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 1.31 2003/09/25 23:49:09 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;
115 _lwkt_wantresched(thread_t ntd, thread_t cur)
117 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
121 * LWKTs operate on a per-cpu basis
123 * WARNING! Called from early boot, 'mycpu' may not work yet.
126 lwkt_gdinit(struct globaldata *gd)
130 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
131 TAILQ_INIT(&gd->gd_tdrunq[i]);
133 TAILQ_INIT(&gd->gd_tdallq);
137 * Initialize a thread wait structure prior to first use.
139 * NOTE! called from low level boot code, we cannot do anything fancy!
142 lwkt_init_wait(lwkt_wait_t w)
144 TAILQ_INIT(&w->wa_waitq);
148 * Create a new thread. The thread must be associated with a process context
149 * or LWKT start address before it can be scheduled.
151 * If you intend to create a thread without a process context this function
152 * does everything except load the startup and switcher function.
155 lwkt_alloc_thread(struct thread *td)
162 if (mycpu->gd_tdfreecount > 0) {
163 --mycpu->gd_tdfreecount;
164 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
165 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
166 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
167 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
169 stack = td->td_kstack;
170 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
173 td = zalloc(thread_zone);
174 td->td_kstack = NULL;
175 flags |= TDF_ALLOCATED_THREAD;
178 if ((stack = td->td_kstack) == NULL) {
179 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
180 flags |= TDF_ALLOCATED_STACK;
182 lwkt_init_thread(td, stack, flags, mycpu);
187 * Initialize a preexisting thread structure. This function is used by
188 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
190 * NOTE! called from low level boot code, we cannot do anything fancy!
191 * Only the low level boot code will call this function with gd != mycpu.
193 * All threads start out in a critical section at a priority of
194 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
198 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
200 bzero(td, sizeof(struct thread));
201 td->td_kstack = stack;
202 td->td_flags |= flags;
204 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
205 lwkt_init_port(&td->td_msgport, td);
206 pmap_init_thread(td);
208 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
213 lwkt_set_comm(thread_t td, const char *ctl, ...)
218 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
223 lwkt_hold(thread_t td)
229 lwkt_rele(thread_t td)
231 KKASSERT(td->td_refs > 0);
236 lwkt_wait_free(thread_t td)
239 tsleep(td, 0, "tdreap", hz);
243 lwkt_free_thread(thread_t td)
245 struct globaldata *gd = mycpu;
247 KASSERT((td->td_flags & TDF_RUNNING) == 0,
248 ("lwkt_free_thread: did not exit! %p", td));
251 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
252 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
253 (td->td_flags & TDF_ALLOCATED_THREAD)
255 ++gd->gd_tdfreecount;
256 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
260 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
261 kmem_free(kernel_map,
262 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
264 td->td_kstack = NULL;
266 if (td->td_flags & TDF_ALLOCATED_THREAD)
267 zfree(thread_zone, td);
273 * Switch to the next runnable lwkt. If no LWKTs are runnable then
274 * switch to the idlethread. Switching must occur within a critical
275 * section to avoid races with the scheduling queue.
277 * We always have full control over our cpu's run queue. Other cpus
278 * that wish to manipulate our queue must use the cpu_*msg() calls to
279 * talk to our cpu, so a critical section is all that is needed and
280 * the result is very, very fast thread switching.
282 * The LWKT scheduler uses a fixed priority model and round-robins at
283 * each priority level. User process scheduling is a totally
284 * different beast and LWKT priorities should not be confused with
285 * user process priorities.
287 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
288 * cleans it up. Note that the td_switch() function cannot do anything that
289 * requires the MP lock since the MP lock will have already been setup for
290 * the target thread (not the current thread). It's nice to have a scheduler
291 * that does not need the MP lock to work because it allows us to do some
292 * really cool high-performance MP lock optimizations.
298 struct globaldata *gd;
299 thread_t td = curthread;
305 if (mycpu->gd_intr_nesting_level &&
306 td->td_preempted == NULL && panicstr == NULL
308 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
312 * Passive release (used to transition from user to kernel mode
313 * when we block or switch rather then when we enter the kernel).
