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 1.18 2003/07/10 04:47:54 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_EXITED),
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 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_EXITED,
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");
317 * td_mpcount cannot be used to determine if we currently hold the
318 * MP lock because get_mplock() will increment it prior to attempting
319 * to get the lock, and switch out if it can't. Look at the actual lock.
321 mpheld = MP_LOCK_HELD();
323 if ((ntd = td->td_preempted) != NULL) {
325 * We had preempted another thread on this cpu, resume the preempted
326 * thread. This occurs transparently, whether the preempted thread
327 * was scheduled or not (it may have been preempted after descheduling
330 * We have to setup the MP lock for the original thread after backing
331 * out the adjustment that was made to curthread when the original
334 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
336 if (ntd->td_mpcount && mpheld == 0) {
337 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
338 td, ntd, td->td_mpcount, ntd->td_mpcount);
340 if (ntd->td_mpcount) {
341 td->td_mpcount -= ntd->td_mpcount;
342 KKASSERT(td->td_mpcount >= 0);
345 ntd->td_flags |= TDF_PREEMPT_DONE;
346 /* YYY release mp lock on switchback if original doesn't need it */
349 * Priority queue / round-robin at each priority. Note that user
350 * processes run at a fixed, low priority and the user process
351 * scheduler deals with interactions between user processes
352 * by scheduling and descheduling them from the LWKT queue as
355 * We have to adjust the MP lock for the target thread. If we
356 * need the MP lock and cannot obtain it we try to locate a
357 * thread that does not need the MP lock.
361 if (gd->gd_runqmask) {
362 int nq = bsrl(gd->gd_runqmask);
363 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
364 gd->gd_runqmask &= ~(1 << nq);
368 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
370 * Target needs MP lock and we couldn't get it, try
371 * to locate a thread which does not need the MP lock
374 u_int32_t rqmask = gd->gd_runqmask;
376 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
377 if (ntd->td_mpcount == 0)
382 rqmask &= ~(1 << nq);
386 ntd = &gd->gd_idlethread;
387 ntd->td_flags |= TDF_IDLE_NOHLT;
389 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
390 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
393 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
394 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
397 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
398 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
401 ntd = &gd->gd_idlethread;
404 KASSERT(ntd->td_pri >= TDPRI_CRIT,
405 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
408 * Passive release (used to transition from user to kernel mode
409 * when we block or switch rather then when we enter the kernel).
410 * This function is NOT called if we are switching into a preemption
411 * or returning from a preemption.
417 * Do the actual switch. If the new target does not need the MP lock
418 * and we are holding it, release the MP lock. If the new target requires
419 * the MP lock we have already acquired it for the target.
422 if (ntd->td_mpcount == 0 ) {
426 ASSERT_MP_LOCK_HELD();
437 * Request that the target thread preempt the current thread. Preemption
438 * only works under a specific set of conditions:
440 * - We are not preempting ourselves
441 * - The target thread is owned by the current cpu
442 * - We are not currently being preempted
443 * - The target is not currently being preempted
444 * - We are able to satisfy the target's MP lock requirements (if any).
446 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
447 * this is called via lwkt_schedule() through the td_preemptable callback.
448 * critpri is the managed critical priority that we should ignore in order
449 * to determine whether preemption is possible (aka usually just the crit
450 * priority of lwkt_schedule() itself).
452 * XXX at the moment we run the target thread in a critical section during
453 * the preemption in order to prevent the target from taking interrupts
454 * that *WE* can't. Preemption is strictly limited to interrupt threads
455 * and interrupt-like threads, outside of a critical section, and the
456 * preempted source thread will be resumed the instant the target blocks
457 * whether or not the source is scheduled (i.e. preemption is supposed to
458 * be as transparent as possible).
460 * The target thread inherits our MP count (added to its own) for the
461 * duration of the preemption in order to preserve the atomicy of the
462 * MP lock during the preemption. Therefore, any preempting targets must be
463 * careful in regards to MP assertions. Note that the MP count may be
464 * out of sync with the physical mp_lock. If we preempt we have to preserve
465 * the expected situation.
468 lwkt_preempt(thread_t ntd, int critpri)
470 thread_t td = curthread;
477 * The caller has put us in a critical section. We can only preempt
478 * if the caller of the caller was not in a critical section (basically
479 * a local interrupt), as determined by the 'critpri' parameter. If
480 * we are unable to preempt
482 * YYY The target thread must be in a critical section (else it must
483 * inherit our critical section? I dunno yet).
