2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * Each cpu in a system has its own self-contained light weight kernel
27 * thread scheduler, which means that generally speaking we only need
28 * to use a critical section to avoid problems. Foreign thread
29 * scheduling is queued via (async) IPIs.
31 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.22 2003/07/11 22:30:09 dillon Exp $
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
38 #include <sys/rtprio.h>
39 #include <sys/queue.h>
40 #include <sys/thread2.h>
41 #include <sys/sysctl.h>
42 #include <sys/kthread.h>
43 #include <machine/cpu.h>
47 #include <vm/vm_param.h>
48 #include <vm/vm_kern.h>
49 #include <vm/vm_object.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_map.h>
52 #include <vm/vm_pager.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_zone.h>
56 #include <machine/stdarg.h>
57 #include <machine/ipl.h>
59 #include <machine/smp.h>
62 static int untimely_switch = 0;
63 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
65 static int token_debug = 0;
66 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
68 static quad_t switch_count = 0;
69 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
70 static quad_t preempt_hit = 0;
71 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
72 static quad_t preempt_miss = 0;
73 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
74 static quad_t preempt_weird = 0;
75 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
76 static quad_t ipiq_count = 0;
77 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
78 static quad_t ipiq_fifofull = 0;
79 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
82 * These helper procedures handle the runq, they can only be called from
83 * within a critical section.
87 _lwkt_dequeue(thread_t td)
89 if (td->td_flags & TDF_RUNQ) {
90 int nq = td->td_pri & TDPRI_MASK;
91 struct globaldata *gd = mycpu;
93 td->td_flags &= ~TDF_RUNQ;
94 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
95 /* runqmask is passively cleaned up by the switcher */
101 _lwkt_enqueue(thread_t td)
103 if ((td->td_flags & TDF_RUNQ) == 0) {
104 int nq = td->td_pri & TDPRI_MASK;
105 struct globaldata *gd = mycpu;
107 td->td_flags |= TDF_RUNQ;
108 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
109 gd->gd_runqmask |= 1 << nq;
112 * YYY needs cli/sti protection? gd_reqpri set by interrupt
113 * when made pending. need better mechanism.
115 if (gd->gd_reqpri < (td->td_pri & TDPRI_MASK))
116 gd->gd_reqpri = (td->td_pri & TDPRI_MASK);
123 _lwkt_wantresched(thread_t ntd, thread_t cur)
125 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
129 * LWKTs operate on a per-cpu basis
131 * WARNING! Called from early boot, 'mycpu' may not work yet.
134 lwkt_gdinit(struct globaldata *gd)
138 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
139 TAILQ_INIT(&gd->gd_tdrunq[i]);
141 TAILQ_INIT(&gd->gd_tdallq);
145 * Initialize a thread wait structure prior to first use.
147 * NOTE! called from low level boot code, we cannot do anything fancy!
150 lwkt_init_wait(lwkt_wait_t w)
152 TAILQ_INIT(&w->wa_waitq);
156 * Create a new thread. The thread must be associated with a process context
157 * or LWKT start address before it can be scheduled.
159 * If you intend to create a thread without a process context this function
160 * does everything except load the startup and switcher function.
163 lwkt_alloc_thread(struct thread *td)
170 if (mycpu->gd_tdfreecount > 0) {
171 --mycpu->gd_tdfreecount;
172 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
173 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
174 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
175 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
177 stack = td->td_kstack;
178 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
181 td = zalloc(thread_zone);
182 td->td_kstack = NULL;
183 flags |= TDF_ALLOCATED_THREAD;
186 if ((stack = td->td_kstack) == NULL) {
187 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
188 flags |= TDF_ALLOCATED_STACK;
190 lwkt_init_thread(td, stack, flags, mycpu);
195 * Initialize a preexisting thread structure. This function is used by
196 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
198 * NOTE! called from low level boot code, we cannot do anything fancy!
201 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
203 bzero(td, sizeof(struct thread));
204 td->td_kstack = stack;
205 td->td_flags |= flags;
207 td->td_pri = TDPRI_CRIT;
208 td->td_cpu = gd->gd_cpuid; /* YYY don't really need this if have td_gd */
209 pmap_init_thread(td);
211 TAILQ_INSERT_TAIL(&mycpu->gd_tdallq, td, td_allq);
216 lwkt_set_comm(thread_t td, const char *ctl, ...)