314 * This function is NOT called if we are switching into a preemption
315 * or returning from a preemption. Typically this causes us to lose
316 * our P_CURPROC designation (if we have one) and become a true LWKT
317 * thread, and may also hand P_CURPROC to another process and schedule
328 * td_mpcount cannot be used to determine if we currently hold the
329 * MP lock because get_mplock() will increment it prior to attempting
330 * to get the lock, and switch out if it can't. Our ownership of
331 * the actual lock will remain stable while we are in a critical section
332 * (but, of course, another cpu may own or release the lock so the
333 * actual value of mp_lock is not stable).
335 mpheld = MP_LOCK_HELD();
337 if ((ntd = td->td_preempted) != NULL) {
339 * We had preempted another thread on this cpu, resume the preempted
340 * thread. This occurs transparently, whether the preempted thread
341 * was scheduled or not (it may have been preempted after descheduling
344 * We have to setup the MP lock for the original thread after backing
345 * out the adjustment that was made to curthread when the original
348 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
350 if (ntd->td_mpcount && mpheld == 0) {
351 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
352 td, ntd, td->td_mpcount, ntd->td_mpcount);
354 if (ntd->td_mpcount) {
355 td->td_mpcount -= ntd->td_mpcount;
356 KKASSERT(td->td_mpcount >= 0);
359 ntd->td_flags |= TDF_PREEMPT_DONE;
360 /* YYY release mp lock on switchback if original doesn't need it */
363 * Priority queue / round-robin at each priority. Note that user
364 * processes run at a fixed, low priority and the user process
365 * scheduler deals with interactions between user processes
366 * by scheduling and descheduling them from the LWKT queue as
369 * We have to adjust the MP lock for the target thread. If we
370 * need the MP lock and cannot obtain it we try to locate a
371 * thread that does not need the MP lock.
375 if (gd->gd_runqmask) {
376 int nq = bsrl(gd->gd_runqmask);
377 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
378 gd->gd_runqmask &= ~(1 << nq);
382 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
384 * Target needs MP lock and we couldn't get it, try
385 * to locate a thread which does not need the MP lock
386 * to run. If we cannot locate a thread spin in idle.
388 u_int32_t rqmask = gd->gd_runqmask;
390 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
391 if (ntd->td_mpcount == 0)
396 rqmask &= ~(1 << nq);
400 ntd = &gd->gd_idlethread;
401 ntd->td_flags |= TDF_IDLE_NOHLT;
403 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
404 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
407 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
408 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
411 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
412 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
416 * Nothing to run but we may still need the BGL to deal with
417 * pending interrupts, spin in idle if so.
419 ntd = &gd->gd_idlethread;
421 ntd->td_flags |= TDF_IDLE_NOHLT;
424 KASSERT(ntd->td_pri >= TDPRI_CRIT,
425 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
428 * Do the actual switch. If the new target does not need the MP lock
429 * and we are holding it, release the MP lock. If the new target requires
430 * the MP lock we have already acquired it for the target.
433 if (ntd->td_mpcount == 0 ) {
437 ASSERT_MP_LOCK_HELD();
448 * Switch if another thread has a higher priority. Do not switch to other
449 * threads at the same priority.
454 struct globaldata *gd = mycpu;
455 struct thread *td = gd->gd_curthread;
457 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
463 * Request that the target thread preempt the current thread. Preemption
464 * only works under a specific set of conditions:
466 * - We are not preempting ourselves
467 * - The target thread is owned by the current cpu
468 * - We are not currently being preempted
469 * - The target is not currently being preempted
470 * - We are able to satisfy the target's MP lock requirements (if any).
472 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
473 * this is called via lwkt_schedule() through the td_preemptable callback.
474 * critpri is the managed critical priority that we should ignore in order
475 * to determine whether preemption is possible (aka usually just the crit
476 * priority of lwkt_schedule() itself).
478 * XXX at the moment we run the target thread in a critical section during
479 * the preemption in order to prevent the target from taking interrupts
480 * that *WE* can't. Preemption is strictly limited to interrupt threads
481 * and interrupt-like threads, outside of a critical section, and the
482 * preempted source thread will be resumed the instant the target blocks
483 * whether or not the source is scheduled (i.e. preemption is supposed to
484 * be as transparent as possible).
486 * The target thread inherits our MP count (added to its own) for the
487 * duration of the preemption in order to preserve the atomicy of the
488 * MP lock during the preemption. Therefore, any preempting targets must be
489 * careful in regards to MP assertions. Note that the MP count may be
490 * out of sync with the physical mp_lock, but we do not have to preserve
491 * the original ownership of the lock if it was out of synch (that is, we
492 * can leave it synchronized on return).
