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 1.69 2004/09/21 18:09:57 joerg 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>
75 #include <sys/stdint.h>
76 #include <libcaps/thread.h>
77 #include <sys/thread.h>
78 #include <sys/msgport.h>
79 #include <sys/errno.h>
80 #include <libcaps/globaldata.h>
81 #include <machine/cpufunc.h>
82 #include <sys/thread2.h>
83 #include <sys/msgport2.h>
87 #include <machine/lock.h>
91 static int untimely_switch = 0;
93 static int panic_on_cscount = 0;
95 static __int64_t switch_count = 0;
96 static __int64_t preempt_hit = 0;
97 static __int64_t preempt_miss = 0;
98 static __int64_t preempt_weird = 0;
102 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
104 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
106 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
107 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
109 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
114 * These helper procedures handle the runq, they can only be called from
115 * within a critical section.
117 * WARNING! Prior to SMP being brought up it is possible to enqueue and
118 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
119 * instead of 'mycpu' when referencing the globaldata structure. Once
120 * SMP live enqueuing and dequeueing only occurs on the current cpu.
124 _lwkt_dequeue(thread_t td)
126 if (td->td_flags & TDF_RUNQ) {
127 int nq = td->td_pri & TDPRI_MASK;
128 struct globaldata *gd = td->td_gd;
130 td->td_flags &= ~TDF_RUNQ;
131 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
132 /* runqmask is passively cleaned up by the switcher */
138 _lwkt_enqueue(thread_t td)
140 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
141 int nq = td->td_pri & TDPRI_MASK;
142 struct globaldata *gd = td->td_gd;
144 td->td_flags |= TDF_RUNQ;
145 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
146 gd->gd_runqmask |= 1 << nq;
151 * Schedule a thread to run. As the current thread we can always safely
152 * schedule ourselves, and a shortcut procedure is provided for that
155 * (non-blocking, self contained on a per cpu basis)
158 lwkt_schedule_self(thread_t td)
160 crit_enter_quick(td);
161 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
162 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
165 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
166 panic("SCHED SELF PANIC");
172 * Deschedule a thread.
174 * (non-blocking, self contained on a per cpu basis)
177 lwkt_deschedule_self(thread_t td)
179 crit_enter_quick(td);
180 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
188 * LWKTs operate on a per-cpu basis
190 * WARNING! Called from early boot, 'mycpu' may not work yet.
193 lwkt_gdinit(struct globaldata *gd)
197 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
198 TAILQ_INIT(&gd->gd_tdrunq[i]);
200 TAILQ_INIT(&gd->gd_tdallq);
206 * Initialize a thread wait structure prior to first use.
208 * NOTE! called from low level boot code, we cannot do anything fancy!
211 lwkt_wait_init(lwkt_wait_t w)
213 lwkt_token_init(&w->wa_token);
214 TAILQ_INIT(&w->wa_waitq);
220 * Create a new thread. The thread must be associated with a process context
221 * or LWKT start address before it can be scheduled. If the target cpu is
222 * -1 the thread will be created on the current cpu.
224 * If you intend to create a thread without a process context this function
225 * does everything except load the startup and switcher function.
228 lwkt_alloc_thread(struct thread *td, int stksize, int cpu)
232 globaldata_t gd = mycpu;
236 if (gd->gd_tdfreecount > 0) {
237 --gd->gd_tdfreecount;
238 td = TAILQ_FIRST(&gd->gd_tdfreeq);
239 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
240 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
241 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
243 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
247 td = zalloc(thread_zone);
249 td = malloc(sizeof(struct thread));
251 td->td_kstack = NULL;
252 td->td_kstack_size = 0;
253 flags |= TDF_ALLOCATED_THREAD;
256 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
257 if (flags & TDF_ALLOCATED_STACK) {
259 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
261 libcaps_free_stack(stack, td->td_kstack_size);
268 stack = (void *)kmem_alloc(kernel_map, stksize);
270 stack = libcaps_alloc_stack(stksize);
272 flags |= TDF_ALLOCATED_STACK;
275 lwkt_init_thread(td, stack, stksize, flags, mycpu);
277 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
284 * Initialize a preexisting thread structure. This function is used by
285 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
287 * All threads start out in a critical section at a priority of
288 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
289 * appropriate. This function may send an IPI message when the
290 * requested cpu is not the current cpu and consequently gd_tdallq may
291 * not be initialized synchronously from the point of view of the originating
294 * NOTE! we have to be careful in regards to creating threads for other cpus
295 * if SMP has not yet been activated.