485 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
487 if (!_lwkt_wantresched(ntd, td)) {
491 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
497 if (ntd->td_cpu != mycpu->gd_cpuid) {
502 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
507 if (ntd->td_preempted) {
514 * note: an interrupt might have occured just as we were transitioning
515 * to the MP lock, with the lock held but our mpcount still 0. We have
516 * to be sure we restore the same condition when the preemption returns.
518 mpheld = MP_LOCK_HELD(); /* 0 or 1 */
519 if (mpheld && td->td_mpcount == 0)
520 panic("lwkt_preempt(): held and no count");
521 savecnt = td->td_mpcount;
522 td->td_mpcount += mpheld;
523 ntd->td_mpcount += td->td_mpcount;
524 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
525 td->td_mpcount -= mpheld;
526 ntd->td_mpcount -= td->td_mpcount;
534 ntd->td_preempted = td;
535 td->td_flags |= TDF_PREEMPT_LOCK;
537 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
539 td->td_mpcount -= mpheld;
540 KKASSERT(savecnt == td->td_mpcount);
541 if (mpheld == 0 && MP_LOCK_HELD())
543 else if (mpheld && !MP_LOCK_HELD())
544 panic("lwkt_preempt(): MP lock was not held through");
546 ntd->td_preempted = NULL;
547 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
551 * Yield our thread while higher priority threads are pending. This is
552 * typically called when we leave a critical section but it can be safely
553 * called while we are in a critical section.
555 * This function will not generally yield to equal priority threads but it
556 * can occur as a side effect. Note that lwkt_switch() is called from
557 * inside the critical section to pervent its own crit_exit() from reentering
558 * lwkt_yield_quick().
560 * gd_reqpri indicates that *something* changed, e.g. an interrupt or softint
561 * came along but was blocked and made pending.
563 * (self contained on a per cpu basis)
566 lwkt_yield_quick(void)
568 thread_t td = curthread;
571 * gd_reqpri is cleared in splz if the cpl is 0. If we were to clear
572 * it with a non-zero cpl then we might not wind up calling splz after
573 * a task switch when the critical section is exited even though the
574 * new task could accept the interrupt. YYY alternative is to have
575 * lwkt_switch() just call splz unconditionally.
577 * XXX from crit_exit() only called after last crit section is released.
578 * If called directly will run splz() even if in a critical section.
580 if ((td->td_pri & TDPRI_MASK) < mycpu->gd_reqpri) {
585 * YYY enabling will cause wakeup() to task-switch, which really
586 * confused the old 4.x code. This is a good way to simulate
587 * preemption and MP without actually doing preemption or MP, because a
588 * lot of code assumes that wakeup() does not block.
590 if (untimely_switch && mycpu->gd_intr_nesting_level == 0) {
593 * YYY temporary hacks until we disassociate the userland scheduler
594 * from the LWKT scheduler.
596 if (td->td_flags & TDF_RUNQ) {
597 lwkt_switch(); /* will not reenter yield function */
599 lwkt_schedule_self(); /* make sure we are scheduled */
600 lwkt_switch(); /* will not reenter yield function */
601 lwkt_deschedule_self(); /* make sure we are descheduled */
608 * This implements a normal yield which, unlike _quick, will yield to equal
609 * priority threads as well. Note that gd_reqpri tests will be handled by
610 * the crit_exit() call in lwkt_switch().
612 * (self contained on a per cpu basis)
617 lwkt_schedule_self();
622 * Schedule a thread to run. As the current thread we can always safely
623 * schedule ourselves, and a shortcut procedure is provided for that
626 * (non-blocking, self contained on a per cpu basis)
629 lwkt_schedule_self(void)
631 thread_t td = curthread;
634 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
636 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
637 panic("SCHED SELF PANIC");
642 * Generic schedule. Possibly schedule threads belonging to other cpus and
643 * deal with threads that might be blocked on a wait queue.
645 * YYY this is one of the best places to implement load balancing code.
646 * Load balancing can be accomplished by requesting other sorts of actions
647 * for the thread in question.