221 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
226 lwkt_hold(thread_t td)
232 lwkt_rele(thread_t td)
234 KKASSERT(td->td_refs > 0);
239 lwkt_wait_free(thread_t td)
242 tsleep(td, PWAIT, "tdreap", hz);
246 lwkt_free_thread(thread_t td)
248 struct globaldata *gd = mycpu;
250 KASSERT((td->td_flags & TDF_RUNNING) == 0,
251 ("lwkt_free_thread: did not exit! %p", td));
254 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
255 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
256 (td->td_flags & TDF_ALLOCATED_THREAD)
258 ++gd->gd_tdfreecount;
259 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
263 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
264 kmem_free(kernel_map,
265 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
267 td->td_kstack = NULL;
269 if (td->td_flags & TDF_ALLOCATED_THREAD)
270 zfree(thread_zone, td);
276 * Switch to the next runnable lwkt. If no LWKTs are runnable then
277 * switch to the idlethread. Switching must occur within a critical
278 * section to avoid races with the scheduling queue.
280 * We always have full control over our cpu's run queue. Other cpus
281 * that wish to manipulate our queue must use the cpu_*msg() calls to
282 * talk to our cpu, so a critical section is all that is needed and
283 * the result is very, very fast thread switching.
285 * The LWKT scheduler uses a fixed priority model and round-robins at
286 * each priority level. User process scheduling is a totally
287 * different beast and LWKT priorities should not be confused with
288 * user process priorities.
290 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
291 * cleans it up. Note that the td_switch() function cannot do anything that
292 * requires the MP lock since the MP lock will have already been setup for
293 * the target thread (not the current thread).
299 struct globaldata *gd;
300 thread_t td = curthread;
306 if (mycpu->gd_intr_nesting_level &&
307 td->td_preempted == NULL && panicstr == NULL
309 panic("lwkt_switch: cannot switch from within an interrupt, yet\n");
313 * Passive release (used to transition from user to kernel mode
314 * when we block or switch rather then when we enter the kernel).
315 * This function is NOT called if we are switching into a preemption
316 * or returning from a preemption. Typically this causes us to lose
317 * our P_CURPROC designation (if we have one) and become a true LWKT
318 * thread, and may also hand P_CURPROC to another process and schedule
329 * td_mpcount cannot be used to determine if we currently hold the
330 * MP lock because get_mplock() will increment it prior to attempting
331 * to get the lock, and switch out if it can't. Look at the actual lock.
333 mpheld = MP_LOCK_HELD();
335 if ((ntd = td->td_preempted) != NULL) {
337 * We had preempted another thread on this cpu, resume the preempted
338 * thread. This occurs transparently, whether the preempted thread
339 * was scheduled or not (it may have been preempted after descheduling
342 * We have to setup the MP lock for the original thread after backing
343 * out the adjustment that was made to curthread when the original
346 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
348 if (ntd->td_mpcount && mpheld == 0) {
349 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
350 td, ntd, td->td_mpcount, ntd->td_mpcount);
352 if (ntd->td_mpcount) {
353 td->td_mpcount -= ntd->td_mpcount;
354 KKASSERT(td->td_mpcount >= 0);
357 ntd->td_flags |= TDF_PREEMPT_DONE;
358 /* YYY release mp lock on switchback if original doesn't need it */
361 * Priority queue / round-robin at each priority. Note that user
362 * processes run at a fixed, low priority and the user process
363 * scheduler deals with interactions between user processes
364 * by scheduling and descheduling them from the LWKT queue as
367 * We have to adjust the MP lock for the target thread. If we
368 * need the MP lock and cannot obtain it we try to locate a
369 * thread that does not need the MP lock.
373 if (gd->gd_runqmask) {
374 int nq = bsrl(gd->gd_runqmask);
375 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
376 gd->gd_runqmask &= ~(1 << nq);
380 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
382 * Target needs MP lock and we couldn't get it, try
383 * to locate a thread which does not need the MP lock
384 * to run. If we cannot locate a thread spin in idle.