300 lwkt_init_thread_remote(void *arg)
304 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
310 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
311 struct globaldata *gd)
313 globaldata_t mygd = mycpu;
315 bzero(td, sizeof(struct thread));
316 td->td_kstack = stack;
317 td->td_kstack_size = stksize;
318 td->td_flags |= flags;
320 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
321 lwkt_initport(&td->td_msgport, td);
322 pmap_init_thread(td);
325 * Normally initializing a thread for a remote cpu requires sending an
326 * IPI. However, the idlethread is setup before the other cpus are
327 * activated so we have to treat it as a special case. XXX manipulation
328 * of gd_tdallq requires the BGL.
330 if (gd == mygd || td == &gd->gd_idlethread) {
332 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
335 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
339 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
347 lwkt_set_comm(thread_t td, const char *ctl, ...)
352 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
357 lwkt_hold(thread_t td)
363 lwkt_rele(thread_t td)
365 KKASSERT(td->td_refs > 0);
372 lwkt_wait_free(thread_t td)
375 tsleep(td, 0, "tdreap", hz);
381 lwkt_free_thread(thread_t td)
383 struct globaldata *gd = mycpu;
385 KASSERT((td->td_flags & TDF_RUNNING) == 0,
386 ("lwkt_free_thread: did not exit! %p", td));
389 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
390 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
391 (td->td_flags & TDF_ALLOCATED_THREAD)
393 ++gd->gd_tdfreecount;
394 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
398 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
400 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
402 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
405 td->td_kstack = NULL;
406 td->td_kstack_size = 0;
408 if (td->td_flags & TDF_ALLOCATED_THREAD) {
410 zfree(thread_zone, td);
420 * Switch to the next runnable lwkt. If no LWKTs are runnable then
421 * switch to the idlethread. Switching must occur within a critical
422 * section to avoid races with the scheduling queue.
424 * We always have full control over our cpu's run queue. Other cpus
425 * that wish to manipulate our queue must use the cpu_*msg() calls to
426 * talk to our cpu, so a critical section is all that is needed and
427 * the result is very, very fast thread switching.
429 * The LWKT scheduler uses a fixed priority model and round-robins at
430 * each priority level. User process scheduling is a totally
431 * different beast and LWKT priorities should not be confused with
432 * user process priorities.
434 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
435 * cleans it up. Note that the td_switch() function cannot do anything that
436 * requires the MP lock since the MP lock will have already been setup for
437 * the target thread (not the current thread). It's nice to have a scheduler
438 * that does not need the MP lock to work because it allows us to do some
439 * really cool high-performance MP lock optimizations.
445 globaldata_t gd = mycpu;
446 thread_t td = gd->gd_curthread;
453 * Switching from within a 'fast' (non thread switched) interrupt is
456 if (gd->gd_intr_nesting_level && panicstr == NULL) {
457 panic("lwkt_switch: cannot switch from within a fast interrupt, yet");
461 * Passive release (used to transition from user to kernel mode
462 * when we block or switch rather then when we enter the kernel).
463 * This function is NOT called if we are switching into a preemption
464 * or returning from a preemption. Typically this causes us to lose
465 * our current process designation (if we have one) and become a true
466 * LWKT thread, and may also hand the current process designation to
467 * another process and schedule thread.
477 * td_mpcount cannot be used to determine if we currently hold the
478 * MP lock because get_mplock() will increment it prior to attempting
479 * to get the lock, and switch out if it can't. Our ownership of
480 * the actual lock will remain stable while we are in a critical section
481 * (but, of course, another cpu may own or release the lock so the
482 * actual value of mp_lock is not stable).
