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 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.45 2003/12/04 00:12:40 dillon Exp $
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
47 #include <sys/rtprio.h>
48 #include <sys/queue.h>
49 #include <sys/thread2.h>
50 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
56 #include <vm/vm_param.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_pager.h>
62 #include <vm/vm_extern.h>
63 #include <vm/vm_zone.h>
65 #include <machine/stdarg.h>
66 #include <machine/ipl.h>
67 #include <machine/smp.h>
69 #define THREAD_STACK (UPAGES * PAGE_SIZE)
73 #include <sys/stdint.h>
74 #include <libcaps/thread.h>
75 #include <sys/thread.h>
76 #include <sys/msgport.h>
77 #include <sys/errno.h>
78 #include <libcaps/globaldata.h>
79 #include <sys/thread2.h>
80 #include <sys/msgport2.h>
82 #include <machine/cpufunc.h>
86 static int untimely_switch = 0;
88 static int token_debug = 0;
90 static __int64_t switch_count = 0;
91 static __int64_t preempt_hit = 0;
92 static __int64_t preempt_miss = 0;
93 static __int64_t preempt_weird = 0;
94 static __int64_t ipiq_count = 0;
95 static __int64_t ipiq_fifofull = 0;
99 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
101 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
103 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
104 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
105 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
106 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
107 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
113 * These helper procedures handle the runq, they can only be called from
114 * within a critical section.
116 * WARNING! Prior to SMP being brought up it is possible to enqueue and
117 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
118 * instead of 'mycpu' when referencing the globaldata structure. Once
119 * SMP live enqueuing and dequeueing only occurs on the current cpu.
123 _lwkt_dequeue(thread_t td)
125 if (td->td_flags & TDF_RUNQ) {
126 int nq = td->td_pri & TDPRI_MASK;
127 struct globaldata *gd = td->td_gd;
129 td->td_flags &= ~TDF_RUNQ;
130 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
131 /* runqmask is passively cleaned up by the switcher */
137 _lwkt_enqueue(thread_t td)
139 if ((td->td_flags & TDF_RUNQ) == 0) {
140 int nq = td->td_pri & TDPRI_MASK;
141 struct globaldata *gd = td->td_gd;
143 td->td_flags |= TDF_RUNQ;
144 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
145 gd->gd_runqmask |= 1 << nq;
151 _lwkt_wantresched(thread_t ntd, thread_t cur)
153 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
156 /* lwkt_gdinit() has a userland override */
160 * LWKTs operate on a per-cpu basis
162 * WARNING! Called from early boot, 'mycpu' may not work yet.
165 lwkt_gdinit(struct globaldata *gd)
169 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
170 TAILQ_INIT(&gd->gd_tdrunq[i]);
172 TAILQ_INIT(&gd->gd_tdallq);
178 * Initialize a thread wait structure prior to first use.
180 * NOTE! called from low level boot code, we cannot do anything fancy!
183 lwkt_init_wait(lwkt_wait_t w)
185 TAILQ_INIT(&w->wa_waitq);
189 * Create a new thread. The thread must be associated with a process context
190 * or LWKT start address before it can be scheduled. If the target cpu is
191 * -1 the thread will be created on the current cpu.
193 * If you intend to create a thread without a process context this function
194 * does everything except load the startup and switcher function.
197 lwkt_alloc_thread(struct thread *td, int cpu)
204 if (mycpu->gd_tdfreecount > 0) {
205 --mycpu->gd_tdfreecount;
206 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
207 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
208 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
209 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
211 stack = td->td_kstack;
212 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
216 td = zalloc(thread_zone);
218 td = malloc(sizeof(struct thread));
220 td->td_kstack = NULL;
221 flags |= TDF_ALLOCATED_THREAD;
224 if ((stack = td->td_kstack) == NULL) {
226 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
228 stack = libcaps_alloc_stack(THREAD_STACK);
230 flags |= TDF_ALLOCATED_STACK;
233 lwkt_init_thread(td, stack, flags, mycpu);
235 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
240 * Initialize a preexisting thread structure. This function is used by
241 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
243 * All threads start out in a critical section at a priority of
244 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
245 * appropriate. This function may send an IPI message when the
246 * requested cpu is not the current cpu and consequently gd_tdallq may
247 * not be initialized synchronously from the point of view of the originating
250 * NOTE! we have to be careful in regards to creating threads for other cpus
251 * if SMP has not yet been activated.