495 lwkt_preempt(thread_t ntd, int critpri)
497 thread_t td = curthread;
504 * The caller has put us in a critical section. We can only preempt
505 * if the caller of the caller was not in a critical section (basically
506 * a local interrupt), as determined by the 'critpri' parameter. If
507 * we are unable to preempt
509 * YYY The target thread must be in a critical section (else it must
510 * inherit our critical section? I dunno yet).
512 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
515 if (!_lwkt_wantresched(ntd, td)) {
519 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
524 if (ntd->td_gd != mycpu) {
529 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
533 if (ntd->td_preempted) {
539 * note: an interrupt might have occured just as we were transitioning
540 * to or from the MP lock. In this case td_mpcount will be pre-disposed
541 * (non-zero) but not actually synchronized with the actual state of the
542 * lock. We can use it to imply an MP lock requirement for the
543 * preemption but we cannot use it to test whether we hold the MP lock
546 savecnt = td->td_mpcount;
547 mpheld = MP_LOCK_HELD();
548 ntd->td_mpcount += td->td_mpcount;
549 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
550 ntd->td_mpcount -= td->td_mpcount;
557 ntd->td_preempted = td;
558 td->td_flags |= TDF_PREEMPT_LOCK;
560 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
562 KKASSERT(savecnt == td->td_mpcount);
563 mpheld = MP_LOCK_HELD();
564 if (mpheld && td->td_mpcount == 0)
566 else if (mpheld == 0 && td->td_mpcount)
567 panic("lwkt_preempt(): MP lock was not held through");
569 ntd->td_preempted = NULL;
570 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
574 * Yield our thread while higher priority threads are pending. This is
575 * typically called when we leave a critical section but it can be safely
576 * called while we are in a critical section.
578 * This function will not generally yield to equal priority threads but it
579 * can occur as a side effect. Note that lwkt_switch() is called from
580 * inside the critical section to pervent its own crit_exit() from reentering
581 * lwkt_yield_quick().
583 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
584 * came along but was blocked and made pending.
586 * (self contained on a per cpu basis)
589 lwkt_yield_quick(void)
591 globaldata_t gd = mycpu;
592 thread_t td = gd->gd_curthread;
595 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
596 * it with a non-zero cpl then we might not wind up calling splz after
597 * a task switch when the critical section is exited even though the
598 * new task could accept the interrupt. YYY alternative is to have
599 * lwkt_switch() just call splz unconditionally.
601 * XXX from crit_exit() only called after last crit section is released.
602 * If called directly will run splz() even if in a critical section.
608 * YYY enabling will cause wakeup() to task-switch, which really
609 * confused the old 4.x code. This is a good way to simulate
610 * preemption and MP without actually doing preemption or MP, because a
611 * lot of code assumes that wakeup() does not block.
613 if (untimely_switch && gd->gd_intr_nesting_level == 0) {
616 * YYY temporary hacks until we disassociate the userland scheduler
617 * from the LWKT scheduler.
619 if (td->td_flags & TDF_RUNQ) {
620 lwkt_switch(); /* will not reenter yield function */
622 lwkt_schedule_self(); /* make sure we are scheduled */
623 lwkt_switch(); /* will not reenter yield function */
624 lwkt_deschedule_self(); /* make sure we are descheduled */
626 crit_exit_noyield(td);
631 * This implements a normal yield which, unlike _quick, will yield to equal
632 * priority threads as well. Note that gd_reqflags tests will be handled by
633 * the crit_exit() call in lwkt_switch().
635 * (self contained on a per cpu basis)
640 lwkt_schedule_self();
645 * Schedule a thread to run. As the current thread we can always safely
646 * schedule ourselves, and a shortcut procedure is provided for that
649 * (non-blocking, self contained on a per cpu basis)
652 lwkt_schedule_self(void)
654 thread_t td = curthread;
657 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
659 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
660 panic("SCHED SELF PANIC");
665 * Generic schedule. Possibly schedule threads belonging to other cpus and
666 * deal with threads that might be blocked on a wait queue.
668 * YYY this is one of the best places to implement load balancing code.
669 * Load balancing can be accomplished by requesting other sorts of actions
670 * for the thread in question.
