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 1.37 2003/10/17 07:30:42 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;
306 * Switching from within a 'fast' (non thread switched) interrupt is
309 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
310 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
314 * Passive release (used to transition from user to kernel mode
315 * when we block or switch rather then when we enter the kernel).
316 * This function is NOT called if we are switching into a preemption
317 * or returning from a preemption. Typically this causes us to lose
318 * our P_CURPROC designation (if we have one) and become a true LWKT
319 * thread, and may also hand P_CURPROC to another process and schedule
330 * td_mpcount cannot be used to determine if we currently hold the
331 * MP lock because get_mplock() will increment it prior to attempting
332 * to get the lock, and switch out if it can't. Our ownership of
333 * the actual lock will remain stable while we are in a critical section
334 * (but, of course, another cpu may own or release the lock so the
335 * actual value of mp_lock is not stable).
337 mpheld = MP_LOCK_HELD();
339 if ((ntd = td->td_preempted) != NULL) {
341 * We had preempted another thread on this cpu, resume the preempted
342 * thread. This occurs transparently, whether the preempted thread
343 * was scheduled or not (it may have been preempted after descheduling
346 * We have to setup the MP lock for the original thread after backing
347 * out the adjustment that was made to curthread when the original
350 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
352 if (ntd->td_mpcount && mpheld == 0) {
353 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
354 td, ntd, td->td_mpcount, ntd->td_mpcount);
356 if (ntd->td_mpcount) {
357 td->td_mpcount -= ntd->td_mpcount;
358 KKASSERT(td->td_mpcount >= 0);
361 ntd->td_flags |= TDF_PREEMPT_DONE;
362 /* YYY release mp lock on switchback if original doesn't need it */
365 * Priority queue / round-robin at each priority. Note that user
366 * processes run at a fixed, low priority and the user process
367 * scheduler deals with interactions between user processes
368 * by scheduling and descheduling them from the LWKT queue as
371 * We have to adjust the MP lock for the target thread. If we
372 * need the MP lock and cannot obtain it we try to locate a
373 * thread that does not need the MP lock.
377 if (gd->gd_runqmask) {
378 int nq = bsrl(gd->gd_runqmask);
379 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
380 gd->gd_runqmask &= ~(1 << nq);
384 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
386 * Target needs MP lock and we couldn't get it, try
387 * to locate a thread which does not need the MP lock
388 * to run. If we cannot locate a thread spin in idle.
390 u_int32_t rqmask = gd->gd_runqmask;
392 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
393 if (ntd->td_mpcount == 0)
398 rqmask &= ~(1 << nq);
402 ntd = &gd->gd_idlethread;
403 ntd->td_flags |= TDF_IDLE_NOHLT;
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);
413 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
414 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
418 * Nothing to run but we may still need the BGL to deal with
419 * pending interrupts, spin in idle if so.
421 ntd = &gd->gd_idlethread;
423 ntd->td_flags |= TDF_IDLE_NOHLT;
426 KASSERT(ntd->td_pri >= TDPRI_CRIT,
427 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
430 * Do the actual switch. If the new target does not need the MP lock
431 * and we are holding it, release the MP lock. If the new target requires
432 * the MP lock we have already acquired it for the target.
435 if (ntd->td_mpcount == 0 ) {
439 ASSERT_MP_LOCK_HELD();
450 * Switch if another thread has a higher priority. Do not switch to other
451 * threads at the same priority.
456 struct globaldata *gd = mycpu;
457 struct thread *td = gd->gd_curthread;
459 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
465 * Request that the target thread preempt the current thread. Preemption
466 * only works under a specific set of conditions:
468 * - We are not preempting ourselves
469 * - The target thread is owned by the current cpu
470 * - We are not currently being preempted
471 * - The target is not currently being preempted
472 * - We are able to satisfy the target's MP lock requirements (if any).
474 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
475 * this is called via lwkt_schedule() through the td_preemptable callback.
476 * critpri is the managed critical priority that we should ignore in order
477 * to determine whether preemption is possible (aka usually just the crit
478 * priority of lwkt_schedule() itself).