254 lwkt_init_thread_remote(void *arg)
258 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
261 /* lwkt_init_thread has a userland override */
265 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
267 bzero(td, sizeof(struct thread));
268 td->td_kstack = stack;
269 td->td_flags |= flags;
271 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
272 lwkt_initport(&td->td_msgport, td);
273 pmap_init_thread(td);
274 if (smp_active == 0 || gd == mycpu) {
276 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
279 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
286 lwkt_set_comm(thread_t td, const char *ctl, ...)
291 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
296 lwkt_hold(thread_t td)
302 lwkt_rele(thread_t td)
304 KKASSERT(td->td_refs > 0);
311 lwkt_wait_free(thread_t td)
314 tsleep(td, 0, "tdreap", hz);
320 lwkt_free_thread(thread_t td)
322 struct globaldata *gd = mycpu;
324 KASSERT((td->td_flags & TDF_RUNNING) == 0,
325 ("lwkt_free_thread: did not exit! %p", td));
328 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
329 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
330 (td->td_flags & TDF_ALLOCATED_THREAD)
332 ++gd->gd_tdfreecount;
333 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
337 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
339 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
341 libcaps_free_stack(td->td_kstack, THREAD_STACK);
344 td->td_kstack = NULL;
346 if (td->td_flags & TDF_ALLOCATED_THREAD) {
348 zfree(thread_zone, td);
358 * Switch to the next runnable lwkt. If no LWKTs are runnable then
359 * switch to the idlethread. Switching must occur within a critical
360 * section to avoid races with the scheduling queue.
362 * We always have full control over our cpu's run queue. Other cpus
363 * that wish to manipulate our queue must use the cpu_*msg() calls to
364 * talk to our cpu, so a critical section is all that is needed and
365 * the result is very, very fast thread switching.
367 * The LWKT scheduler uses a fixed priority model and round-robins at
368 * each priority level. User process scheduling is a totally
369 * different beast and LWKT priorities should not be confused with
370 * user process priorities.
372 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
373 * cleans it up. Note that the td_switch() function cannot do anything that
374 * requires the MP lock since the MP lock will have already been setup for
375 * the target thread (not the current thread). It's nice to have a scheduler
376 * that does not need the MP lock to work because it allows us to do some
377 * really cool high-performance MP lock optimizations.
383 struct globaldata *gd;
384 thread_t td = curthread;
391 * Switching from within a 'fast' (non thread switched) interrupt is
394 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
395 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
399 * Passive release (used to transition from user to kernel mode
400 * when we block or switch rather then when we enter the kernel).
401 * This function is NOT called if we are switching into a preemption
402 * or returning from a preemption. Typically this causes us to lose
403 * our P_CURPROC designation (if we have one) and become a true LWKT
404 * thread, and may also hand P_CURPROC to another process and schedule
415 * td_mpcount cannot be used to determine if we currently hold the
416 * MP lock because get_mplock() will increment it prior to attempting
417 * to get the lock, and switch out if it can't. Our ownership of
418 * the actual lock will remain stable while we are in a critical section
419 * (but, of course, another cpu may own or release the lock so the
420 * actual value of mp_lock is not stable).