673 lwkt_schedule(thread_t td)
676 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
677 && td->td_proc->p_stat == SSLEEP
679 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
681 curthread->td_proc ? curthread->td_proc->p_pid : -1,
682 curthread->td_proc ? curthread->td_proc->p_stat : -1,
684 td->td_proc ? curthread->td_proc->p_pid : -1,
685 td->td_proc ? curthread->td_proc->p_stat : -1
687 panic("SCHED PANIC");
691 if (td == curthread) {
697 * If the thread is on a wait list we have to send our scheduling
698 * request to the owner of the wait structure. Otherwise we send
699 * the scheduling request to the cpu owning the thread. Races
700 * are ok, the target will forward the message as necessary (the
701 * message may chase the thread around before it finally gets
704 * (remember, wait structures use stable storage)
706 if ((w = td->td_wait) != NULL) {
707 if (lwkt_trytoken(&w->wa_token)) {
708 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
711 if (td->td_gd == mycpu) {
713 if (td->td_preemptable) {
714 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
715 } else if (_lwkt_wantresched(td, curthread)) {
719 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
721 lwkt_reltoken(&w->wa_token);
723 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
727 * If the wait structure is NULL and we own the thread, there
728 * is no race (since we are in a critical section). If we
729 * do not own the thread there might be a race but the
730 * target cpu will deal with it.
732 if (td->td_gd == mycpu) {
734 if (td->td_preemptable) {
735 td->td_preemptable(td, TDPRI_CRIT);
736 } else if (_lwkt_wantresched(td, curthread)) {
740 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
748 * Managed acquisition. This code assumes that the MP lock is held for
749 * the tdallq operation and that the thread has been descheduled from its
750 * original cpu. We also have to wait for the thread to be entirely switched
751 * out on its original cpu (this is usually fast enough that we never loop)
752 * since the LWKT system does not have to hold the MP lock while switching
753 * and the target may have released it before switching.
756 lwkt_acquire(thread_t td)
758 struct globaldata *gd;
761 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
762 while (td->td_flags & TDF_RUNNING) /* XXX spin */
766 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
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_gd == mycpu) {
807 lwkt_send_ipiq(td->td_gd->gd_cpuid, (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_gd == mycpu);
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;
864 typedef struct lwkt_gettoken_req {
872 * This function deschedules the current thread and blocks on the specified
873 * wait queue. We obtain ownership of the wait queue in order to block
874 * on it. A generation number is used to interlock the wait queue in case
875 * it gets signalled while we are blocked waiting on the token.
877 * Note: alternatively we could dequeue our thread and then message the
878 * target cpu owning the wait queue. YYY implement as sysctl.
880 * Note: wait queue signals normally ping-pong the cpu as an optimization.
884 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
886 thread_t td = curthread;
888 lwkt_gettoken(&w->wa_token);
889 if (w->wa_gen == *gen) {
891 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
894 td->td_wmesg = wmesg;
897 lwkt_regettoken(&w->wa_token);
898 if (td->td_wmesg != NULL) {
903 /* token might be lost, doesn't matter for gen update */
905 lwkt_reltoken(&w->wa_token);
909 * Signal a wait queue. We gain ownership of the wait queue in order to
910 * signal it. Once a thread is removed from the wait queue we have to
911 * deal with the cpu owning the thread.
913 * Note: alternatively we could message the target cpu owning the wait
914 * queue. YYY implement as sysctl.
917 lwkt_signal(lwkt_wait_t w, int count)
922 lwkt_gettoken(&w->wa_token);
926 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
929 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
932 if (td->td_gd == mycpu) {
935 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
937 lwkt_regettoken(&w->wa_token);
939 lwkt_reltoken(&w->wa_token);
945 * Acquire ownership of a token
947 * Acquire ownership of a token. The token may have spl and/or critical
948 * section side effects, depending on its purpose. These side effects
949 * guarentee that you will maintain ownership of the token as long as you
950 * do not block. If you block you may lose access to the token (but you
951 * must still release it even if you lose your access to it).
953 * YYY for now we use a critical section to prevent IPIs from taking away
954 * a token, but do we really only need to disable IPIs ?
956 * YYY certain tokens could be made to act like mutexes when performance
957 * would be better (e.g. t_cpu == -1). This is not yet implemented.
959 * YYY the tokens replace 4.x's simplelocks for the most part, but this
960 * means that 4.x does not expect a switch so for now we cannot switch
961 * when waiting for an IPI to be returned.