480 * XXX at the moment we run the target thread in a critical section during
481 * the preemption in order to prevent the target from taking interrupts
482 * that *WE* can't. Preemption is strictly limited to interrupt threads
483 * and interrupt-like threads, outside of a critical section, and the
484 * preempted source thread will be resumed the instant the target blocks
485 * whether or not the source is scheduled (i.e. preemption is supposed to
486 * be as transparent as possible).
488 * The target thread inherits our MP count (added to its own) for the
489 * duration of the preemption in order to preserve the atomicy of the
490 * MP lock during the preemption. Therefore, any preempting targets must be
491 * careful in regards to MP assertions. Note that the MP count may be
492 * out of sync with the physical mp_lock, but we do not have to preserve
493 * the original ownership of the lock if it was out of synch (that is, we
494 * can leave it synchronized on return).
497 lwkt_preempt(thread_t ntd, int critpri)
499 struct globaldata *gd = mycpu;
500 thread_t td = gd->gd_curthread;
507 * The caller has put us in a critical section. We can only preempt
508 * if the caller of the caller was not in a critical section (basically
509 * a local interrupt), as determined by the 'critpri' parameter. If
510 * we are unable to preempt
512 * YYY The target thread must be in a critical section (else it must
513 * inherit our critical section? I dunno yet).
515 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
518 if (!_lwkt_wantresched(ntd, td)) {
522 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
527 if (ntd->td_gd != gd) {
532 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
536 if (ntd->td_preempted) {
542 * note: an interrupt might have occured just as we were transitioning
543 * to or from the MP lock. In this case td_mpcount will be pre-disposed
544 * (non-zero) but not actually synchronized with the actual state of the
545 * lock. We can use it to imply an MP lock requirement for the
546 * preemption but we cannot use it to test whether we hold the MP lock
549 savecnt = td->td_mpcount;
550 mpheld = MP_LOCK_HELD();
551 ntd->td_mpcount += td->td_mpcount;
552 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
553 ntd->td_mpcount -= td->td_mpcount;
560 ntd->td_preempted = td;
561 td->td_flags |= TDF_PREEMPT_LOCK;
563 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
565 KKASSERT(savecnt == td->td_mpcount);
566 mpheld = MP_LOCK_HELD();
567 if (mpheld && td->td_mpcount == 0)
569 else if (mpheld == 0 && td->td_mpcount)
570 panic("lwkt_preempt(): MP lock was not held through");
572 ntd->td_preempted = NULL;
573 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
577 * Yield our thread while higher priority threads are pending. This is
578 * typically called when we leave a critical section but it can be safely
579 * called while we are in a critical section.
581 * This function will not generally yield to equal priority threads but it
582 * can occur as a side effect. Note that lwkt_switch() is called from
583 * inside the critical section to prevent its own crit_exit() from reentering
584 * lwkt_yield_quick().
586 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
587 * came along but was blocked and made pending.
589 * (self contained on a per cpu basis)
592 lwkt_yield_quick(void)
594 globaldata_t gd = mycpu;
595 thread_t td = gd->gd_curthread;
598 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
599 * it with a non-zero cpl then we might not wind up calling splz after
600 * a task switch when the critical section is exited even though the
601 * new task could accept the interrupt.
603 * XXX from crit_exit() only called after last crit section is released.
604 * If called directly will run splz() even if in a critical section.
606 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
607 * except for this special case, we MUST call splz() here to handle any
608 * pending ints, particularly after we switch, or we might accidently
609 * halt the cpu with interrupts pending.
611 if (gd->gd_reqflags && td->td_nest_count < 2)
615 * YYY enabling will cause wakeup() to task-switch, which really
616 * confused the old 4.x code. This is a good way to simulate
617 * preemption and MP without actually doing preemption or MP, because a
618 * lot of code assumes that wakeup() does not block.
620 if (untimely_switch && td->td_nest_count == 0 &&
621 gd->gd_intr_nesting_level == 0
625 * YYY temporary hacks until we disassociate the userland scheduler
626 * from the LWKT scheduler.