422 mpheld = MP_LOCK_HELD();
424 if ((ntd = td->td_preempted) != NULL) {
426 * We had preempted another thread on this cpu, resume the preempted
427 * thread. This occurs transparently, whether the preempted thread
428 * was scheduled or not (it may have been preempted after descheduling
431 * We have to setup the MP lock for the original thread after backing
432 * out the adjustment that was made to curthread when the original
435 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
437 if (ntd->td_mpcount && mpheld == 0) {
438 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
439 td, ntd, td->td_mpcount, ntd->td_mpcount);
441 if (ntd->td_mpcount) {
442 td->td_mpcount -= ntd->td_mpcount;
443 KKASSERT(td->td_mpcount >= 0);
446 ntd->td_flags |= TDF_PREEMPT_DONE;
447 /* YYY release mp lock on switchback if original doesn't need it */
450 * Priority queue / round-robin at each priority. Note that user
451 * processes run at a fixed, low priority and the user process
452 * scheduler deals with interactions between user processes
453 * by scheduling and descheduling them from the LWKT queue as
456 * We have to adjust the MP lock for the target thread. If we
457 * need the MP lock and cannot obtain it we try to locate a
458 * thread that does not need the MP lock.
462 if (gd->gd_runqmask) {
463 int nq = bsrl(gd->gd_runqmask);
464 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
465 gd->gd_runqmask &= ~(1 << nq);
469 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
471 * Target needs MP lock and we couldn't get it, try
472 * to locate a thread which does not need the MP lock
473 * to run. If we cannot locate a thread spin in idle.
475 u_int32_t rqmask = gd->gd_runqmask;
477 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
478 if (ntd->td_mpcount == 0)
483 rqmask &= ~(1 << nq);
487 ntd = &gd->gd_idlethread;
488 ntd->td_flags |= TDF_IDLE_NOHLT;
490 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
491 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
494 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
495 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
498 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
499 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
503 * Nothing to run but we may still need the BGL to deal with
504 * pending interrupts, spin in idle if so.
506 ntd = &gd->gd_idlethread;
508 ntd->td_flags |= TDF_IDLE_NOHLT;
511 KASSERT(ntd->td_pri >= TDPRI_CRIT,
512 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
515 * Do the actual switch. If the new target does not need the MP lock
516 * and we are holding it, release the MP lock. If the new target requires
517 * the MP lock we have already acquired it for the target.
520 if (ntd->td_mpcount == 0 ) {
524 ASSERT_MP_LOCK_HELD();
535 * Switch if another thread has a higher priority. Do not switch to other
536 * threads at the same priority.
541 struct globaldata *gd = mycpu;
542 struct thread *td = gd->gd_curthread;
544 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
550 * Request that the target thread preempt the current thread. Preemption
551 * only works under a specific set of conditions:
553 * - We are not preempting ourselves
554 * - The target thread is owned by the current cpu
555 * - We are not currently being preempted
556 * - The target is not currently being preempted
557 * - We are able to satisfy the target's MP lock requirements (if any).
559 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
560 * this is called via lwkt_schedule() through the td_preemptable callback.
561 * critpri is the managed critical priority that we should ignore in order
562 * to determine whether preemption is possible (aka usually just the crit
563 * priority of lwkt_schedule() itself).
565 * XXX at the moment we run the target thread in a critical section during
566 * the preemption in order to prevent the target from taking interrupts
567 * that *WE* can't. Preemption is strictly limited to interrupt threads
568 * and interrupt-like threads, outside of a critical section, and the
569 * preempted source thread will be resumed the instant the target blocks
570 * whether or not the source is scheduled (i.e. preemption is supposed to
571 * be as transparent as possible).
573 * The target thread inherits our MP count (added to its own) for the
574 * duration of the preemption in order to preserve the atomicy of the
575 * MP lock during the preemption. Therefore, any preempting targets must be
576 * careful in regards to MP assertions. Note that the MP count may be
577 * out of sync with the physical mp_lock, but we do not have to preserve
578 * the original ownership of the lock if it was out of synch (that is, we
579 * can leave it synchronized on return).
582 lwkt_preempt(thread_t ntd, int critpri)
584 struct globaldata *gd = mycpu;
585 thread_t td = gd->gd_curthread;
592 * The caller has put us in a critical section. We can only preempt
593 * if the caller of the caller was not in a critical section (basically
594 * a local interrupt), as determined by the 'critpri' parameter. If
595 * we are unable to preempt
597 * YYY The target thread must be in a critical section (else it must
598 * inherit our critical section? I dunno yet).
