrupt.
791 * XXX from crit_exit() only called after last crit section is released.
792 * If called directly will run splz() even if in a critical section.
794 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
795 * except for this special case, we MUST call splz() here to handle any
796 * pending ints, particularly after we switch, or we might accidently
797 * halt the cpu with interrupts pending.
799 if (gd->gd_reqflags && td->td_nest_count < 2)
803 * YYY enabling will cause wakeup() to task-switch, which really
804 * confused the old 4.x code. This is a good way to simulate
805 * preemption and MP without actually doing preemption or MP, because a
806 * lot of code assumes that wakeup() does not block.
808 if (untimely_switch && td->td_nest_count == 0 &&
809 gd->gd_intr_nesting_level == 0
811 crit_enter_quick(td);
813 * YYY temporary hacks until we disassociate the userland scheduler
814 * from the LWKT scheduler.
816 if (td->td_flags & TDF_RUNQ) {
817 lwkt_switch(); /* will not reenter yield function */
819 lwkt_schedule_self(td); /* make sure we are scheduled */
820 lwkt_switch(); /* will not reenter yield function */
821 lwkt_deschedule_self(td); /* make sure we are descheduled */
823 crit_exit_noyield(td);
828 * This implements a normal yield which, unlike _quick, will yield to equal
829 * priority threads as well. Note that gd_reqflags tests will be handled by
830 * the crit_exit() call in lwkt_switch().
832 * (self contained on a per cpu basis)
837 lwkt_schedule_self(curthread);
842 * Generic schedule. Possibly schedule threads belonging to other cpus and
843 * deal with threads that might be blocked on a wait queue.
845 * We have a little helper inline function which does additional work after
846 * the thread has been enqueued, including dealing with preemption and
847 * setting need_lwkt_resched() (which prevents the kernel from returning
848 * to userland until it has processed higher priority threads).
852 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
854 if (ntd->td_preemptable) {
855 ntd->td_preemptable(ntd, cpri); /* YYY +token */
856 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
857 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
864 lwkt_schedule(thread_t td)
866 globaldata_t mygd = mycpu;
869 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
870 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
871 && td->td_proc->p_stat == SSLEEP
873 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
875 curthread->td_proc ? curthread->td_proc->p_pid : -1,
876 curthread->td_proc ? curthread->td_proc->p_stat : -1,
878 td->td_proc ? curthread->td_proc->p_pid : -1,
879 td->td_proc ? curthread->td_proc->p_stat : -1
881 panic("SCHED PANIC");
885 if (td == mygd->gd_curthread) {
891 * If the thread is on a wait list we have to send our scheduling
892 * request to the owner of the wait structure. Otherwise we send
893 * the scheduling request to the cpu owning the thread. Races
894 * are ok, the target will forward the message as necessary (the
895 * message may chase the thread around before it finally gets
898 * (remember, wait structures use stable storage)
900 * NOTE: tokens no longer enter a critical section, so we only need
901 * to account for the crit_enter() above when calling
902 * _lwkt_schedule_post().
904 if ((w = td->td_wait) != NULL) {
907 if (lwkt_trytoken(&wref, &w->wa_token)) {
908 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
912 if (td->td_gd == mygd) {
914 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
916 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
920 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
922 lwkt_reltoken(&wref);
924 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
928 * If the wait structure is NULL and we own the thread, there
929 * is no race (since we are in a critical section). If we
930 * do not own the thread there might be a race but the
931 * target cpu will deal with it.
934 if (td->td_gd == mygd) {
936 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
938 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
942 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
950 * Managed acquisition. This code assumes that the MP lock is held for
951 * the tdallq operation and that the thread has been descheduled from its
952 * original cpu. We also have to wait for the thread to be entirely switched
953 * out on its original cpu (this is usually fast enough that we never loop)
954 * since the LWKT system does not have to hold the MP lock while switching
955 * and the target may have released it before switching.
958 lwkt_acquire(thread_t td)
965 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
966 while (td->td_flags & TDF_RUNNING) /* XXX spin */
970 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
972 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
978 * Generic deschedule. Descheduling threads other then your own should be
979 * done only in carefully controlled circumstances. Descheduling is
982 * This function may block if the cpu has run out of messages.
985 lwkt_deschedule(thread_t td)
988 if (td == curthread) {
991 if (td->td_gd == mycpu) {
994 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
1001 * Set the target thread's priority. This routine does not automatically
1002 * switch to a higher priority thread, LWKT threads are not designed for
1003 * continuous priority changes. Yield if you want to switch.
1005 * We have to retain the critical section count which uses the high bits
1006 * of the td_pri field. The specified priority may also indicate zero or
1007 * more critical sections by adding TDPRI_CRIT*N.
1009 * Note that we requeue the thread whether it winds up on a different runq
1010 * or not. uio_yield() depends on this and the routine is not normally
1011 * called with the same priority otherwise.
1014 lwkt_setpri(thread_t td, int pri)
1017 KKASSERT(td->td_gd == mycpu);
1019 if (td->td_flags & TDF_RUNQ) {
1021 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1024 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1030 lwkt_setpri_self(int pri)
1032 thread_t td = curthread;
1034 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1036 if (td->td_flags & TDF_RUNQ) {
1038 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1041 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1047 * Determine if there is a runnable thread at a higher priority then
1048 * the current thread. lwkt_setpri() does not check this automatically.
1049 * Return 1 if there is, 0 if there isn't.
1051 * Example: if bit 31 of runqmask is set and the current thread is priority
1052 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1054 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1055 * up comparing against 0xffffffff, a comparison that will always be false.