628 if (td->td_flags & TDF_RUNQ) {
629 lwkt_switch(); /* will not reenter yield function */
631 lwkt_schedule_self(); /* make sure we are scheduled */
632 lwkt_switch(); /* will not reenter yield function */
633 lwkt_deschedule_self(); /* make sure we are descheduled */
635 crit_exit_noyield(td);
640 * This implements a normal yield which, unlike _quick, will yield to equal
641 * priority threads as well. Note that gd_reqflags tests will be handled by
642 * the crit_exit() call in lwkt_switch().
644 * (self contained on a per cpu basis)
649 lwkt_schedule_self();
654 * Schedule a thread to run. As the current thread we can always safely
655 * schedule ourselves, and a shortcut procedure is provided for that
658 * (non-blocking, self contained on a per cpu basis)
661 lwkt_schedule_self(void)
663 thread_t td = curthread;
666 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
668 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
669 panic("SCHED SELF PANIC");
674 * Generic schedule. Possibly schedule threads belonging to other cpus and
675 * deal with threads that might be blocked on a wait queue.
677 * YYY this is one of the best places to implement load balancing code.
678 * Load balancing can be accomplished by requesting other sorts of actions
679 * for the thread in question.
682 lwkt_schedule(thread_t td)
685 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
686 && td->td_proc->p_stat == SSLEEP
688 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
690 curthread->td_proc ? curthread->td_proc->p_pid : -1,
691 curthread->td_proc ? curthread->td_proc->p_stat : -1,
693 td->td_proc ? curthread->td_proc->p_pid : -1,
694 td->td_proc ? curthread->td_proc->p_stat : -1
696 panic("SCHED PANIC");
700 if (td == curthread) {
706 * If the thread is on a wait list we have to send our scheduling
707 * request to the owner of the wait structure. Otherwise we send
708 * the scheduling request to the cpu owning the thread. Races
709 * are ok, the target will forward the message as necessary (the
710 * message may chase the thread around before it finally gets
713 * (remember, wait structures use stable storage)
715 if ((w = td->td_wait) != NULL) {
716 if (lwkt_trytoken(&w->wa_token)) {
717 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
720 if (td->td_gd == mycpu) {
722 if (td->td_preemptable) {
723 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
724 } else if (_lwkt_wantresched(td, curthread)) {
728 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
730 lwkt_reltoken(&w->wa_token);
732 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
736 * If the wait structure is NULL and we own the thread, there
737 * is no race (since we are in a critical section). If we
738 * do not own the thread there might be a race but the
739 * target cpu will deal with it.
741 if (td->td_gd == mycpu) {
743 if (td->td_preemptable) {
744 td->td_preemptable(td, TDPRI_CRIT);
745 } else if (_lwkt_wantresched(td, curthread)) {
749 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
757 * Managed acquisition. This code assumes that the MP lock is held for
758 * the tdallq operation and that the thread has been descheduled from its
759 * original cpu. We also have to wait for the thread to be entirely switched
760 * out on its original cpu (this is usually fast enough that we never loop)
761 * since the LWKT system does not have to hold the MP lock while switching
762 * and the target may have released it before switching.
765 lwkt_acquire(thread_t td)
767 struct globaldata *gd;
770 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
771 while (td->td_flags & TDF_RUNNING) /* XXX spin */
775 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
778 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
784 * Deschedule a thread.
786 * (non-blocking, self contained on a per cpu basis)
789 lwkt_deschedule_self(void)
791 thread_t td = curthread;
794 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
800 * Generic deschedule. Descheduling threads other then your own should be
801 * done only in carefully controlled circumstances. Descheduling is
804 * This function may block if the cpu has run out of messages.
807 lwkt_deschedule(thread_t td)
810 if (td == curthread) {
813 if (td->td_gd == mycpu) {
816 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
823 * Set the target thread's priority. This routine does not automatically
824 * switch to a higher priority thread, LWKT threads are not designed for
825 * continuous priority changes. Yield if you want to switch.
827 * We have to retain the critical section count which uses the high bits
828 * of the td_pri field. The specified priority may also indicate zero or
829 * more critical sections by adding TDPRI_CRIT*N.