600 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
603 if (!_lwkt_wantresched(ntd, td)) {
607 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
612 if (ntd->td_gd != gd) {
617 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
621 if (ntd->td_preempted) {
627 * note: an interrupt might have occured just as we were transitioning
628 * to or from the MP lock. In this case td_mpcount will be pre-disposed
629 * (non-zero) but not actually synchronized with the actual state of the
630 * lock. We can use it to imply an MP lock requirement for the
631 * preemption but we cannot use it to test whether we hold the MP lock
634 savecnt = td->td_mpcount;
635 mpheld = MP_LOCK_HELD();
636 ntd->td_mpcount += td->td_mpcount;
637 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
638 ntd->td_mpcount -= td->td_mpcount;
645 ntd->td_preempted = td;
646 td->td_flags |= TDF_PREEMPT_LOCK;
648 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
650 KKASSERT(savecnt == td->td_mpcount);
651 mpheld = MP_LOCK_HELD();
652 if (mpheld && td->td_mpcount == 0)
654 else if (mpheld == 0 && td->td_mpcount)
655 panic("lwkt_preempt(): MP lock was not held through");
657 ntd->td_preempted = NULL;
658 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
662 * Yield our thread while higher priority threads are pending. This is
663 * typically called when we leave a critical section but it can be safely
664 * called while we are in a critical section.
666 * This function will not generally yield to equal priority threads but it
667 * can occur as a side effect. Note that lwkt_switch() is called from
668 * inside the critical section to prevent its own crit_exit() from reentering
669 * lwkt_yield_quick().
671 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
672 * came along but was blocked and made pending.
674 * (self contained on a per cpu basis)
677 lwkt_yield_quick(void)
679 globaldata_t gd = mycpu;
680 thread_t td = gd->gd_curthread;
683 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
684 * it with a non-zero cpl then we might not wind up calling splz after
685 * a task switch when the critical section is exited even though the
686 * new task could accept the interrupt.
688 * XXX from crit_exit() only called after last crit section is released.
689 * If called directly will run splz() even if in a critical section.
691 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
692 * except for this special case, we MUST call splz() here to handle any
693 * pending ints, particularly after we switch, or we might accidently
694 * halt the cpu with interrupts pending.
696 if (gd->gd_reqflags && td->td_nest_count < 2)
700 * YYY enabling will cause wakeup() to task-switch, which really
701 * confused the old 4.x code. This is a good way to simulate
702 * preemption and MP without actually doing preemption or MP, because a
703 * lot of code assumes that wakeup() does not block.
705 if (untimely_switch && td->td_nest_count == 0 &&
706 gd->gd_intr_nesting_level == 0
710 * YYY temporary hacks until we disassociate the userland scheduler
711 * from the LWKT scheduler.
713 if (td->td_flags & TDF_RUNQ) {
714 lwkt_switch(); /* will not reenter yield function */
716 lwkt_schedule_self(); /* make sure we are scheduled */
717 lwkt_switch(); /* will not reenter yield function */
718 lwkt_deschedule_self(); /* make sure we are descheduled */
720 crit_exit_noyield(td);
725 * This implements a normal yield which, unlike _quick, will yield to equal
726 * priority threads as well. Note that gd_reqflags tests will be handled by
727 * the crit_exit() call in lwkt_switch().
729 * (self contained on a per cpu basis)
734 lwkt_schedule_self();
739 * Schedule a thread to run. As the current thread we can always safely
740 * schedule ourselves, and a shortcut procedure is provided for that
743 * (non-blocking, self contained on a per cpu basis)
746 lwkt_schedule_self(void)
748 thread_t td = curthread;
751 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
754 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
755 panic("SCHED SELF PANIC");
761 * Generic schedule. Possibly schedule threads belonging to other cpus and
762 * deal with threads that might be blocked on a wait queue.