963 * YYY If the token is owned by another cpu we may have to send an IPI to
964 * it and then block. The IPI causes the token to be given away to the
965 * requesting cpu, unless it has already changed hands. Since only the
966 * current cpu can give away a token it owns we do not need a memory barrier.
967 * This needs serious optimization.
974 lwkt_gettoken_remote(void *arg)
976 lwkt_gettoken_req *req = arg;
977 if (req->tok->t_cpu == mycpu->gd_cpuid) {
980 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
982 req->tok->t_cpu = req->cpu;
983 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
984 /* else set reqcpu to point to current cpu for release */
991 lwkt_gettoken(lwkt_token_t tok)
994 * Prevent preemption so the token can't be taken away from us once
995 * we gain ownership of it. Use a synchronous request which might
996 * block. The request will be forwarded as necessary playing catchup
1002 if (curthread->td_pri > 2000) {
1003 curthread->td_pri = 1000;
1008 while (tok->t_cpu != mycpu->gd_cpuid) {
1009 struct lwkt_gettoken_req req;
1013 req.cpu = mycpu->gd_cpuid;
1015 dcpu = (volatile int)tok->t_cpu;
1016 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1019 printf("REQT%d ", dcpu);
1021 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1022 lwkt_wait_ipiq(dcpu, seq);
1025 printf("REQR%d ", tok->t_cpu);
1030 * leave us in a critical section on return. This will be undone
1031 * by lwkt_reltoken(). Bump the generation number.
1033 return(++tok->t_gen);
1037 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1041 lwkt_trytoken(lwkt_token_t tok)
1045 if (tok->t_cpu != mycpu->gd_cpuid) {
1049 /* leave us in the critical section */
1055 * Release your ownership of a token. Releases must occur in reverse
1056 * order to aquisitions, eventually so priorities can be unwound properly
1057 * like SPLs. At the moment the actual implemention doesn't care.
1059 * We can safely hand a token that we own to another cpu without notifying
1060 * it, but once we do we can't get it back without requesting it (unless
1061 * the other cpu hands it back to us before we check).
1063 * We might have lost the token, so check that.
1066 lwkt_reltoken(lwkt_token_t tok)
1068 if (tok->t_cpu == mycpu->gd_cpuid) {
1069 tok->t_cpu = tok->t_reqcpu;
1075 * Reacquire a token that might have been lost and compare and update the
1076 * generation number. 0 is returned if the generation has not changed
1077 * (nobody else obtained the token while we were blocked, on this cpu or
1080 * This function returns with the token re-held whether the generation
1081 * number changed or not.
1084 lwkt_gentoken(lwkt_token_t tok, int *gen)
1086 if (lwkt_regettoken(tok) == *gen) {
1096 * Re-acquire a token that might have been lost. Returns the generation
1097 * number of the token.
1100 lwkt_regettoken(lwkt_token_t tok)
1102 /* assert we are in a critical section */
1103 if (tok->t_cpu != mycpu->gd_cpuid) {
1105 while (tok->t_cpu != mycpu->gd_cpuid) {
1106 struct lwkt_gettoken_req req;
1110 req.cpu = mycpu->gd_cpuid;
1112 dcpu = (volatile int)tok->t_cpu;
1113 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1116 printf("REQT%d ", dcpu);
1118 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1119 lwkt_wait_ipiq(dcpu, seq);
1122 printf("REQR%d ", tok->t_cpu);
1132 lwkt_inittoken(lwkt_token_t tok)
1135 * Zero structure and set cpu owner and reqcpu to cpu 0.
1137 bzero(tok, sizeof(*tok));
1141 * Create a kernel process/thread/whatever. It shares it's address space
1142 * with proc0 - ie: kernel only.
1144 * XXX should be renamed to lwkt_create()
1146 * The thread will be entered with the MP lock held.
1149 lwkt_create(void (*func)(void *), void *arg,
1150 struct thread **tdp, thread_t template, int tdflags,
1151 const char *fmt, ...)
1156 td = lwkt_alloc_thread(template);
1159 cpu_set_thread_handler(td, kthread_exit, func, arg);
1160 td->td_flags |= TDF_VERBOSE | tdflags;
1166 * Set up arg0 for 'ps' etc
1169 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1173 * Schedule the thread to run
1175 if ((td->td_flags & TDF_STOPREQ) == 0)
1178 td->td_flags &= ~TDF_STOPREQ;
1183 * Destroy an LWKT thread. Warning! This function is not called when
1184 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1185 * uses a different reaping mechanism.