832 lwkt_setpri(thread_t td, int pri)
835 KKASSERT(td->td_gd == mycpu);
837 if (td->td_flags & TDF_RUNQ) {
839 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
842 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
848 lwkt_setpri_self(int pri)
850 thread_t td = curthread;
852 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
854 if (td->td_flags & TDF_RUNQ) {
856 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
859 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
865 lwkt_preempted_proc(void)
867 thread_t td = curthread;
868 while (td->td_preempted)
869 td = td->td_preempted;
873 typedef struct lwkt_gettoken_req {
881 * This function deschedules the current thread and blocks on the specified
882 * wait queue. We obtain ownership of the wait queue in order to block
883 * on it. A generation number is used to interlock the wait queue in case
884 * it gets signalled while we are blocked waiting on the token.
886 * Note: alternatively we could dequeue our thread and then message the
887 * target cpu owning the wait queue. YYY implement as sysctl.
889 * Note: wait queue signals normally ping-pong the cpu as an optimization.
893 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
895 thread_t td = curthread;
897 lwkt_gettoken(&w->wa_token);
898 if (w->wa_gen == *gen) {
900 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
903 td->td_wmesg = wmesg;
906 lwkt_regettoken(&w->wa_token);
907 if (td->td_wmesg != NULL) {
912 /* token might be lost, doesn't matter for gen update */
914 lwkt_reltoken(&w->wa_token);
918 * Signal a wait queue. We gain ownership of the wait queue in order to
919 * signal it. Once a thread is removed from the wait queue we have to
920 * deal with the cpu owning the thread.
922 * Note: alternatively we could message the target cpu owning the wait
923 * queue. YYY implement as sysctl.
926 lwkt_signal(lwkt_wait_t w, int count)
931 lwkt_gettoken(&w->wa_token);
935 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
938 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
941 if (td->td_gd == mycpu) {
944 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
946 lwkt_regettoken(&w->wa_token);
948 lwkt_reltoken(&w->wa_token);
954 * Acquire ownership of a token
956 * Acquire ownership of a token. The token may have spl and/or critical
957 * section side effects, depending on its purpose. These side effects
958 * guarentee that you will maintain ownership of the token as long as you
959 * do not block. If you block you may lose access to the token (but you
960 * must still release it even if you lose your access to it).
962 * YYY for now we use a critical section to prevent IPIs from taking away
963 * a token, but do we really only need to disable IPIs ?
965 * YYY certain tokens could be made to act like mutexes when performance
966 * would be better (e.g. t_cpu == -1). This is not yet implemented.
968 * YYY the tokens replace 4.x's simplelocks for the most part, but this
969 * means that 4.x does not expect a switch so for now we cannot switch
970 * when waiting for an IPI to be returned.
972 * YYY If the token is owned by another cpu we may have to send an IPI to
973 * it and then block. The IPI causes the token to be given away to the
974 * requesting cpu, unless it has already changed hands. Since only the
975 * current cpu can give away a token it owns we do not need a memory barrier.
976 * This needs serious optimization.
983 lwkt_gettoken_remote(void *arg)
985 lwkt_gettoken_req *req = arg;
986 if (req->tok->t_cpu == mycpu->gd_cpuid) {
989 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
991 req->tok->t_cpu = req->cpu;
992 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
993 /* else set reqcpu to point to current cpu for release */
1000 lwkt_gettoken(lwkt_token_t tok)
1003 * Prevent preemption so the token can't be taken away from us once
1004 * we gain ownership of it. Use a synchronous request which might
1005 * block. The request will be forwarded as necessary playing catchup
1011 if (curthread->td_pri > 1800) {
1012 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1013 tok, ((int **)&tok)[-1]);
1015 if (curthread->td_pri > 2000) {
1016 curthread->td_pri = 1000;
1021 while (tok->t_cpu != mycpu->gd_cpuid) {
1022 struct lwkt_gettoken_req req;
1026 req.cpu = mycpu->gd_cpuid;
1028 dcpu = (volatile int)tok->t_cpu;
1029 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1032 printf("REQT%d ", dcpu);
1034 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1035 lwkt_wait_ipiq(dcpu, seq);
1038 printf("REQR%d ", tok->t_cpu);
1043 * leave us in a critical section on return. This will be undone
1044 * by lwkt_reltoken(). Bump the generation number.