764 * YYY this is one of the best places to implement load balancing code.
765 * Load balancing can be accomplished by requesting other sorts of actions
766 * for the thread in question.
769 lwkt_schedule(thread_t td)
772 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
773 && td->td_proc->p_stat == SSLEEP
775 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
777 curthread->td_proc ? curthread->td_proc->p_pid : -1,
778 curthread->td_proc ? curthread->td_proc->p_stat : -1,
780 td->td_proc ? curthread->td_proc->p_pid : -1,
781 td->td_proc ? curthread->td_proc->p_stat : -1
783 panic("SCHED PANIC");
787 if (td == curthread) {
793 * If the thread is on a wait list we have to send our scheduling
794 * request to the owner of the wait structure. Otherwise we send
795 * the scheduling request to the cpu owning the thread. Races
796 * are ok, the target will forward the message as necessary (the
797 * message may chase the thread around before it finally gets
800 * (remember, wait structures use stable storage)
802 if ((w = td->td_wait) != NULL) {
803 if (lwkt_trytoken(&w->wa_token)) {
804 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
807 if (smp_active == 0 || td->td_gd == mycpu) {
809 if (td->td_preemptable) {
810 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
811 } else if (_lwkt_wantresched(td, curthread)) {
815 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
817 lwkt_reltoken(&w->wa_token);
819 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
823 * If the wait structure is NULL and we own the thread, there
824 * is no race (since we are in a critical section). If we
825 * do not own the thread there might be a race but the
826 * target cpu will deal with it.
828 if (smp_active == 0 || td->td_gd == mycpu) {
830 if (td->td_preemptable) {
831 td->td_preemptable(td, TDPRI_CRIT);
832 } else if (_lwkt_wantresched(td, curthread)) {
836 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
844 * Managed acquisition. This code assumes that the MP lock is held for
845 * the tdallq operation and that the thread has been descheduled from its
846 * original cpu. We also have to wait for the thread to be entirely switched
847 * out on its original cpu (this is usually fast enough that we never loop)
848 * since the LWKT system does not have to hold the MP lock while switching
849 * and the target may have released it before switching.
852 lwkt_acquire(thread_t td)
854 struct globaldata *gd;
857 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
858 while (td->td_flags & TDF_RUNNING) /* XXX spin */
862 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
865 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
871 * Deschedule a thread.
873 * (non-blocking, self contained on a per cpu basis)
876 lwkt_deschedule_self(void)
878 thread_t td = curthread;
881 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
887 * Generic deschedule. Descheduling threads other then your own should be
888 * done only in carefully controlled circumstances. Descheduling is
891 * This function may block if the cpu has run out of messages.
894 lwkt_deschedule(thread_t td)
897 if (td == curthread) {
900 if (td->td_gd == mycpu) {
903 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
910 * Set the target thread's priority. This routine does not automatically
911 * switch to a higher priority thread, LWKT threads are not designed for
912 * continuous priority changes. Yield if you want to switch.
914 * We have to retain the critical section count which uses the high bits
915 * of the td_pri field. The specified priority may also indicate zero or
916 * more critical sections by adding TDPRI_CRIT*N.
919 lwkt_setpri(thread_t td, int pri)
922 KKASSERT(td->td_gd == mycpu);
924 if (td->td_flags & TDF_RUNQ) {
926 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
929 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
935 lwkt_setpri_self(int pri)
937 thread_t td = curthread;
939 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
941 if (td->td_flags & TDF_RUNQ) {
943 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
946 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
952 lwkt_preempted_proc(void)
954 thread_t td = curthread;
955 while (td->td_preempted)
956 td = td->td_preempted;
960 typedef struct lwkt_gettoken_req {
968 * This function deschedules the current thread and blocks on the specified
969 * wait queue. We obtain ownership of the wait queue in order to block
970 * on it. A generation number is used to interlock the wait queue in case
971 * it gets signalled while we are blocked waiting on the token.