484 mpheld = MP_LOCK_HELD();
486 if (td->td_cscount) {
487 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
489 if (panic_on_cscount)
490 panic("switching while mastering cpusync");
494 if ((ntd = td->td_preempted) != NULL) {
496 * We had preempted another thread on this cpu, resume the preempted
497 * thread. This occurs transparently, whether the preempted thread
498 * was scheduled or not (it may have been preempted after descheduling
501 * We have to setup the MP lock for the original thread after backing
502 * out the adjustment that was made to curthread when the original
505 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
507 if (ntd->td_mpcount && mpheld == 0) {
508 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
509 td, ntd, td->td_mpcount, ntd->td_mpcount);
511 if (ntd->td_mpcount) {
512 td->td_mpcount -= ntd->td_mpcount;
513 KKASSERT(td->td_mpcount >= 0);
516 ntd->td_flags |= TDF_PREEMPT_DONE;
519 * XXX. The interrupt may have woken a thread up, we need to properly
520 * set the reschedule flag if the originally interrupted thread is at
523 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
525 /* YYY release mp lock on switchback if original doesn't need it */
528 * Priority queue / round-robin at each priority. Note that user
529 * processes run at a fixed, low priority and the user process
530 * scheduler deals with interactions between user processes
531 * by scheduling and descheduling them from the LWKT queue as
534 * We have to adjust the MP lock for the target thread. If we
535 * need the MP lock and cannot obtain it we try to locate a
536 * thread that does not need the MP lock. If we cannot, we spin
539 * A similar issue exists for the tokens held by the target thread.
540 * If we cannot obtain ownership of the tokens we cannot immediately
541 * schedule the thread.
545 * We are switching threads. If there are any pending requests for
546 * tokens we can satisfy all of them here.
549 if (gd->gd_tokreqbase)
550 lwkt_drain_token_requests();
554 * If an LWKT reschedule was requested, well that is what we are
555 * doing now so clear it.
557 clear_lwkt_resched();
559 if (gd->gd_runqmask) {
560 int nq = bsrl(gd->gd_runqmask);
561 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
562 gd->gd_runqmask &= ~(1 << nq);
567 * If the target needs the MP lock and we couldn't get it,
568 * or if the target is holding tokens and we could not
569 * gain ownership of the tokens, continue looking for a
570 * thread to schedule and spin instead of HLT if we can't.
572 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
573 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
575 u_int32_t rqmask = gd->gd_runqmask;
577 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
578 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
580 mpheld = MP_LOCK_HELD();
581 if (ntd->td_toks && !lwkt_chktokens(ntd))
587 rqmask &= ~(1 << nq);
591 ntd = &gd->gd_idlethread;
592 ntd->td_flags |= TDF_IDLE_NOHLT;
594 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
595 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
598 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
599 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
602 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
603 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
607 * We have nothing to run but only let the idle loop halt
608 * the cpu if there are no pending interrupts.
610 ntd = &gd->gd_idlethread;
611 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
612 ntd->td_flags |= TDF_IDLE_NOHLT;
615 KASSERT(ntd->td_pri >= TDPRI_CRIT,
616 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
619 * Do the actual switch. If the new target does not need the MP lock
620 * and we are holding it, release the MP lock. If the new target requires
621 * the MP lock we have already acquired it for the target.
624 if (ntd->td_mpcount == 0 ) {
628 ASSERT_MP_LOCK_HELD();
633 /* NOTE: current cpu may have changed after switch */
638 * Request that the target thread preempt the current thread. Preemption
639 * only works under a specific set of conditions:
641 * - We are not preempting ourselves
642 * - The target thread is owned by the current cpu
643 * - We are not currently being preempted
644 * - The target is not currently being preempted
645 * - We are able to satisfy the target's MP lock requirements (if any).
647 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
648 * this is called via lwkt_schedule() through the td_preemptable callback.
649 * critpri is the managed critical priority that we should ignore in order
650 * to determine whether preemption is possible (aka usually just the crit
651 * priority of lwkt_schedule() itself).
653 * XXX at the moment we run the target thread in a critical section during
654 * the preemption in order to prevent the target from taking interrupts
655 * that *WE* can't. Preemption is strictly limited to interrupt threads
656 * and interrupt-like threads, outside of a critical section, and the
657 * preempted source thread will be resumed the instant the target blocks
658 * whether or not the source is scheduled (i.e. preemption is supposed to
659 * be as transparent as possible).
661 * The target thread inherits our MP count (added to its own) for the
662 * duration of the preemption in order to preserve the atomicy of the
663 * MP lock during the preemption. Therefore, any preempting targets must be
664 * careful in regards to MP assertions. Note that the MP count may be
665 * out of sync with the physical mp_lock, but we do not have to preserve
666 * the original ownership of the lock if it was out of synch (that is, we
667 * can leave it synchronized on return).