386 u_int32_t rqmask = gd->gd_runqmask;
388 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
389 if (ntd->td_mpcount == 0)
394 rqmask &= ~(1 << nq);
398 ntd = &gd->gd_idlethread;
399 ntd->td_flags |= TDF_IDLE_NOHLT;
401 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
402 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
405 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
406 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
409 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
410 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
414 * Nothing to run but we may still need the BGL to deal with
415 * pending interrupts, spin in idle if so.
417 ntd = &gd->gd_idlethread;
419 ntd->td_flags |= TDF_IDLE_NOHLT;
422 KASSERT(ntd->td_pri >= TDPRI_CRIT,
423 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
426 * Do the actual switch. If the new target does not need the MP lock
427 * and we are holding it, release the MP lock. If the new target requires
428 * the MP lock we have already acquired it for the target.
431 if (ntd->td_mpcount == 0 ) {
435 ASSERT_MP_LOCK_HELD();
446 * Switch if another thread has a higher priority. Do not switch to other
447 * threads at the same priority.
452 struct globaldata *gd = mycpu;
453 struct thread *td = gd->gd_curthread;
455 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
461 * Request that the target thread preempt the current thread. Preemption
462 * only works under a specific set of conditions:
464 * - We are not preempting ourselves
465 * - The target thread is owned by the current cpu
466 * - We are not currently being preempted
467 * - The target is not currently being preempted
468 * - We are able to satisfy the target's MP lock requirements (if any).
470 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
471 * this is called via lwkt_schedule() through the td_preemptable callback.
472 * critpri is the managed critical priority that we should ignore in order
473 * to determine whether preemption is possible (aka usually just the crit
474 * priority of lwkt_schedule() itself).
476 * XXX at the moment we run the target thread in a critical section during
477 * the preemption in order to prevent the target from taking interrupts
478 * that *WE* can't. Preemption is strictly limited to interrupt threads
479 * and interrupt-like threads, outside of a critical section, and the
480 * preempted source thread will be resumed the instant the target blocks
481 * whether or not the source is scheduled (i.e. preemption is supposed to
482 * be as transparent as possible).
484 * The target thread inherits our MP count (added to its own) for the
485 * duration of the preemption in order to preserve the atomicy of the
486 * MP lock during the preemption. Therefore, any preempting targets must be
487 * careful in regards to MP assertions. Note that the MP count may be
488 * out of sync with the physical mp_lock. If we preempt we have to preserve
489 * the expected situation.
492 lwkt_preempt(thread_t ntd, int critpri)
494 thread_t td = curthread;
501 * The caller has put us in a critical section. We can only preempt
502 * if the caller of the caller was not in a critical section (basically
503 * a local interrupt), as determined by the 'critpri' parameter. If
504 * we are unable to preempt
506 * YYY The target thread must be in a critical section (else it must
507 * inherit our critical section? I dunno yet).
509 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
512 if (!_lwkt_wantresched(ntd, td)) {
516 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
521 if (ntd->td_cpu != mycpu->gd_cpuid) {
526 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
530 if (ntd->td_preempted) {
536 * note: an interrupt might have occured just as we were transitioning
537 * to the MP lock. In this case td_mpcount will be pre-disposed but
538 * not actually synchronized with the actual state of the lock. We
539 * can use it to imply an MP lock requirement for the preemption but
540 * we cannot use it to test whether we hold the MP lock or not.
542 mpheld = MP_LOCK_HELD();
543 if (mpheld && td->td_mpcount == 0)
544 panic("lwkt_preempt(): held and no count");
545 savecnt = td->td_mpcount;
546 ntd->td_mpcount += td->td_mpcount;
547 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
548 ntd->td_mpcount -= td->td_mpcount;
555 ntd->td_preempted = td;
556 td->td_flags |= TDF_PREEMPT_LOCK;
558 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
560 KKASSERT(savecnt == td->td_mpcount);
561 if (mpheld == 0 && MP_LOCK_HELD())
563 else if (mpheld && !MP_LOCK_HELD())
564 panic("lwkt_preempt(): MP lock was not held through");
566 ntd->td_preempted = NULL;
567 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
571 * Yield our thread while higher priority threads are pending. This is
572 * typically called when we leave a critical section but it can be safely
573 * called while we are in a critical section.