973 * Note: alternatively we could dequeue our thread and then message the
974 * target cpu owning the wait queue. YYY implement as sysctl.
976 * Note: wait queue signals normally ping-pong the cpu as an optimization.
980 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
982 thread_t td = curthread;
984 lwkt_gettoken(&w->wa_token);
985 if (w->wa_gen == *gen) {
987 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
990 td->td_wmesg = wmesg;
993 lwkt_regettoken(&w->wa_token);
994 if (td->td_wmesg != NULL) {
999 /* token might be lost, doesn't matter for gen update */
1001 lwkt_reltoken(&w->wa_token);
1005 * Signal a wait queue. We gain ownership of the wait queue in order to
1006 * signal it. Once a thread is removed from the wait queue we have to
1007 * deal with the cpu owning the thread.
1009 * Note: alternatively we could message the target cpu owning the wait
1010 * queue. YYY implement as sysctl.
1013 lwkt_signal(lwkt_wait_t w, int count)
1018 lwkt_gettoken(&w->wa_token);
1021 count = w->wa_count;
1022 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1025 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1027 td->td_wmesg = NULL;
1028 if (td->td_gd == mycpu) {
1031 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
1033 lwkt_regettoken(&w->wa_token);
1035 lwkt_reltoken(&w->wa_token);
1041 * Acquire ownership of a token
1043 * Acquire ownership of a token. The token may have spl and/or critical
1044 * section side effects, depending on its purpose. These side effects
1045 * guarentee that you will maintain ownership of the token as long as you
1046 * do not block. If you block you may lose access to the token (but you
1047 * must still release it even if you lose your access to it).
1049 * YYY for now we use a critical section to prevent IPIs from taking away
1050 * a token, but do we really only need to disable IPIs ?
1052 * YYY certain tokens could be made to act like mutexes when performance
1053 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1055 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1056 * means that 4.x does not expect a switch so for now we cannot switch
1057 * when waiting for an IPI to be returned.
1059 * YYY If the token is owned by another cpu we may have to send an IPI to
1060 * it and then block. The IPI causes the token to be given away to the
1061 * requesting cpu, unless it has already changed hands. Since only the
1062 * current cpu can give away a token it owns we do not need a memory barrier.
1063 * This needs serious optimization.
1070 lwkt_gettoken_remote(void *arg)
1072 lwkt_gettoken_req *req = arg;
1073 if (req->tok->t_cpu == mycpu->gd_cpuid) {
1076 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
1078 req->tok->t_cpu = req->cpu;
1079 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1080 /* else set reqcpu to point to current cpu for release */
1087 lwkt_gettoken(lwkt_token_t tok)
1090 * Prevent preemption so the token can't be taken away from us once
1091 * we gain ownership of it. Use a synchronous request which might
1092 * block. The request will be forwarded as necessary playing catchup
1098 if (curthread->td_pri > 1800) {
1099 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1100 tok, ((int **)&tok)[-1]);
1102 if (curthread->td_pri > 2000) {
1103 curthread->td_pri = 1000;
1108 while (tok->t_cpu != mycpu->gd_cpuid) {
1109 struct lwkt_gettoken_req req;
1113 req.cpu = mycpu->gd_cpuid;
1115 dcpu = (volatile int)tok->t_cpu;
1116 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1119 printf("REQT%d ", dcpu);
1121 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1122 lwkt_wait_ipiq(dcpu, seq);
1125 printf("REQR%d ", tok->t_cpu);
1130 * leave us in a critical section on return. This will be undone
1131 * by lwkt_reltoken(). Bump the generation number.
1133 return(++tok->t_gen);
1137 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1141 lwkt_trytoken(lwkt_token_t tok)
1145 if (tok->t_cpu != mycpu->gd_cpuid) {
1150 /* leave us in the critical section */
1156 * Release your ownership of a token. Releases must occur in reverse
1157 * order to aquisitions, eventually so priorities can be unwound properly
1158 * like SPLs. At the moment the actual implemention doesn't care.