1046 return(++tok->t_gen);
1050 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1054 lwkt_trytoken(lwkt_token_t tok)
1058 if (tok->t_cpu != mycpu->gd_cpuid) {
1063 /* leave us in the critical section */
1069 * Release your ownership of a token. Releases must occur in reverse
1070 * order to aquisitions, eventually so priorities can be unwound properly
1071 * like SPLs. At the moment the actual implemention doesn't care.
1073 * We can safely hand a token that we own to another cpu without notifying
1074 * it, but once we do we can't get it back without requesting it (unless
1075 * the other cpu hands it back to us before we check).
1077 * We might have lost the token, so check that.
1079 * Return the token's generation number. The number is useful to callers
1080 * who may want to know if the token was stolen during potential blockages.
1083 lwkt_reltoken(lwkt_token_t tok)
1087 if (tok->t_cpu == mycpu->gd_cpuid) {
1088 tok->t_cpu = tok->t_reqcpu;
1096 * Reacquire a token that might have been lost. 0 is returned if the
1097 * generation has not changed (nobody stole the token from us), -1 is
1098 * returned otherwise. The token is reacquired regardless but the
1099 * generation number is not bumped further if we already own the token.
1101 * For efficiency we inline the best-case situation for lwkt_regettoken()
1102 * (i.e .we still own the token).
1105 lwkt_gentoken(lwkt_token_t tok, int *gen)
1107 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
1109 *gen = lwkt_regettoken(tok);
1114 * Re-acquire a token that might have been lost. The generation number
1115 * is bumped and returned regardless of whether the token had been lost
1116 * or not (because we only have cpu granularity we have to bump the token
1120 lwkt_regettoken(lwkt_token_t tok)
1122 /* assert we are in a critical section */
1123 if (tok->t_cpu != mycpu->gd_cpuid) {
1125 while (tok->t_cpu != mycpu->gd_cpuid) {
1126 struct lwkt_gettoken_req req;
1130 req.cpu = mycpu->gd_cpuid;
1132 dcpu = (volatile int)tok->t_cpu;
1133 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1136 printf("REQT%d ", dcpu);
1138 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1139 lwkt_wait_ipiq(dcpu, seq);
1142 printf("REQR%d ", tok->t_cpu);
1152 lwkt_inittoken(lwkt_token_t tok)
1155 * Zero structure and set cpu owner and reqcpu to cpu 0.
1157 bzero(tok, sizeof(*tok));
1161 * Create a kernel process/thread/whatever. It shares it's address space
1162 * with proc0 - ie: kernel only.
1164 * XXX should be renamed to lwkt_create()
1166 * The thread will be entered with the MP lock held.
1169 lwkt_create(void (*func)(void *), void *arg,
1170 struct thread **tdp, thread_t template, int tdflags,
1171 const char *fmt, ...)
1176 td = lwkt_alloc_thread(template);
1179 cpu_set_thread_handler(td, kthread_exit, func, arg);
1180 td->td_flags |= TDF_VERBOSE | tdflags;
1186 * Set up arg0 for 'ps' etc
1189 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1193 * Schedule the thread to run
1195 if ((td->td_flags & TDF_STOPREQ) == 0)
1198 td->td_flags &= ~TDF_STOPREQ;
1203 * Destroy an LWKT thread. Warning! This function is not called when
1204 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1205 * uses a different reaping mechanism.
1210 thread_t td = curthread;
1212 if (td->td_flags & TDF_VERBOSE)
1213 printf("kthread %p %s has exited\n", td, td->td_comm);
1215 lwkt_deschedule_self();
1216 ++mycpu->gd_tdfreecount;
1217 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1222 * Create a kernel process/thread/whatever. It shares it's address space
1223 * with proc0 - ie: kernel only. 5.x compatible.