670 lwkt_preempt(thread_t ntd, int critpri)
672 struct globaldata *gd = mycpu;
680 * The caller has put us in a critical section. We can only preempt
681 * if the caller of the caller was not in a critical section (basically
682 * a local interrupt), as determined by the 'critpri' parameter.
684 * YYY The target thread must be in a critical section (else it must
685 * inherit our critical section? I dunno yet).
687 * Any tokens held by the target may not be held by thread(s) being
688 * preempted. We take the easy way out and do not preempt if
689 * the target is holding tokens.
691 * Set need_lwkt_resched() unconditionally for now YYY.
693 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
695 td = gd->gd_curthread;
696 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
700 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
706 if (ntd->td_gd != gd) {
713 * Take the easy way out and do not preempt if the target is holding
714 * one or more tokens. We could test whether the thread(s) being
715 * preempted interlock against the target thread's tokens and whether
716 * we can get all the target thread's tokens, but this situation
717 * should not occur very often so its easier to simply not preempt.
719 if (ntd->td_toks != NULL) {
724 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
729 if (ntd->td_preempted) {
736 * note: an interrupt might have occured just as we were transitioning
737 * to or from the MP lock. In this case td_mpcount will be pre-disposed
738 * (non-zero) but not actually synchronized with the actual state of the
739 * lock. We can use it to imply an MP lock requirement for the
740 * preemption but we cannot use it to test whether we hold the MP lock
743 savecnt = td->td_mpcount;
744 mpheld = MP_LOCK_HELD();
745 ntd->td_mpcount += td->td_mpcount;
746 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
747 ntd->td_mpcount -= td->td_mpcount;
755 * Since we are able to preempt the current thread, there is no need to
756 * call need_lwkt_resched().
759 ntd->td_preempted = td;
760 td->td_flags |= TDF_PREEMPT_LOCK;
762 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
764 KKASSERT(savecnt == td->td_mpcount);
765 mpheld = MP_LOCK_HELD();
766 if (mpheld && td->td_mpcount == 0)
768 else if (mpheld == 0 && td->td_mpcount)
769 panic("lwkt_preempt(): MP lock was not held through");
771 ntd->td_preempted = NULL;
772 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
776 * Yield our thread while higher priority threads are pending. This is
777 * typically called when we leave a critical section but it can be safely
778 * called while we are in a critical section.
780 * This function will not generally yield to equal priority threads but it
781 * can occur as a side effect. Note that lwkt_switch() is called from
782 * inside the critical section to prevent its own crit_exit() from reentering
783 * lwkt_yield_quick().
785 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
786 * came along but was blocked and made pending.
788 * (self contained on a per cpu basis)
791 lwkt_yield_quick(void)
793 globaldata_t gd = mycpu;
794 thread_t td = gd->gd_curthread;
797 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
798 * it with a non-zero cpl then we might not wind up calling splz after
799 * a task switch when the critical section is exited even though the
800 * new task could accept the interrupt.
802 * XXX from crit_exit() only called after last crit section is released.
803 * If called directly will run splz() even if in a critical section.
805 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
806 * except for this special case, we MUST call splz() here to handle any
807 * pending ints, particularly after we switch, or we might accidently
808 * halt the cpu with interrupts pending.
810 if (gd->gd_reqflags && td->td_nest_count < 2)
814 * YYY enabling will cause wakeup() to task-switch, which really
815 * confused the old 4.x code. This is a good way to simulate
816 * preemption and MP without actually doing preemption or MP, because a
817 * lot of code assumes that wakeup() does not block.
819 if (untimely_switch && td->td_nest_count == 0 &&
820 gd->gd_intr_nesting_level == 0
822 crit_enter_quick(td);
824 * YYY temporary hacks until we disassociate the userland scheduler
825 * from the LWKT scheduler.
827 if (td->td_flags & TDF_RUNQ) {
828 lwkt_switch(); /* will not reenter yield function */
830 lwkt_schedule_self(td); /* make sure we are scheduled */
831 lwkt_switch(); /* will not reenter yield function */
832 lwkt_deschedule_self(td); /* make sure we are descheduled */
834 crit_exit_noyield(td);
839 * This implements a normal yield which, unlike _quick, will yield to equal
840 * priority threads as well. Note that gd_reqflags tests will be handled by
841 * the crit_exit() call in lwkt_switch().