650 lwkt_schedule(thread_t td)
653 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
654 && td->td_proc->p_stat == SSLEEP
656 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
658 curthread->td_proc ? curthread->td_proc->p_pid : -1,
659 curthread->td_proc ? curthread->td_proc->p_stat : -1,
661 td->td_proc ? curthread->td_proc->p_pid : -1,
662 td->td_proc ? curthread->td_proc->p_stat : -1
664 panic("SCHED PANIC");
668 if (td == curthread) {
674 * If the thread is on a wait list we have to send our scheduling
675 * request to the owner of the wait structure. Otherwise we send
676 * the scheduling request to the cpu owning the thread. Races
677 * are ok, the target will forward the message as necessary (the
678 * message may chase the thread around before it finally gets
681 * (remember, wait structures use stable storage)
683 if ((w = td->td_wait) != NULL) {
684 if (lwkt_trytoken(&w->wa_token)) {
685 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
688 if (td->td_cpu == mycpu->gd_cpuid) {
690 if (td->td_preemptable) {
691 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
692 } else if (_lwkt_wantresched(td, curthread)) {
696 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
698 lwkt_reltoken(&w->wa_token);
700 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
704 * If the wait structure is NULL and we own the thread, there
705 * is no race (since we are in a critical section). If we
706 * do not own the thread there might be a race but the
707 * target cpu will deal with it.
709 if (td->td_cpu == mycpu->gd_cpuid) {
711 if (td->td_preemptable) {
712 td->td_preemptable(td, TDPRI_CRIT);
713 } else if (_lwkt_wantresched(td, curthread)) {
717 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
725 lwkt_acquire(thread_t td)
727 struct globaldata *gd;
730 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
733 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
736 td->td_cpu = gd->gd_cpuid;
737 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
743 * Deschedule a thread.
745 * (non-blocking, self contained on a per cpu basis)
748 lwkt_deschedule_self(void)
750 thread_t td = curthread;
753 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
759 * Generic deschedule. Descheduling threads other then your own should be
760 * done only in carefully controlled circumstances. Descheduling is
763 * This function may block if the cpu has run out of messages.
766 lwkt_deschedule(thread_t td)
769 if (td == curthread) {
772 if (td->td_cpu == mycpu->gd_cpuid) {
775 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_deschedule, td);
782 * Set the target thread's priority. This routine does not automatically
783 * switch to a higher priority thread, LWKT threads are not designed for
784 * continuous priority changes. Yield if you want to switch.
786 * We have to retain the critical section count which uses the high bits
787 * of the td_pri field. The specified priority may also indicate zero or
788 * more critical sections by adding TDPRI_CRIT*N.
791 lwkt_setpri(thread_t td, int pri)
794 KKASSERT(td->td_cpu == mycpu->gd_cpuid);
796 if (td->td_flags & TDF_RUNQ) {
798 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
801 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
807 lwkt_setpri_self(int pri)
809 thread_t td = curthread;
811 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
813 if (td->td_flags & TDF_RUNQ) {
815 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
818 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
824 lwkt_preempted_proc(void)
826 thread_t td = curthread;
827 while (td->td_preempted)
828 td = td->td_preempted;
834 * This function deschedules the current thread and blocks on the specified
835 * wait queue. We obtain ownership of the wait queue in order to block
836 * on it. A generation number is used to interlock the wait queue in case
837 * it gets signalled while we are blocked waiting on the token.
839 * Note: alternatively we could dequeue our thread and then message the
840 * target cpu owning the wait queue. YYY implement as sysctl.
842 * Note: wait queue signals normally ping-pong the cpu as an optimization.
844 typedef struct lwkt_gettoken_req {
850 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
852 thread_t td = curthread;
854 lwkt_gettoken(&w->wa_token);
855 if (w->wa_gen == *gen) {
857 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
860 td->td_wmesg = wmesg;
863 /* token might be lost, doesn't matter for gen update */
865 lwkt_reltoken(&w->wa_token);
869 * Signal a wait queue. We gain ownership of the wait queue in order to
870 * signal it. Once a thread is removed from the wait queue we have to
871 * deal with the cpu owning the thread.
873 * Note: alternatively we could message the target cpu owning the wait
874 * queue. YYY implement as sysctl.
877 lwkt_signal(lwkt_wait_t w)
882 lwkt_gettoken(&w->wa_token);
885 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
888 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
891 if (td->td_cpu == mycpu->gd_cpuid) {
894 lwkt_send_ipiq(td->td_cpu, (ipifunc_t)lwkt_schedule, td);
896 lwkt_regettoken(&w->wa_token);
898 lwkt_reltoken(&w->wa_token);
902 * Acquire ownership of a token
904 * Acquire ownership of a token. The token may have spl and/or critical
905 * section side effects, depending on its purpose. These side effects
906 * guarentee that you will maintain ownership of the token as long as you
907 * do not block. If you block you may lose access to the token (but you
908 * must still release it even if you lose your access to it).