1160 * We can safely hand a token that we own to another cpu without notifying
1161 * it, but once we do we can't get it back without requesting it (unless
1162 * the other cpu hands it back to us before we check).
1164 * We might have lost the token, so check that.
1166 * Return the token's generation number. The number is useful to callers
1167 * who may want to know if the token was stolen during potential blockages.
1170 lwkt_reltoken(lwkt_token_t tok)
1174 if (tok->t_cpu == mycpu->gd_cpuid) {
1175 tok->t_cpu = tok->t_reqcpu;
1183 * Reacquire a token that might have been lost. 0 is returned if the
1184 * generation has not changed (nobody stole the token from us), -1 is
1185 * returned otherwise. The token is reacquired regardless but the
1186 * generation number is not bumped further if we already own the token.
1188 * For efficiency we inline the best-case situation for lwkt_regettoken()
1189 * (i.e .we still own the token).
1192 lwkt_gentoken(lwkt_token_t tok, int *gen)
1194 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
1196 *gen = lwkt_regettoken(tok);
1201 * Re-acquire a token that might have been lost. The generation number
1202 * is bumped and returned regardless of whether the token had been lost
1203 * or not (because we only have cpu granularity we have to bump the token
1207 lwkt_regettoken(lwkt_token_t tok)
1209 /* assert we are in a critical section */
1210 if (tok->t_cpu != mycpu->gd_cpuid) {
1212 while (tok->t_cpu != mycpu->gd_cpuid) {
1213 struct lwkt_gettoken_req req;
1217 req.cpu = mycpu->gd_cpuid;
1219 dcpu = (volatile int)tok->t_cpu;
1220 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1223 printf("REQT%d ", dcpu);
1225 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1226 lwkt_wait_ipiq(dcpu, seq);
1229 printf("REQR%d ", tok->t_cpu);
1239 lwkt_inittoken(lwkt_token_t tok)
1242 * Zero structure and set cpu owner and reqcpu to cpu 0.
1244 bzero(tok, sizeof(*tok));
1248 * Create a kernel process/thread/whatever. It shares it's address space
1249 * with proc0 - ie: kernel only.
1251 * NOTE! By default new threads are created with the MP lock held. A
1252 * thread which does not require the MP lock should release it by calling
1253 * rel_mplock() at the start of the new thread.
1256 lwkt_create(void (*func)(void *), void *arg,
1257 struct thread **tdp, thread_t template, int tdflags, int cpu,
1258 const char *fmt, ...)
1263 td = lwkt_alloc_thread(template, cpu);
1266 cpu_set_thread_handler(td, kthread_exit, func, arg);
1267 td->td_flags |= TDF_VERBOSE | tdflags;
1273 * Set up arg0 for 'ps' etc
1275 __va_start(ap, fmt);
1276 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1280 * Schedule the thread to run
1282 if ((td->td_flags & TDF_STOPREQ) == 0)
1285 td->td_flags &= ~TDF_STOPREQ;
1290 * lwkt_exit() has a userland override.
1291 * kthread_* is specific to the kernel and is not needed by userland.
1296 * Destroy an LWKT thread. Warning! This function is not called when
1297 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1298 * uses a different reaping mechanism.
1303 thread_t td = curthread;
1305 if (td->td_flags & TDF_VERBOSE)
1306 printf("kthread %p %s has exited\n", td, td->td_comm);
1308 lwkt_deschedule_self();
1309 ++mycpu->gd_tdfreecount;
1310 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1315 * Create a kernel process/thread/whatever. It shares it's address space
1316 * with proc0 - ie: kernel only. 5.x compatible.
1318 * NOTE! By default kthreads are created with the MP lock held. A
1319 * thread which does not require the MP lock should release it by calling
1320 * rel_mplock() at the start of the new thread.
1323 kthread_create(void (*func)(void *), void *arg,
1324 struct thread **tdp, const char *fmt, ...)
