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 1.75 2005/06/20 07:31:05 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>
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>
88 #include <machine/atomic.h>
89 #include <machine/cpu.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 stksize, 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 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
249 td = zalloc(thread_zone);
251 td = malloc(sizeof(struct thread));
253 td->td_kstack = NULL;
254 td->td_kstack_size = 0;
255 flags |= TDF_ALLOCATED_THREAD;
258 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
259 if (flags & TDF_ALLOCATED_STACK) {
261 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
263 libcaps_free_stack(stack, td->td_kstack_size);
270 stack = (void *)kmem_alloc(kernel_map, stksize);
272 stack = libcaps_alloc_stack(stksize);
274 flags |= TDF_ALLOCATED_STACK;
277 lwkt_init_thread(td, stack, stksize, flags, mycpu);
279 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
286 * Initialize a preexisting thread structure. This function is used by
287 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
289 * All threads start out in a critical section at a priority of
290 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
291 * appropriate. This function may send an IPI message when the
292 * requested cpu is not the current cpu and consequently gd_tdallq may
293 * not be initialized synchronously from the point of view of the originating
296 * NOTE! we have to be careful in regards to creating threads for other cpus
297 * if SMP has not yet been activated.
302 lwkt_init_thread_remote(void *arg)
306 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
312 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
313 struct globaldata *gd)
315 globaldata_t mygd = mycpu;
317 bzero(td, sizeof(struct thread));
318 td->td_kstack = stack;
319 td->td_kstack_size = stksize;
320 td->td_flags |= flags;
322 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
323 lwkt_initport(&td->td_msgport, td);
324 pmap_init_thread(td);
327 * Normally initializing a thread for a remote cpu requires sending an
328 * IPI. However, the idlethread is setup before the other cpus are
329 * activated so we have to treat it as a special case. XXX manipulation
330 * of gd_tdallq requires the BGL.
332 if (gd == mygd || td == &gd->gd_idlethread) {
334 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
337 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
349 lwkt_set_comm(thread_t td, const char *ctl, ...)
354 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
359 lwkt_hold(thread_t td)
365 lwkt_rele(thread_t td)
367 KKASSERT(td->td_refs > 0);
374 lwkt_wait_free(thread_t td)
377 tsleep(td, 0, "tdreap", hz);
383 lwkt_free_thread(thread_t td)
385 struct globaldata *gd = mycpu;
387 KASSERT((td->td_flags & TDF_RUNNING) == 0,
388 ("lwkt_free_thread: did not exit! %p", td));
391 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
392 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
393 (td->td_flags & TDF_ALLOCATED_THREAD)
395 ++gd->gd_tdfreecount;
396 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
400 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
402 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
404 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
407 td->td_kstack = NULL;
408 td->td_kstack_size = 0;
410 if (td->td_flags & TDF_ALLOCATED_THREAD) {
412 zfree(thread_zone, td);
422 * Switch to the next runnable lwkt. If no LWKTs are runnable then
423 * switch to the idlethread. Switching must occur within a critical
424 * section to avoid races with the scheduling queue.
426 * We always have full control over our cpu's run queue. Other cpus
427 * that wish to manipulate our queue must use the cpu_*msg() calls to
428 * talk to our cpu, so a critical section is all that is needed and
429 * the result is very, very fast thread switching.
431 * The LWKT scheduler uses a fixed priority model and round-robins at
432 * each priority level. User process scheduling is a totally
433 * different beast and LWKT priorities should not be confused with
434 * user process priorities.
436 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
437 * cleans it up. Note that the td_switch() function cannot do anything that
438 * requires the MP lock since the MP lock will have already been setup for
439 * the target thread (not the current thread). It's nice to have a scheduler
440 * that does not need the MP lock to work because it allows us to do some
441 * really cool high-performance MP lock optimizations.
447 globaldata_t gd = mycpu;
448 thread_t td = gd->gd_curthread;
455 * Switching from within a 'fast' (non thread switched) interrupt is
458 if (gd->gd_intr_nesting_level && panicstr == NULL) {
459 panic("lwkt_switch: cannot switch from within a fast interrupt, yet, td %p\n", td);
463 * Passive release (used to transition from user to kernel mode
464 * when we block or switch rather then when we enter the kernel).
465 * This function is NOT called if we are switching into a preemption
466 * or returning from a preemption. Typically this causes us to lose
467 * our current process designation (if we have one) and become a true
468 * LWKT thread, and may also hand the current process designation to
469 * another process and schedule thread.