1190 thread_t td = curthread;
1192 if (td->td_flags & TDF_VERBOSE)
1193 printf("kthread %p %s has exited\n", td, td->td_comm);
1195 lwkt_deschedule_self();
1196 ++mycpu->gd_tdfreecount;
1197 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1202 * Create a kernel process/thread/whatever. It shares it's address space
1203 * with proc0 - ie: kernel only. 5.x compatible.
1206 kthread_create(void (*func)(void *), void *arg,
1207 struct thread **tdp, const char *fmt, ...)
1212 td = lwkt_alloc_thread(NULL);
1215 cpu_set_thread_handler(td, kthread_exit, func, arg);
1216 td->td_flags |= TDF_VERBOSE;
1222 * Set up arg0 for 'ps' etc
1225 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1229 * Schedule the thread to run
1238 thread_t td = curthread;
1239 int lpri = td->td_pri;
1242 panic("td_pri is/would-go negative! %p %d", td, lpri);
1246 * Destroy an LWKT thread. Warning! This function is not called when
1247 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1248 * uses a different reaping mechanism.
1250 * XXX duplicates lwkt_exit()
1261 * Send a function execution request to another cpu. The request is queued
1262 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1263 * possible target cpu. The FIFO can be written.
1265 * YYY If the FIFO fills up we have to enable interrupts and process the
1266 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1267 * Create a CPU_*() function to do this!
1269 * Must be called from a critical section.
1272 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1276 struct globaldata *gd = mycpu;
1278 if (dcpu == gd->gd_cpuid) {
1283 ++gd->gd_intr_nesting_level;
1285 if (gd->gd_intr_nesting_level > 20)
1286 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1288 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1289 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1291 ip = &gd->gd_ipiq[dcpu];
1294 * We always drain before the FIFO becomes full so it should never
1295 * become full. We need to leave enough entries to deal with
1298 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1299 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1300 ip->ip_func[windex] = func;
1301 ip->ip_arg[windex] = arg;
1302 /* YYY memory barrier */
1304 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1305 unsigned int eflags = read_eflags();
1308 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1309 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1310 lwkt_process_ipiq();
1312 write_eflags(eflags);
1314 --gd->gd_intr_nesting_level;
1315 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1317 return(ip->ip_windex);
1321 * Send a message to several target cpus. Typically used for scheduling.
1324 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1330 lwkt_send_ipiq(cpuid, func, arg);
1331 mask &= ~(1 << cpuid);
1336 * Wait for the remote cpu to finish processing a function.
1338 * YYY we have to enable interrupts and process the IPIQ while waiting
1339 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1340 * function to do this! YYY we really should 'block' here.
1342 * Must be called from a critical section. Thsi routine may be called
1343 * from an interrupt (for example, if an interrupt wakes a foreign thread
1347 lwkt_wait_ipiq(int dcpu, int seq)
1350 int maxc = 100000000;
1352 if (dcpu != mycpu->gd_cpuid) {
1353 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1354 ip = &mycpu->gd_ipiq[dcpu];
1355 if ((int)(ip->ip_xindex - seq) < 0) {
1356 unsigned int eflags = read_eflags();
1358 while ((int)(ip->ip_xindex - seq) < 0) {
1359 lwkt_process_ipiq();
1361 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1362 if (maxc < -1000000)
1363 panic("LWKT_WAIT_IPIQ");
1365 write_eflags(eflags);
1371 * Called from IPI interrupt (like a fast interrupt), which has placed
1372 * us in a critical section. The MP lock may or may not be held.
1373 * May also be called from doreti or splz, or be reentrantly called
1374 * indirectly through the ip_func[] we run.
1377 lwkt_process_ipiq(void)
1380 int cpuid = mycpu->gd_cpuid;
1382 for (n = 0; n < ncpus; ++n) {
1388 ip = globaldata_find(n)->gd_ipiq;
1394 * Note: xindex is only updated after we are sure the function has
1395 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1396 * function may send an IPI which may block/drain.
1398 while (ip->ip_rindex != ip->ip_windex) {
1399 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1401 ip->ip_func[ri](ip->ip_arg[ri]);
1402 /* YYY memory barrier */
1403 ip->ip_xindex = ip->ip_rindex;
1411 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1413 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1414 return(0); /* NOT REACHED */
1418 lwkt_wait_ipiq(int dcpu, int seq)
1420 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);