575 * This function will not generally yield to equal priority threads but it
576 * can occur as a side effect. Note that lwkt_switch() is called from
577 * inside the critical section to pervent its own crit_exit() from reentering
578 * lwkt_yield_quick().
580 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
581 * came along but was blocked and made pending.
583 * (self contained on a per cpu basis)
586 lwkt_yield_quick(void)
588 thread_t td = curthread;
591 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
592 * it with a non-zero cpl then we might not wind up calling splz after
593 * a task switch when the critical section is exited even though the
594 * new task could accept the interrupt. YYY alternative is to have
595 * lwkt_switch() just call splz unconditionally.
597 * XXX from crit_exit() only called after last crit section is released.
598 * If called directly will run splz() even if in a critical section.
600 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
605 * YYY enabling will cause wakeup() to task-switch, which really
606 * confused the old 4.x code. This is a good way to simulate
607 * preemption and MP without actually doing preemption or MP, because a
608 * lot of code assumes that wakeup() does not block.
610 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
613 * YYY temporary hacks until we disassociate the userland scheduler
614 * from the LWKT scheduler.
616 if (td->td_flags & TDF_RUNQ) {
617 lwkt_switch(); /* will not reenter yield function */
619 lwkt_schedule_self(); /* make sure we are scheduled */
620 lwkt_switch(); /* will not reenter yield function */
621 lwkt_deschedule_self(); /* make sure we are descheduled */
628 * This implements a normal yield which, unlike _quick, will yield to equal
629 * priority threads as well. Note that gd_reqpri tests will be handled by
630 * the crit_exit() call in lwkt_switch().
632 * (self contained on a per cpu basis)
637 lwkt_schedule_self();
642 * Schedule a thread to run. As the current thread we can always safely
643 * schedule ourselves, and a shortcut procedure is provided for that
646 * (non-blocking, self contained on a per cpu basis)
649 lwkt_schedule_self(void)
651 thread_t td = curthread;
654 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
656 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
657 panic("SCHED SELF PANIC");
662 * Generic schedule. Possibly schedule threads belonging to other cpus and
663 * deal with threads that might be blocked on a wait queue.
665 * YYY this is one of the best places to implement load balancing code.
666 * Load balancing can be accomplished by requesting other sorts of actions
667 * for the thread in question.
670 lwkt_schedule(thread_t td)
673 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
674 && td->td_proc->p_stat == SSLEEP
676 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
678 curthread->td_proc ? curthread->td_proc->p_pid : -1,
679 curthread->td_proc ? curthread->td_proc->p_stat : -1,
681 td->td_proc ? curthread->td_proc->p_pid : -1,
682 td->td_proc ? curthread->td_proc->p_stat : -1
684 panic("SCHED PANIC");
688 if (td == curthread) {
694 * If the thread is on a wait list we have to send our scheduling
695 * request to the owner of the wait structure. Otherwise we send
696 * the scheduling request to the cpu owning the thread. Races
697 * are ok, the target will forward the message as necessary (the
698 * message may chase the thread around before it finally gets
701 * (remember, wait structures use stable storage)
703 if ((w = td->td_wait) != NULL) {
704 if (lwkt_trytoken(&w->wa_token)) {
705 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
708 if (td->td_cpu == mycpu->gd_cpuid) {
710 if (td->td_preemptable) {
711 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
712 } else if (_lwkt_wantresched(td, curthread)) {
716 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
718 lwkt_reltoken(&w->wa_token);
720 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
724 * If the wait structure is NULL and we own the thread, there
725 * is no race (since we are in a critical section). If we
726 * do not own the thread there might be a race but the
727 * target cpu will deal with it.
729 if (td->td_cpu == mycpu->gd_cpuid) {
731 if (td->td_preemptable) {
732 td->td_preemptable(td, TDPRI_CRIT);
733 } else if (_lwkt_wantresched(td, curthread)) {
737 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
745 * Managed acquisition. This code assumes that the MP lock is held for
746 * the tdallq operation and that the thread has been descheduled from its
747 * original cpu. We also have to wait for the thread to be entirely switched
748 * out on its original cpu (this is usually fast enough that we never loop)
749 * since the LWKT system does not have to hold the MP lock while switching
750 * and the target may have released it before switching.