478 * td_mpcount cannot be used to determine if we currently hold the
479 * MP lock because get_mplock() will increment it prior to attempting
480 * to get the lock, and switch out if it can't. Our ownership of
481 * the actual lock will remain stable while we are in a critical section
482 * (but, of course, another cpu may own or release the lock so the
483 * actual value of mp_lock is not stable).
485 mpheld = MP_LOCK_HELD();
487 if (td->td_cscount) {
488 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
490 if (panic_on_cscount)
491 panic("switching while mastering cpusync");
495 if ((ntd = td->td_preempted) != NULL) {
497 * We had preempted another thread on this cpu, resume the preempted
498 * thread. This occurs transparently, whether the preempted thread
499 * was scheduled or not (it may have been preempted after descheduling
502 * We have to setup the MP lock for the original thread after backing
503 * out the adjustment that was made to curthread when the original
506 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
508 if (ntd->td_mpcount && mpheld == 0) {
509 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
510 td, ntd, td->td_mpcount, ntd->td_mpcount);
512 if (ntd->td_mpcount) {
513 td->td_mpcount -= ntd->td_mpcount;
514 KKASSERT(td->td_mpcount >= 0);
517 ntd->td_flags |= TDF_PREEMPT_DONE;
520 * XXX. The interrupt may have woken a thread up, we need to properly
521 * set the reschedule flag if the originally interrupted thread is at
524 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
526 /* YYY release mp lock on switchback if original doesn't need it */
529 * Priority queue / round-robin at each priority. Note that user
530 * processes run at a fixed, low priority and the user process
531 * scheduler deals with interactions between user processes
532 * by scheduling and descheduling them from the LWKT queue as
535 * We have to adjust the MP lock for the target thread. If we
536 * need the MP lock and cannot obtain it we try to locate a
537 * thread that does not need the MP lock. If we cannot, we spin
540 * A similar issue exists for the tokens held by the target thread.
541 * If we cannot obtain ownership of the tokens we cannot immediately
542 * schedule the thread.
546 * We are switching threads. If there are any pending requests for
547 * tokens we can satisfy all of them here.
550 if (gd->gd_tokreqbase)
551 lwkt_drain_token_requests();
555 * If an LWKT reschedule was requested, well that is what we are
556 * doing now so clear it.
558 clear_lwkt_resched();
560 if (gd->gd_runqmask) {
561 int nq = bsrl(gd->gd_runqmask);
562 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
563 gd->gd_runqmask &= ~(1 << nq);
568 * If the target needs the MP lock and we couldn't get it,
569 * or if the target is holding tokens and we could not
570 * gain ownership of the tokens, continue looking for a
571 * thread to schedule and spin instead of HLT if we can't.
573 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
574 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
576 u_int32_t rqmask = gd->gd_runqmask;
578 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
579 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock())
581 mpheld = MP_LOCK_HELD();
582 if (ntd->td_toks && !lwkt_chktokens(ntd))
588 rqmask &= ~(1 << nq);
592 ntd = &gd->gd_idlethread;
593 ntd->td_flags |= TDF_IDLE_NOHLT;
595 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
596 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
599 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
600 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
603 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
604 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
608 * We have nothing to run but only let the idle loop halt
609 * the cpu if there are no pending interrupts.
611 ntd = &gd->gd_idlethread;
612 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
613 ntd->td_flags |= TDF_IDLE_NOHLT;
616 KASSERT(ntd->td_pri >= TDPRI_CRIT,
617 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
620 * Do the actual switch. If the new target does not need the MP lock
621 * and we are holding it, release the MP lock. If the new target requires
622 * the MP lock we have already acquired it for the target.
625 if (ntd->td_mpcount == 0 ) {
629 ASSERT_MP_LOCK_HELD();
636 /* NOTE: current cpu may have changed after switch */
641 * Request that the target thread preempt the current thread. Preemption
642 * only works under a specific set of conditions:
644 * - We are not preempting ourselves
645 * - The target thread is owned by the current cpu
646 * - We are not currently being preempted
647 * - The target is not currently being preempted
648 * - We are able to satisfy the target's MP lock requirements (if any).
650 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
651 * this is called via lwkt_schedule() through the td_preemptable callback.
652 * critpri is the managed critical priority that we should ignore in order
653 * to determine whether preemption is possible (aka usually just the crit
654 * priority of lwkt_schedule() itself).
656 * XXX at the moment we run the target thread in a critical section during
657 * the preemption in order to prevent the target from taking interrupts
658 * that *WE* can't. Preemption is strictly limited to interrupt threads
659 * and interrupt-like threads, outside of a critical section, and the
660 * preempted source thread will be resumed the instant the target blocks
661 * whether or not the source is scheduled (i.e. preemption is supposed to
662 * be as transparent as possible).