1058 lwkt_checkpri_self(void)
1060 globaldata_t gd = mycpu;
1061 thread_t td = gd->gd_curthread;
1062 int nq = td->td_pri & TDPRI_MASK;
1064 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1065 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1073 * Migrate the current thread to the specified cpu. The BGL must be held
1074 * (for the gd_tdallq manipulation XXX). This is accomplished by
1075 * descheduling ourselves from the current cpu, moving our thread to the
1076 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1077 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1080 static void lwkt_setcpu_remote(void *arg);
1084 lwkt_setcpu_self(globaldata_t rgd)
1087 thread_t td = curthread;
1089 if (td->td_gd != rgd) {
1090 crit_enter_quick(td);
1091 td->td_flags |= TDF_MIGRATING;
1092 lwkt_deschedule_self(td);
1093 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1094 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1095 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1097 /* we are now on the target cpu */
1098 crit_exit_quick(td);
1104 * Remote IPI for cpu migration (called while in a critical section so we
1105 * do not have to enter another one). The thread has already been moved to
1106 * our cpu's allq, but we must wait for the thread to be completely switched
1107 * out on the originating cpu before we schedule it on ours or the stack
1108 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1109 * change to main memory.
1111 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1112 * against wakeups. It is best if this interface is used only when there
1113 * are no pending events that might try to schedule the thread.
1117 lwkt_setcpu_remote(void *arg)
1120 globaldata_t gd = mycpu;
1122 while (td->td_flags & TDF_RUNNING)
1126 td->td_flags &= ~TDF_MIGRATING;
1132 lwkt_preempted_proc(void)
1134 thread_t td = curthread;
1135 while (td->td_preempted)
1136 td = td->td_preempted;
1137 return(td->td_proc);
1141 * Block on the specified wait queue until signaled. A generation number
1142 * must be supplied to interlock the wait queue. The function will
1143 * return immediately if the generation number does not match the wait
1144 * structure's generation number.
1147 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1149 thread_t td = curthread;
1152 lwkt_gettoken(&ilock, &w->wa_token);
1154 if (w->wa_gen == *gen) {
1156 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1159 td->td_wmesg = wmesg;
1162 if (td->td_wmesg != NULL) {
1169 lwkt_reltoken(&ilock);
1173 * Signal a wait queue. We gain ownership of the wait queue in order to
1174 * signal it. Once a thread is removed from the wait queue we have to
1175 * deal with the cpu owning the thread.
1177 * Note: alternatively we could message the target cpu owning the wait
1178 * queue. YYY implement as sysctl.
1181 lwkt_signal(lwkt_wait_t w, int count)
1186 lwkt_gettoken(&ilock, &w->wa_token);
1190 count = w->wa_count;
1191 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1194 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1196 td->td_wmesg = NULL;
1197 if (td->td_gd == mycpu) {
1200 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1204 lwkt_reltoken(&ilock);
1208 * Create a kernel process/thread/whatever. It shares it's address space
1209 * with proc0 - ie: kernel only.
1211 * NOTE! By default new threads are created with the MP lock held. A
1212 * thread which does not require the MP lock should release it by calling
1213 * rel_mplock() at the start of the new thread.
1216 lwkt_create(void (*func)(void *), void *arg,
1217 struct thread **tdp, thread_t template, int tdflags, int cpu,
1218 const char *fmt, ...)
1223 td = lwkt_alloc_thread(template, cpu);
1226 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1227 td->td_flags |= TDF_VERBOSE | tdflags;
1233 * Set up arg0 for 'ps' etc
1235 __va_start(ap, fmt);
1236 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1240 * Schedule the thread to run
1242 if ((td->td_flags & TDF_STOPREQ) == 0)
1245 td->td_flags &= ~TDF_STOPREQ;
1250 * kthread_* is specific to the kernel and is not needed by userland.
1255 * Destroy an LWKT thread. Warning! This function is not called when
1256 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1257 * uses a different reaping mechanism.
1262 thread_t td = curthread;
1265 if (td->td_flags & TDF_VERBOSE)
1266 printf("kthread %p %s has exited\n", td, td->td_comm);
1268 crit_enter_quick(td);
1269 lwkt_deschedule_self(td);
1271 KKASSERT(gd == td->td_gd);
1272 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1273 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1274 ++gd->gd_tdfreecount;
1275 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1281 * Create a kernel process/thread/whatever. It shares it's address space
1282 * with proc0 - ie: kernel only. 5.x compatible.
1284 * NOTE! By default kthreads are created with the MP lock held. A
1285 * thread which does not require the MP lock should release it by calling
1286 * rel_mplock() at the start of the new thread.
1289 kthread_create(void (*func)(void *), void *arg,
1290 struct thread **tdp, const char *fmt, ...)
1295 td = lwkt_alloc_thread(NULL, -1);
1298 cpu_set_thread_handler(td, kthread_exit, func, arg);
1299 td->td_flags |= TDF_VERBOSE;
1305 * Set up arg0 for 'ps' etc
1307 __va_start(ap, fmt);
1308 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1312 * Schedule the thread to run
1319 * Destroy an LWKT thread. Warning! This function is not called when
1320 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1321 * uses a different reaping mechanism.
1323 * XXX duplicates lwkt_exit()
1331 #endif /* _KERNEL */
1336 thread_t td = curthread;
1337 int lpri = td->td_pri;
1340 panic("td_pri is/would-go negative! %p %d", td, lpri);