753 lwkt_acquire(thread_t td)
755 struct globaldata *gd;
759 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
760 while (td->td_flags & TDF_RUNNING) /* XXX spin */
765 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
768 td->td_cpu = gd->gd_cpuid;
769 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
775 * Deschedule a thread.
777 * (non-blocking, self contained on a per cpu basis)
780 lwkt_deschedule_self(void)
782 thread_t td = curthread;
785 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
791 * Generic deschedule. Descheduling threads other then your own should be
792 * done only in carefully controlled circumstances. Descheduling is
795 * This function may block if the cpu has run out of messages.
798 lwkt_deschedule(thread_t td)
801 if (td == curthread) {
804 if (td->td_cpu == mycpu->gd_cpuid) {
807 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
814 * Set the target thread's priority. This routine does not automatically
815 * switch to a higher priority thread, LWKT threads are not designed for
816 * continuous priority changes. Yield if you want to switch.
818 * We have to retain the critical section count which uses the high bits
819 * of the td_pri field. The specified priority may also indicate zero or
820 * more critical sections by adding TDPRI_CRIT*N.
823 lwkt_setpri(thread_t td, int pri)
826 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
828 if (td->td_flags & TDF_RUNQ) {
830 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
833 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
839 lwkt_setpri_self(int pri)
841 thread_t td = curthread;
843 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
845 if (td->td_flags & TDF_RUNQ) {
847 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
850 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
856 lwkt_preempted_proc(void)
858 thread_t td = curthread;
859 while (td->td_preempted)
860 td = td->td_preempted;
866 * This function deschedules the current thread and blocks on the specified
867 * wait queue. We obtain ownership of the wait queue in order to block
868 * on it. A generation number is used to interlock the wait queue in case
869 * it gets signalled while we are blocked waiting on the token.
871 * Note: alternatively we could dequeue our thread and then message the
872 * target cpu owning the wait queue. YYY implement as sysctl.
874 * Note: wait queue signals normally ping-pong the cpu as an optimization.
876 typedef struct lwkt_gettoken_req {
882 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
884 thread_t td = curthread;
886 lwkt_gettoken(&w->wa_token);
887 if (w->wa_gen == *gen) {
889 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
892 td->td_wmesg = wmesg;
895 /* token might be lost, doesn't matter for gen update */
897 lwkt_reltoken(&w->wa_token);
901 * Signal a wait queue. We gain ownership of the wait queue in order to
902 * signal it. Once a thread is removed from the wait queue we have to
903 * deal with the cpu owning the thread.
905 * Note: alternatively we could message the target cpu owning the wait
906 * queue. YYY implement as sysctl.
909 lwkt_signal(lwkt_wait_t w)
914 lwkt_gettoken(&w->wa_token);
917 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
920 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
923 if (td->td_cpu == mycpu->gd_cpuid) {
926 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
928 lwkt_regettoken(&w->wa_token);
930 lwkt_reltoken(&w->wa_token);
934 * Acquire ownership of a token
936 * Acquire ownership of a token. The token may have spl and/or critical
937 * section side effects, depending on its purpose. These side effects
938 * guarentee that you will maintain ownership of the token as long as you
939 * do not block. If you block you may lose access to the token (but you
940 * must still release it even if you lose your access to it).
942 * YYY for now we use a critical section to prevent IPIs from taking away
943 * a token, but do we really only need to disable IPIs ?
945 * YYY certain tokens could be made to act like mutexes when performance
946 * would be better (e.g. t_cpu == -1). This is not yet implemented.
948 * YYY the tokens replace 4.x's simplelocks for the most part, but this
949 * means that 4.x does not expect a switch so for now we cannot switch
950 * when waiting for an IPI to be returned.
952 * YYY If the token is owned by another cpu we may have to send an IPI to
953 * it and then block. The IPI causes the token to be given away to the
954 * requesting cpu, unless it has already changed hands. Since only the
955 * current cpu can give away a token it owns we do not need a memory barrier.
956 * This needs serious optimization.