664 * The target thread inherits our MP count (added to its own) for the
665 * duration of the preemption in order to preserve the atomicy of the
666 * MP lock during the preemption. Therefore, any preempting targets must be
667 * careful in regards to MP assertions. Note that the MP count may be
668 * out of sync with the physical mp_lock, but we do not have to preserve
669 * the original ownership of the lock if it was out of synch (that is, we
670 * can leave it synchronized on return).
673 lwkt_preempt(thread_t ntd, int critpri)
675 struct globaldata *gd = mycpu;
683 * The caller has put us in a critical section. We can only preempt
684 * if the caller of the caller was not in a critical section (basically
685 * a local interrupt), as determined by the 'critpri' parameter.
687 * YYY The target thread must be in a critical section (else it must
688 * inherit our critical section? I dunno yet).
690 * Any tokens held by the target may not be held by thread(s) being
691 * preempted. We take the easy way out and do not preempt if
692 * the target is holding tokens.
694 * Set need_lwkt_resched() unconditionally for now YYY.
696 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
698 td = gd->gd_curthread;
699 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
703 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
709 if (ntd->td_gd != gd) {
716 * Take the easy way out and do not preempt if the target is holding
717 * one or more tokens. We could test whether the thread(s) being
718 * preempted interlock against the target thread's tokens and whether
719 * we can get all the target thread's tokens, but this situation
720 * should not occur very often so its easier to simply not preempt.
722 if (ntd->td_toks != NULL) {
727 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
732 if (ntd->td_preempted) {
739 * note: an interrupt might have occured just as we were transitioning
740 * to or from the MP lock. In this case td_mpcount will be pre-disposed
741 * (non-zero) but not actually synchronized with the actual state of the
742 * lock. We can use it to imply an MP lock requirement for the
743 * preemption but we cannot use it to test whether we hold the MP lock
746 savecnt = td->td_mpcount;
747 mpheld = MP_LOCK_HELD();
748 ntd->td_mpcount += td->td_mpcount;
749 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
750 ntd->td_mpcount -= td->td_mpcount;
758 * Since we are able to preempt the current thread, there is no need to
759 * call need_lwkt_resched().
762 ntd->td_preempted = td;
763 td->td_flags |= TDF_PREEMPT_LOCK;
765 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
767 KKASSERT(savecnt == td->td_mpcount);
768 mpheld = MP_LOCK_HELD();
769 if (mpheld && td->td_mpcount == 0)
771 else if (mpheld == 0 && td->td_mpcount)
772 panic("lwkt_preempt(): MP lock was not held through");
774 ntd->td_preempted = NULL;
775 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
779 * Yield our thread while higher priority threads are pending. This is
780 * typically called when we leave a critical section but it can be safely
781 * called while we are in a critical section.
783 * This function will not generally yield to equal priority threads but it
784 * can occur as a side effect. Note that lwkt_switch() is called from
785 * inside the critical section to prevent its own crit_exit() from reentering
786 * lwkt_yield_quick().
788 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
789 * came along but was blocked and made pending.
791 * (self contained on a per cpu basis)
794 lwkt_yield_quick(void)
796 globaldata_t gd = mycpu;
797 thread_t td = gd->gd_curthread;
800 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
801 * it with a non-zero cpl then we might not wind up calling splz after
802 * a task switch when the critical section is exited even though the
803 * new task could accept the interrupt.
805 * XXX from crit_exit() only called after last crit section is released.
806 * If called directly will run splz() even if in a critical section.
808 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
809 * except for this special case, we MUST call splz() here to handle any
810 * pending ints, particularly after we switch, or we might accidently
811 * halt the cpu with interrupts pending.
813 if (gd->gd_reqflags && td->td_nest_count < 2)
817 * YYY enabling will cause wakeup() to task-switch, which really
818 * confused the old 4.x code. This is a good way to simulate
819 * preemption and MP without actually doing preemption or MP, because a
820 * lot of code assumes that wakeup() does not block.
822 if (untimely_switch && td->td_nest_count == 0 &&
823 gd->gd_intr_nesting_level == 0
825 crit_enter_quick(td);
827 * YYY temporary hacks until we disassociate the userland scheduler
828 * from the LWKT scheduler.
830 if (td->td_flags & TDF_RUNQ) {
831 lwkt_switch(); /* will not reenter yield function */
833 lwkt_schedule_self(td); /* make sure we are scheduled */
834 lwkt_switch(); /* will not reenter yield function */
835 lwkt_deschedule_self(td); /* make sure we are descheduled */
837 crit_exit_noyield(td);
842 * This implements a normal yield which, unlike _quick, will yield to equal
843 * priority threads as well. Note that gd_reqflags tests will be handled by
844 * the crit_exit() call in lwkt_switch().