843 * (self contained on a per cpu basis)
848 lwkt_schedule_self(curthread);
853 * Generic schedule. Possibly schedule threads belonging to other cpus and
854 * deal with threads that might be blocked on a wait queue.
856 * We have a little helper inline function which does additional work after
857 * the thread has been enqueued, including dealing with preemption and
858 * setting need_lwkt_resched() (which prevents the kernel from returning
859 * to userland until it has processed higher priority threads).
863 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
865 if (ntd->td_preemptable) {
866 ntd->td_preemptable(ntd, cpri); /* YYY +token */
867 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
868 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
875 lwkt_schedule(thread_t td)
877 globaldata_t mygd = mycpu;
880 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
881 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
882 && td->td_proc->p_stat == SSLEEP
884 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
886 curthread->td_proc ? curthread->td_proc->p_pid : -1,
887 curthread->td_proc ? curthread->td_proc->p_stat : -1,
889 td->td_proc ? curthread->td_proc->p_pid : -1,
890 td->td_proc ? curthread->td_proc->p_stat : -1
892 panic("SCHED PANIC");
896 if (td == mygd->gd_curthread) {
902 * If the thread is on a wait list we have to send our scheduling
903 * request to the owner of the wait structure. Otherwise we send
904 * the scheduling request to the cpu owning the thread. Races
905 * are ok, the target will forward the message as necessary (the
906 * message may chase the thread around before it finally gets
909 * (remember, wait structures use stable storage)
911 * NOTE: tokens no longer enter a critical section, so we only need
912 * to account for the crit_enter() above when calling
913 * _lwkt_schedule_post().
915 if ((w = td->td_wait) != NULL) {
918 if (lwkt_trytoken(&wref, &w->wa_token)) {
919 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
923 if (td->td_gd == mygd) {
925 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
927 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
931 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
933 lwkt_reltoken(&wref);
935 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
939 * If the wait structure is NULL and we own the thread, there
940 * is no race (since we are in a critical section). If we
941 * do not own the thread there might be a race but the
942 * target cpu will deal with it.
945 if (td->td_gd == mygd) {
947 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
949 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
953 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
961 * Managed acquisition. This code assumes that the MP lock is held for
962 * the tdallq operation and that the thread has been descheduled from its
963 * original cpu. We also have to wait for the thread to be entirely switched
964 * out on its original cpu (this is usually fast enough that we never loop)
965 * since the LWKT system does not have to hold the MP lock while switching
966 * and the target may have released it before switching.
969 lwkt_acquire(thread_t td)
976 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
977 while (td->td_flags & TDF_RUNNING) /* XXX spin */
981 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
983 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
989 * Generic deschedule. Descheduling threads other then your own should be
990 * done only in carefully controlled circumstances. Descheduling is
993 * This function may block if the cpu has run out of messages.
996 lwkt_deschedule(thread_t td)
999 if (td == curthread) {
1002 if (td->td_gd == mycpu) {
1005 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
1012 * Set the target thread's priority. This routine does not automatically
1013 * switch to a higher priority thread, LWKT threads are not designed for
1014 * continuous priority changes. Yield if you want to switch.
1016 * We have to retain the critical section count which uses the high bits
1017 * of the td_pri field. The specified priority may also indicate zero or
1018 * more critical sections by adding TDPRI_CRIT*N.
1020 * Note that we requeue the thread whether it winds up on a different runq
1021 * or not. uio_yield() depends on this and the routine is not normally
1022 * called with the same priority otherwise.
1025 lwkt_setpri(thread_t td, int pri)
1028 KKASSERT(td->td_gd == mycpu);
1030 if (td->td_flags & TDF_RUNQ) {
1032 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1035 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1041 lwkt_setpri_self(int pri)
1043 thread_t td = curthread;
1045 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1047 if (td->td_flags & TDF_RUNQ) {
1049 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1052 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1058 * Determine if there is a runnable thread at a higher priority then
1059 * the current thread. lwkt_setpri() does not check this automatically.
1060 * Return 1 if there is, 0 if there isn't.
1062 * Example: if bit 31 of runqmask is set and the current thread is priority
1063 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1065 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1066 * up comparing against 0xffffffff, a comparison that will always be false.