910 * YYY for now we use a critical section to prevent IPIs from taking away
911 * a token, but do we really only need to disable IPIs ?
913 * YYY certain tokens could be made to act like mutexes when performance
914 * would be better (e.g. t_cpu == -1). This is not yet implemented.
916 * YYY the tokens replace 4.x's simplelocks for the most part, but this
917 * means that 4.x does not expect a switch so for now we cannot switch
918 * when waiting for an IPI to be returned.
920 * YYY If the token is owned by another cpu we may have to send an IPI to
921 * it and then block. The IPI causes the token to be given away to the
922 * requesting cpu, unless it has already changed hands. Since only the
923 * current cpu can give away a token it owns we do not need a memory barrier.
924 * This needs serious optimization.
931 lwkt_gettoken_remote(void *arg)
933 lwkt_gettoken_req *req = arg;
934 if (req->tok->t_cpu == mycpu->gd_cpuid) {
936 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
937 req->tok->t_cpu = req->cpu;
938 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
939 /* else set reqcpu to point to current cpu for release */
946 lwkt_gettoken(lwkt_token_t tok)
949 * Prevent preemption so the token can't be taken away from us once
950 * we gain ownership of it. Use a synchronous request which might
951 * block. The request will be forwarded as necessary playing catchup
957 if (curthread->td_pri > 2000) {
958 curthread->td_pri = 1000;
963 while (tok->t_cpu != mycpu->gd_cpuid) {
964 struct lwkt_gettoken_req req;
968 req.cpu = mycpu->gd_cpuid;
970 dcpu = (volatile int)tok->t_cpu;
971 KKASSERT(dcpu >= 0 && dcpu < ncpus);
973 printf("REQT%d ", dcpu);
974 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
975 lwkt_wait_ipiq(dcpu, seq);
977 printf("REQR%d ", tok->t_cpu);
981 * leave us in a critical section on return. This will be undone
982 * by lwkt_reltoken(). Bump the generation number.
984 return(++tok->t_gen);
988 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
992 lwkt_trytoken(lwkt_token_t tok)
996 if (tok->t_cpu != mycpu->gd_cpuid) {
1000 /* leave us in the critical section */
1006 * Release your ownership of a token. Releases must occur in reverse
1007 * order to aquisitions, eventually so priorities can be unwound properly
1008 * like SPLs. At the moment the actual implemention doesn't care.
1010 * We can safely hand a token that we own to another cpu without notifying
1011 * it, but once we do we can't get it back without requesting it (unless
1012 * the other cpu hands it back to us before we check).
1014 * We might have lost the token, so check that.
1017 lwkt_reltoken(lwkt_token_t tok)
1019 if (tok->t_cpu == mycpu->gd_cpuid) {
1020 tok->t_cpu = tok->t_reqcpu;
1026 * Reacquire a token that might have been lost and compare and update the
1027 * generation number. 0 is returned if the generation has not changed
1028 * (nobody else obtained the token while we were blocked, on this cpu or
1031 * This function returns with the token re-held whether the generation
1032 * number changed or not.
1035 lwkt_gentoken(lwkt_token_t tok, int *gen)
1037 if (lwkt_regettoken(tok) == *gen) {
1047 * Re-acquire a token that might have been lost. Returns the generation
1048 * number of the token.
1051 lwkt_regettoken(lwkt_token_t tok)
1053 /* assert we are in a critical section */
1054 if (tok->t_cpu != mycpu->gd_cpuid) {
1056 while (tok->t_cpu != mycpu->gd_cpuid) {
1057 struct lwkt_gettoken_req req;
1061 req.cpu = mycpu->gd_cpuid;
1063 dcpu = (volatile int)tok->t_cpu;
1064 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1065 printf("REQT%d ", dcpu);
1066 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1067 lwkt_wait_ipiq(dcpu, seq);
1068 printf("REQR%d ", tok->t_cpu);
1077 lwkt_inittoken(lwkt_token_t tok)
1080 * Zero structure and set cpu owner and reqcpu to cpu 0.
1082 bzero(tok, sizeof(*tok));
1086 * Create a kernel process/thread/whatever. It shares it's address space
1087 * with proc0 - ie: kernel only.
1089 * XXX should be renamed to lwkt_create()
1091 * The thread will be entered with the MP lock held.
1094 lwkt_create(void (*func)(void *), void *arg,
1095 struct thread **tdp, thread_t template, int tdflags,
1096 const char *fmt, ...)