846 * (self contained on a per cpu basis)
851 lwkt_schedule_self(curthread);
856 * Generic schedule. Possibly schedule threads belonging to other cpus and
857 * deal with threads that might be blocked on a wait queue.
859 * We have a little helper inline function which does additional work after
860 * the thread has been enqueued, including dealing with preemption and
861 * setting need_lwkt_resched() (which prevents the kernel from returning
862 * to userland until it has processed higher priority threads).
866 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
868 if (ntd->td_preemptable) {
869 ntd->td_preemptable(ntd, cpri); /* YYY +token */
870 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
871 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
878 lwkt_schedule(thread_t td)
880 globaldata_t mygd = mycpu;
883 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
884 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
885 && td->td_proc->p_stat == SSLEEP
887 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
889 curthread->td_proc ? curthread->td_proc->p_pid : -1,
890 curthread->td_proc ? curthread->td_proc->p_stat : -1,
892 td->td_proc ? curthread->td_proc->p_pid : -1,
893 td->td_proc ? curthread->td_proc->p_stat : -1
895 panic("SCHED PANIC");
899 if (td == mygd->gd_curthread) {
905 * If the thread is on a wait list we have to send our scheduling
906 * request to the owner of the wait structure. Otherwise we send
907 * the scheduling request to the cpu owning the thread. Races
908 * are ok, the target will forward the message as necessary (the
909 * message may chase the thread around before it finally gets
912 * (remember, wait structures use stable storage)
914 * NOTE: we have to account for the number of critical sections
915 * under our control when calling _lwkt_schedule_post() so it
916 * can figure out whether preemption is allowed.
918 * NOTE: The wait structure algorithms are a mess and need to be
921 * NOTE: We cannot safely acquire or release a token, even
922 * non-blocking, because this routine may be called in the context
923 * of a thread already holding the token and thus not provide any
924 * interlock protection. We cannot safely manipulate the td_toks
925 * list for the same reason. Instead we depend on our critical
926 * section if the token is owned by our cpu.
928 if ((w = td->td_wait) != NULL) {
929 if (w->wa_token.t_cpu == mygd) {
930 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
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);
945 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
949 * If the wait structure is NULL and we own the thread, there
950 * is no race (since we are in a critical section). If we
951 * do not own the thread there might be a race but the
952 * target cpu will deal with it.
955 if (td->td_gd == mygd) {
957 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
959 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
963 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
971 * Managed acquisition. This code assumes that the MP lock is held for
972 * the tdallq operation and that the thread has been descheduled from its
973 * original cpu. We also have to wait for the thread to be entirely switched
974 * out on its original cpu (this is usually fast enough that we never loop)
975 * since the LWKT system does not have to hold the MP lock while switching
976 * and the target may have released it before switching.
979 lwkt_acquire(thread_t td)
987 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
988 while (td->td_flags & TDF_RUNNING) /* XXX spin */
992 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
994 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1000 * Generic deschedule. Descheduling threads other then your own should be
1001 * done only in carefully controlled circumstances. Descheduling is
1004 * This function may block if the cpu has run out of messages.
1007 lwkt_deschedule(thread_t td)
1010 if (td == curthread) {
1013 if (td->td_gd == mycpu) {
1016 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
1023 * Set the target thread's priority. This routine does not automatically
1024 * switch to a higher priority thread, LWKT threads are not designed for
1025 * continuous priority changes. Yield if you want to switch.
1027 * We have to retain the critical section count which uses the high bits
1028 * of the td_pri field. The specified priority may also indicate zero or
1029 * more critical sections by adding TDPRI_CRIT*N.
1031 * Note that we requeue the thread whether it winds up on a different runq
1032 * or not. uio_yield() depends on this and the routine is not normally
1033 * called with the same priority otherwise.
1036 lwkt_setpri(thread_t td, int pri)
1039 KKASSERT(td->td_gd == mycpu);
1041 if (td->td_flags & TDF_RUNQ) {
1043 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1046 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1052 lwkt_setpri_self(int pri)
1054 thread_t td = curthread;
1056 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1058 if (td->td_flags & TDF_RUNQ) {
1060 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1063 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1069 * Determine if there is a runnable thread at a higher priority then
1070 * the current thread. lwkt_setpri() does not check this automatically.
1071 * Return 1 if there is, 0 if there isn't.