963 lwkt_gettoken_remote(void *arg)
965 lwkt_gettoken_req *req = arg;
966 if (req->tok->t_cpu == mycpu->gd_cpuid) {
968 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
969 req->tok->t_cpu = req->cpu;
970 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
971 /* else set reqcpu to point to current cpu for release */
978 lwkt_gettoken(lwkt_token_t tok)
981 * Prevent preemption so the token can't be taken away from us once
982 * we gain ownership of it. Use a synchronous request which might
983 * block. The request will be forwarded as necessary playing catchup
989 if (curthread->td_pri > 2000) {
990 curthread->td_pri = 1000;
995 while (tok->t_cpu != mycpu->gd_cpuid) {
996 struct lwkt_gettoken_req req;
1000 req.cpu = mycpu->gd_cpuid;
1002 dcpu = (volatile int)tok->t_cpu;
1003 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1005 printf("REQT%d ", dcpu);
1006 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1007 lwkt_wait_ipiq(dcpu, seq);
1009 printf("REQR%d ", tok->t_cpu);
1013 * leave us in a critical section on return. This will be undone
1014 * by lwkt_reltoken(). Bump the generation number.
1016 return(++tok->t_gen);
1020 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1024 lwkt_trytoken(lwkt_token_t tok)
1028 if (tok->t_cpu != mycpu->gd_cpuid) {
1032 /* leave us in the critical section */
1038 * Release your ownership of a token. Releases must occur in reverse
1039 * order to aquisitions, eventually so priorities can be unwound properly
1040 * like SPLs. At the moment the actual implemention doesn't care.
1042 * We can safely hand a token that we own to another cpu without notifying
1043 * it, but once we do we can't get it back without requesting it (unless
1044 * the other cpu hands it back to us before we check).
1046 * We might have lost the token, so check that.
1049 lwkt_reltoken(lwkt_token_t tok)
1051 if (tok->t_cpu == mycpu->gd_cpuid) {
1052 tok->t_cpu = tok->t_reqcpu;
1058 * Reacquire a token that might have been lost and compare and update the
1059 * generation number. 0 is returned if the generation has not changed
1060 * (nobody else obtained the token while we were blocked, on this cpu or
1063 * This function returns with the token re-held whether the generation
1064 * number changed or not.
1067 lwkt_gentoken(lwkt_token_t tok, int *gen)
1069 if (lwkt_regettoken(tok) == *gen) {
1079 * Re-acquire a token that might have been lost. Returns the generation
1080 * number of the token.
1083 lwkt_regettoken(lwkt_token_t tok)
1085 /* assert we are in a critical section */
1086 if (tok->t_cpu != mycpu->gd_cpuid) {
1088 while (tok->t_cpu != mycpu->gd_cpuid) {
1089 struct lwkt_gettoken_req req;
1093 req.cpu = mycpu->gd_cpuid;
1095 dcpu = (volatile int)tok->t_cpu;
1096 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1098 printf("REQT%d ", dcpu);
1099 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1100 lwkt_wait_ipiq(dcpu, seq);
1102 printf("REQR%d ", tok->t_cpu);
1111 lwkt_inittoken(lwkt_token_t tok)
1114 * Zero structure and set cpu owner and reqcpu to cpu 0.
1116 bzero(tok, sizeof(*tok));
1120 * Create a kernel process/thread/whatever. It shares it's address space
1121 * with proc0 - ie: kernel only.
1123 * XXX should be renamed to lwkt_create()
1125 * The thread will be entered with the MP lock held.
1128 lwkt_create(void (*func)(void *), void *arg,
1129 struct thread **tdp, thread_t template, int tdflags,
1130 const char *fmt, ...)
1135 td = lwkt_alloc_thread(template);
1138 cpu_set_thread_handler(td, kthread_exit, func, arg);
1139 td->td_flags |= TDF_VERBOSE | tdflags;
1145 * Set up arg0 for 'ps' etc
1148 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1152 * Schedule the thread to run
1154 if ((td->td_flags & TDF_STOPREQ) == 0)
1157 td->td_flags &= ~TDF_STOPREQ;
1162 * Destroy an LWKT thread. Warning! This function is not called when
1163 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1164 * uses a different reaping mechanism.