1226 kthread_create(void (*func)(void *), void *arg,
1227 struct thread **tdp, const char *fmt, ...)
1232 td = lwkt_alloc_thread(NULL);
1235 cpu_set_thread_handler(td, kthread_exit, func, arg);
1236 td->td_flags |= TDF_VERBOSE;
1242 * Set up arg0 for 'ps' etc
1245 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1249 * Schedule the thread to run
1258 thread_t td = curthread;
1259 int lpri = td->td_pri;
1262 panic("td_pri is/would-go negative! %p %d", td, lpri);
1266 * Destroy an LWKT thread. Warning! This function is not called when
1267 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1268 * uses a different reaping mechanism.
1270 * XXX duplicates lwkt_exit()
1281 * Send a function execution request to another cpu. The request is queued
1282 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1283 * possible target cpu. The FIFO can be written.
1285 * YYY If the FIFO fills up we have to enable interrupts and process the
1286 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1287 * Create a CPU_*() function to do this!
1289 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1290 * end will take care of any pending interrupts.
1292 * Must be called from a critical section.
1295 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1299 struct globaldata *gd = mycpu;
1301 if (dcpu == gd->gd_cpuid) {
1306 ++gd->gd_intr_nesting_level;
1308 if (gd->gd_intr_nesting_level > 20)
1309 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1311 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1312 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1314 ip = &gd->gd_ipiq[dcpu];
1317 * We always drain before the FIFO becomes full so it should never
1318 * become full. We need to leave enough entries to deal with
1321 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1322 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1323 ip->ip_func[windex] = func;
1324 ip->ip_arg[windex] = arg;
1325 /* YYY memory barrier */
1327 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1328 unsigned int eflags = read_eflags();
1331 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1332 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1333 lwkt_process_ipiq();
1335 write_eflags(eflags);
1337 --gd->gd_intr_nesting_level;
1338 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1340 return(ip->ip_windex);
1344 * Send a message to several target cpus. Typically used for scheduling.
1345 * The message will not be sent to stopped cpus.
1348 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1352 mask &= ~stopped_cpus;
1355 lwkt_send_ipiq(cpuid, func, arg);
1356 mask &= ~(1 << cpuid);
1361 * Wait for the remote cpu to finish processing a function.
1363 * YYY we have to enable interrupts and process the IPIQ while waiting
1364 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1365 * function to do this! YYY we really should 'block' here.
1367 * Must be called from a critical section. Thsi routine may be called
1368 * from an interrupt (for example, if an interrupt wakes a foreign thread
1372 lwkt_wait_ipiq(int dcpu, int seq)
1375 int maxc = 100000000;
1377 if (dcpu != mycpu->gd_cpuid) {
1378 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1379 ip = &mycpu->gd_ipiq[dcpu];
1380 if ((int)(ip->ip_xindex - seq) < 0) {
1381 unsigned int eflags = read_eflags();
1383 while ((int)(ip->ip_xindex - seq) < 0) {
1384 lwkt_process_ipiq();
1386 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1387 if (maxc < -1000000)
1388 panic("LWKT_WAIT_IPIQ");
1390 write_eflags(eflags);
1396 * Called from IPI interrupt (like a fast interrupt), which has placed
1397 * us in a critical section. The MP lock may or may not be held.
1398 * May also be called from doreti or splz, or be reentrantly called
1399 * indirectly through the ip_func[] we run.
1402 lwkt_process_ipiq(void)
1405 int cpuid = mycpu->gd_cpuid;
1407 for (n = 0; n < ncpus; ++n) {
1413 ip = globaldata_find(n)->gd_ipiq;
1419 * Note: xindex is only updated after we are sure the function has
1420 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1421 * function may send an IPI which may block/drain.
1423 while (ip->ip_rindex != ip->ip_windex) {
1424 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1426 ip->ip_func[ri](ip->ip_arg[ri]);
1427 /* YYY memory barrier */
1428 ip->ip_xindex = ip->ip_rindex;
1436 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1438 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1439 return(0); /* NOT REACHED */
1443 lwkt_wait_ipiq(int dcpu, int seq)
1445 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);