1101 td = lwkt_alloc_thread(template);
1104 cpu_set_thread_handler(td, kthread_exit, func, arg);
1105 td->td_flags |= TDF_VERBOSE | tdflags;
1111 * Set up arg0 for 'ps' etc
1114 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1118 * Schedule the thread to run
1120 if ((td->td_flags & TDF_STOPREQ) == 0)
1123 td->td_flags &= ~TDF_STOPREQ;
1128 * Destroy an LWKT thread. Warning! This function is not called when
1129 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1130 * uses a different reaping mechanism.
1135 thread_t td = curthread;
1137 if (td->td_flags & TDF_VERBOSE)
1138 printf("kthread %p %s has exited\n", td, td->td_comm);
1140 lwkt_deschedule_self();
1141 ++mycpu->gd_tdfreecount;
1142 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1147 * Create a kernel process/thread/whatever. It shares it's address space
1148 * with proc0 - ie: kernel only. 5.x compatible.
1151 kthread_create(void (*func)(void *), void *arg,
1152 struct thread **tdp, const char *fmt, ...)
1157 td = lwkt_alloc_thread(NULL);
1160 cpu_set_thread_handler(td, kthread_exit, func, arg);
1161 td->td_flags |= TDF_VERBOSE;
1167 * Set up arg0 for 'ps' etc
1170 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1174 * Schedule the thread to run
1183 thread_t td = curthread;
1184 int lpri = td->td_pri;
1187 panic("td_pri is/would-go negative! %p %d", td, lpri);
1191 * Destroy an LWKT thread. Warning! This function is not called when
1192 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1193 * uses a different reaping mechanism.
1195 * XXX duplicates lwkt_exit()
1206 * Send a function execution request to another cpu. The request is queued
1207 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1208 * possible target cpu. The FIFO can be written.
1210 * YYY If the FIFO fills up we have to enable interrupts and process the
1211 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1212 * Create a CPU_*() function to do this!
1214 * Must be called from a critical section.
1217 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1221 struct globaldata *gd = mycpu;
1223 if (dcpu == gd->gd_cpuid) {
1227 ++gd->gd_intr_nesting_level;
1229 if (gd->gd_intr_nesting_level > 20)
1230 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1232 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1233 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1235 ip = &gd->gd_ipiq[dcpu];
1236 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1237 unsigned int eflags = read_eflags();
1238 printf("SEND_IPIQ FIFO FULL\n");
1241 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1242 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1243 lwkt_process_ipiq();
1245 printf("SEND_IPIQ FIFO GOOD\n");
1246 write_eflags(eflags);
1248 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1249 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1250 ip->ip_func[windex] = func;
1251 ip->ip_arg[windex] = arg;
1252 /* YYY memory barrier */
1254 --gd->gd_intr_nesting_level;
1255 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1256 return(ip->ip_windex);
1260 * Wait for the remote cpu to finish processing a function.
1262 * YYY we have to enable interrupts and process the IPIQ while waiting
1263 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1264 * function to do this! YYY we really should 'block' here.
1266 * Must be called from a critical section. Thsi routine may be called
1267 * from an interrupt (for example, if an interrupt wakes a foreign thread
1271 lwkt_wait_ipiq(int dcpu, int seq)
1274 int maxc = 100000000;
1276 if (dcpu != mycpu->gd_cpuid) {
1277 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1278 ip = &mycpu->gd_ipiq[dcpu];
1279 if ((int)(ip->ip_rindex - seq) < 0) {
1280 unsigned int eflags = read_eflags();
1282 while ((int)(ip->ip_rindex - seq) < 0) {
1283 lwkt_process_ipiq();
1285 lwkt_switch(); /* YYY fixme */
1288 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_rindex - seq);
1289 if (maxc < -1000000)
1290 panic("LWKT_WAIT_IPIQ");
1292 write_eflags(eflags);
1298 * Called from IPI interrupt (like a fast interrupt), which has placed
1299 * us in a critical section. The MP lock may or may not be held.
1300 * May also be called from doreti or splz.
1303 lwkt_process_ipiq(void)
1306 int cpuid = mycpu->gd_cpuid;
1308 for (n = 0; n < ncpus; ++n) {
1314 ip = globaldata_find(n)->gd_ipiq;
1318 while (ip->ip_rindex != ip->ip_windex) {
1319 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1320 ip->ip_func[ri](ip->ip_arg[ri]);
1329 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1331 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1332 return(0); /* NOT REACHED */
1336 lwkt_wait_ipiq(int dcpu, int seq)
1338 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);