1329 td = lwkt_alloc_thread(NULL, -1);
1332 cpu_set_thread_handler(td, kthread_exit, func, arg);
1333 td->td_flags |= TDF_VERBOSE;
1339 * Set up arg0 for 'ps' etc
1341 __va_start(ap, fmt);
1342 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1346 * Schedule the thread to run
1353 * Destroy an LWKT thread. Warning! This function is not called when
1354 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1355 * uses a different reaping mechanism.
1357 * XXX duplicates lwkt_exit()
1365 #endif /* _KERNEL */
1370 thread_t td = curthread;
1371 int lpri = td->td_pri;
1374 panic("td_pri is/would-go negative! %p %d", td, lpri);
1380 * Send a function execution request to another cpu. The request is queued
1381 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1382 * possible target cpu. The FIFO can be written.
1384 * YYY If the FIFO fills up we have to enable interrupts and process the
1385 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1386 * Create a CPU_*() function to do this!
1388 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1389 * end will take care of any pending interrupts.
1391 * Must be called from a critical section.
1394 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1398 struct globaldata *gd = mycpu;
1400 if (dcpu == gd->gd_cpuid) {
1405 ++gd->gd_intr_nesting_level;
1407 if (gd->gd_intr_nesting_level > 20)
1408 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1410 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1411 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1413 ip = &gd->gd_ipiq[dcpu];
1416 * We always drain before the FIFO becomes full so it should never
1417 * become full. We need to leave enough entries to deal with
1420 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1421 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1422 ip->ip_func[windex] = func;
1423 ip->ip_arg[windex] = arg;
1424 /* YYY memory barrier */
1426 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1427 unsigned int eflags = read_eflags();
1430 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1431 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1432 lwkt_process_ipiq();
1434 write_eflags(eflags);
1436 --gd->gd_intr_nesting_level;
1437 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1439 return(ip->ip_windex);
1443 * Send a message to several target cpus. Typically used for scheduling.
1444 * The message will not be sent to stopped cpus.
1447 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1451 mask &= ~stopped_cpus;
1454 lwkt_send_ipiq(cpuid, func, arg);
1455 mask &= ~(1 << cpuid);
1460 * Wait for the remote cpu to finish processing a function.
1462 * YYY we have to enable interrupts and process the IPIQ while waiting
1463 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1464 * function to do this! YYY we really should 'block' here.
1466 * Must be called from a critical section. Thsi routine may be called
1467 * from an interrupt (for example, if an interrupt wakes a foreign thread
1471 lwkt_wait_ipiq(int dcpu, int seq)
1474 int maxc = 100000000;
1476 if (dcpu != mycpu->gd_cpuid) {
1477 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1478 ip = &mycpu->gd_ipiq[dcpu];
1479 if ((int)(ip->ip_xindex - seq) < 0) {
1480 unsigned int eflags = read_eflags();
1482 while ((int)(ip->ip_xindex - seq) < 0) {
1483 lwkt_process_ipiq();
1485 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1486 if (maxc < -1000000)
1487 panic("LWKT_WAIT_IPIQ");
1489 write_eflags(eflags);
1495 * Called from IPI interrupt (like a fast interrupt), which has placed
1496 * us in a critical section. The MP lock may or may not be held.
1497 * May also be called from doreti or splz, or be reentrantly called
1498 * indirectly through the ip_func[] we run.
1501 lwkt_process_ipiq(void)
1504 int cpuid = mycpu->gd_cpuid;
1506 for (n = 0; n < ncpus; ++n) {
1512 ip = globaldata_find(n)->gd_ipiq;
1518 * Note: xindex is only updated after we are sure the function has
1519 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1520 * function may send an IPI which may block/drain.
1522 while (ip->ip_rindex != ip->ip_windex) {
1523 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1525 ip->ip_func[ri](ip->ip_arg[ri]);
1526 /* YYY memory barrier */
1527 ip->ip_xindex = ip->ip_rindex;
1535 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1537 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1538 return(0); /* NOT REACHED */
1542 lwkt_wait_ipiq(int dcpu, int seq)
1544 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);