1073 * Example: if bit 31 of runqmask is set and the current thread is priority
1074 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1076 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1077 * up comparing against 0xffffffff, a comparison that will always be false.
1080 lwkt_checkpri_self(void)
1082 globaldata_t gd = mycpu;
1083 thread_t td = gd->gd_curthread;
1084 int nq = td->td_pri & TDPRI_MASK;
1086 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1087 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1095 * Migrate the current thread to the specified cpu. The BGL must be held
1096 * (for the gd_tdallq manipulation XXX). This is accomplished by
1097 * descheduling ourselves from the current cpu, moving our thread to the
1098 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1099 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1102 static void lwkt_setcpu_remote(void *arg);
1106 lwkt_setcpu_self(globaldata_t rgd)
1109 thread_t td = curthread;
1111 if (td->td_gd != rgd) {
1112 crit_enter_quick(td);
1113 td->td_flags |= TDF_MIGRATING;
1114 lwkt_deschedule_self(td);
1115 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1116 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1117 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1119 /* we are now on the target cpu */
1120 crit_exit_quick(td);
1126 * Remote IPI for cpu migration (called while in a critical section so we
1127 * do not have to enter another one). The thread has already been moved to
1128 * our cpu's allq, but we must wait for the thread to be completely switched
1129 * out on the originating cpu before we schedule it on ours or the stack
1130 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1131 * change to main memory.
1133 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1134 * against wakeups. It is best if this interface is used only when there
1135 * are no pending events that might try to schedule the thread.
1139 lwkt_setcpu_remote(void *arg)
1142 globaldata_t gd = mycpu;
1144 while (td->td_flags & TDF_RUNNING)
1148 td->td_flags &= ~TDF_MIGRATING;
1154 lwkt_preempted_proc(void)
1156 thread_t td = curthread;
1157 while (td->td_preempted)
1158 td = td->td_preempted;
1159 return(td->td_proc);
1163 * Block on the specified wait queue until signaled. A generation number
1164 * must be supplied to interlock the wait queue. The function will
1165 * return immediately if the generation number does not match the wait
1166 * structure's generation number.
1169 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1171 thread_t td = curthread;
1174 lwkt_gettoken(&ilock, &w->wa_token);
1176 if (w->wa_gen == *gen) {
1178 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1181 td->td_wmesg = wmesg;
1184 if (td->td_wmesg != NULL) {
1191 lwkt_reltoken(&ilock);
1195 * Signal a wait queue. We gain ownership of the wait queue in order to
1196 * signal it. Once a thread is removed from the wait queue we have to
1197 * deal with the cpu owning the thread.
1199 * Note: alternatively we could message the target cpu owning the wait
1200 * queue. YYY implement as sysctl.
1203 lwkt_signal(lwkt_wait_t w, int count)
1208 lwkt_gettoken(&ilock, &w->wa_token);
1212 count = w->wa_count;
1213 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1216 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1218 td->td_wmesg = NULL;
1219 if (td->td_gd == mycpu) {
1222 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1226 lwkt_reltoken(&ilock);
1230 * Create a kernel process/thread/whatever. It shares it's address space
1231 * with proc0 - ie: kernel only.
1233 * NOTE! By default new threads are created with the MP lock held. A
1234 * thread which does not require the MP lock should release it by calling
1235 * rel_mplock() at the start of the new thread.
1238 lwkt_create(void (*func)(void *), void *arg,
1239 struct thread **tdp, thread_t template, int tdflags, int cpu,
1240 const char *fmt, ...)
1245 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
1248 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1249 td->td_flags |= TDF_VERBOSE | tdflags;
1255 * Set up arg0 for 'ps' etc
1257 __va_start(ap, fmt);
1258 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1262 * Schedule the thread to run
1264 if ((td->td_flags & TDF_STOPREQ) == 0)
1267 td->td_flags &= ~TDF_STOPREQ;
1272 * kthread_* is specific to the kernel and is not needed by userland.
1277 * Destroy an LWKT thread. Warning! This function is not called when
1278 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1279 * uses a different reaping mechanism.
1284 thread_t td = curthread;
1287 if (td->td_flags & TDF_VERBOSE)
1288 printf("kthread %p %s has exited\n", td, td->td_comm);
1290 crit_enter_quick(td);
1291 lwkt_deschedule_self(td);
1293 KKASSERT(gd == td->td_gd);
1294 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1295 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1296 ++gd->gd_tdfreecount;
1297 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1302 #endif /* _KERNEL */
1307 thread_t td = curthread;
1308 int lpri = td->td_pri;
1311 panic("td_pri is/would-go negative! %p %d", td, lpri);