1069 lwkt_checkpri_self(void)
1071 globaldata_t gd = mycpu;
1072 thread_t td = gd->gd_curthread;
1073 int nq = td->td_pri & TDPRI_MASK;
1075 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1076 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1084 * Migrate the current thread to the specified cpu. The BGL must be held
1085 * (for the gd_tdallq manipulation XXX). This is accomplished by
1086 * descheduling ourselves from the current cpu, moving our thread to the
1087 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1088 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1091 static void lwkt_setcpu_remote(void *arg);
1095 lwkt_setcpu_self(globaldata_t rgd)
1098 thread_t td = curthread;
1100 if (td->td_gd != rgd) {
1101 crit_enter_quick(td);
1102 td->td_flags |= TDF_MIGRATING;
1103 lwkt_deschedule_self(td);
1104 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1105 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1106 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1108 /* we are now on the target cpu */
1109 crit_exit_quick(td);
1115 * Remote IPI for cpu migration (called while in a critical section so we
1116 * do not have to enter another one). The thread has already been moved to
1117 * our cpu's allq, but we must wait for the thread to be completely switched
1118 * out on the originating cpu before we schedule it on ours or the stack
1119 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1120 * change to main memory.
1122 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1123 * against wakeups. It is best if this interface is used only when there
1124 * are no pending events that might try to schedule the thread.
1128 lwkt_setcpu_remote(void *arg)
1131 globaldata_t gd = mycpu;
1133 while (td->td_flags & TDF_RUNNING)
1137 td->td_flags &= ~TDF_MIGRATING;
1143 lwkt_preempted_proc(void)
1145 thread_t td = curthread;
1146 while (td->td_preempted)
1147 td = td->td_preempted;
1148 return(td->td_proc);
1152 * Block on the specified wait queue until signaled. A generation number
1153 * must be supplied to interlock the wait queue. The function will
1154 * return immediately if the generation number does not match the wait
1155 * structure's generation number.
1158 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1160 thread_t td = curthread;
1163 lwkt_gettoken(&ilock, &w->wa_token);
1165 if (w->wa_gen == *gen) {
1167 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1170 td->td_wmesg = wmesg;
1173 if (td->td_wmesg != NULL) {
1180 lwkt_reltoken(&ilock);
1184 * Signal a wait queue. We gain ownership of the wait queue in order to
1185 * signal it. Once a thread is removed from the wait queue we have to
1186 * deal with the cpu owning the thread.
1188 * Note: alternatively we could message the target cpu owning the wait
1189 * queue. YYY implement as sysctl.
1192 lwkt_signal(lwkt_wait_t w, int count)
1197 lwkt_gettoken(&ilock, &w->wa_token);
1201 count = w->wa_count;
1202 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1205 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1207 td->td_wmesg = NULL;
1208 if (td->td_gd == mycpu) {
1211 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1215 lwkt_reltoken(&ilock);
1219 * Create a kernel process/thread/whatever. It shares it's address space
1220 * with proc0 - ie: kernel only.
1222 * NOTE! By default new threads are created with the MP lock held. A
1223 * thread which does not require the MP lock should release it by calling
1224 * rel_mplock() at the start of the new thread.
1227 lwkt_create(void (*func)(void *), void *arg,
1228 struct thread **tdp, thread_t template, int tdflags, int cpu,
1229 const char *fmt, ...)
1234 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
1237 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1238 td->td_flags |= TDF_VERBOSE | tdflags;
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
1253 if ((td->td_flags & TDF_STOPREQ) == 0)
1256 td->td_flags &= ~TDF_STOPREQ;
1261 * kthread_* is specific to the kernel and is not needed by userland.
1266 * Destroy an LWKT thread. Warning! This function is not called when
1267 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1268 * uses a different reaping mechanism.
1273 thread_t td = curthread;
1276 if (td->td_flags & TDF_VERBOSE)
1277 printf("kthread %p %s has exited\n", td, td->td_comm);
1279 crit_enter_quick(td);
1280 lwkt_deschedule_self(td);
1282 KKASSERT(gd == td->td_gd);
1283 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1284 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1285 ++gd->gd_tdfreecount;
1286 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1291 #endif /* _KERNEL */
1296 thread_t td = curthread;
1297 int lpri = td->td_pri;
1300 panic("td_pri is/would-go negative! %p %d", td, lpri);