1169 thread_t td = curthread;
1171 if (td->td_flags & TDF_VERBOSE)
1172 printf("kthread %p %s has exited\n", td, td->td_comm);
1174 lwkt_deschedule_self();
1175 ++mycpu->gd_tdfreecount;
1176 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1181 * Create a kernel process/thread/whatever. It shares it's address space
1182 * with proc0 - ie: kernel only. 5.x compatible.
1185 kthread_create(void (*func)(void *), void *arg,
1186 struct thread **tdp, const char *fmt, ...)
1191 td = lwkt_alloc_thread(NULL);
1194 cpu_set_thread_handler(td, kthread_exit, func, arg);
1195 td->td_flags |= TDF_VERBOSE;
1201 * Set up arg0 for 'ps' etc
1204 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1208 * Schedule the thread to run
1217 thread_t td = curthread;
1218 int lpri = td->td_pri;
1221 panic("td_pri is/would-go negative! %p %d", td, lpri);
1225 * Destroy an LWKT thread. Warning! This function is not called when
1226 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1227 * uses a different reaping mechanism.
1229 * XXX duplicates lwkt_exit()
1240 * Send a function execution request to another cpu. The request is queued
1241 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1242 * possible target cpu. The FIFO can be written.
1244 * YYY If the FIFO fills up we have to enable interrupts and process the
1245 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1246 * Create a CPU_*() function to do this!
1248 * Must be called from a critical section.
1251 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1255 struct globaldata *gd = mycpu;
1257 if (dcpu == gd->gd_cpuid) {
1262 ++gd->gd_intr_nesting_level;
1264 if (gd->gd_intr_nesting_level > 20)
1265 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1267 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1268 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1270 ip = &gd->gd_ipiq[dcpu];
1273 * We always drain before the FIFO becomes full so it should never
1274 * become full. We need to leave enough entries to deal with
1277 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1278 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1279 ip->ip_func[windex] = func;
1280 ip->ip_arg[windex] = arg;
1281 /* YYY memory barrier */
1283 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1284 unsigned int eflags = read_eflags();
1287 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1288 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1289 lwkt_process_ipiq();
1291 write_eflags(eflags);
1293 --gd->gd_intr_nesting_level;
1294 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1296 return(ip->ip_windex);
1300 * Send a message to several target cpus. Typically used for scheduling.
1303 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1309 lwkt_send_ipiq(cpuid, func, arg);
1310 mask &= ~(1 << cpuid);
1315 * Wait for the remote cpu to finish processing a function.
1317 * YYY we have to enable interrupts and process the IPIQ while waiting
1318 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1319 * function to do this! YYY we really should 'block' here.
1321 * Must be called from a critical section. Thsi routine may be called
1322 * from an interrupt (for example, if an interrupt wakes a foreign thread
1326 lwkt_wait_ipiq(int dcpu, int seq)
1329 int maxc = 100000000;
1331 if (dcpu != mycpu->gd_cpuid) {
1332 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1333 ip = &mycpu->gd_ipiq[dcpu];
1334 if ((int)(ip->ip_xindex - seq) < 0) {
1335 unsigned int eflags = read_eflags();
1337 while ((int)(ip->ip_xindex - seq) < 0) {
1338 lwkt_process_ipiq();
1340 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1341 if (maxc < -1000000)
1342 panic("LWKT_WAIT_IPIQ");
1344 write_eflags(eflags);
1350 * Called from IPI interrupt (like a fast interrupt), which has placed
1351 * us in a critical section. The MP lock may or may not be held.
1352 * May also be called from doreti or splz, or be reentrantly called
1353 * indirectly through the ip_func[] we run.
1356 lwkt_process_ipiq(void)
1359 int cpuid = mycpu->gd_cpuid;
1361 for (n = 0; n < ncpus; ++n) {
1367 ip = globaldata_find(n)->gd_ipiq;
1373 * Note: xindex is only updated after we are sure the function has
1374 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1375 * function may send an IPI which may block/drain.
1377 while (ip->ip_rindex != ip->ip_windex) {
1378 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1380 ip->ip_func[ri](ip->ip_arg[ri]);
1381 /* YYY memory barrier */
1382 ip->ip_xindex = ip->ip_rindex;
1390 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1392 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1393 return(0); /* NOT REACHED */
1397 lwkt_wait_ipiq(int dcpu, int seq)
1399 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);