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 1.54 2004/02/15 02:14:41 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>
57 #include <vm/vm_param.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_object.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_map.h>
62 #include <vm/vm_pager.h>
63 #include <vm/vm_extern.h>
64 #include <vm/vm_zone.h>
66 #include <machine/stdarg.h>
67 #include <machine/ipl.h>
68 #include <machine/smp.h>
70 #define THREAD_STACK (UPAGES * PAGE_SIZE)
74 #include <sys/stdint.h>
75 #include <libcaps/thread.h>
76 #include <sys/thread.h>
77 #include <sys/msgport.h>
78 #include <sys/errno.h>
79 #include <libcaps/globaldata.h>
80 #include <sys/thread2.h>
81 #include <sys/msgport2.h>
85 #include <machine/cpufunc.h>
86 #include <machine/lock.h>
90 static int untimely_switch = 0;
91 static __int64_t switch_count = 0;
92 static __int64_t preempt_hit = 0;
93 static __int64_t preempt_miss = 0;
94 static __int64_t preempt_weird = 0;
98 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
99 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
100 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
101 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
102 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
107 * These helper procedures handle the runq, they can only be called from
108 * within a critical section.
110 * WARNING! Prior to SMP being brought up it is possible to enqueue and
111 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
112 * instead of 'mycpu' when referencing the globaldata structure. Once
113 * SMP live enqueuing and dequeueing only occurs on the current cpu.
117 _lwkt_dequeue(thread_t td)
119 if (td->td_flags & TDF_RUNQ) {
120 int nq = td->td_pri & TDPRI_MASK;
121 struct globaldata *gd = td->td_gd;
123 td->td_flags &= ~TDF_RUNQ;
124 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
125 /* runqmask is passively cleaned up by the switcher */
131 _lwkt_enqueue(thread_t td)
133 if ((td->td_flags & TDF_RUNQ) == 0) {
134 int nq = td->td_pri & TDPRI_MASK;
135 struct globaldata *gd = td->td_gd;
137 td->td_flags |= TDF_RUNQ;
138 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
139 gd->gd_runqmask |= 1 << nq;
145 _lwkt_wantresched(thread_t ntd, thread_t cur)
147 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
153 * LWKTs operate on a per-cpu basis
155 * WARNING! Called from early boot, 'mycpu' may not work yet.
158 lwkt_gdinit(struct globaldata *gd)
162 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
163 TAILQ_INIT(&gd->gd_tdrunq[i]);
165 TAILQ_INIT(&gd->gd_tdallq);
171 * Initialize a thread wait structure prior to first use.
173 * NOTE! called from low level boot code, we cannot do anything fancy!
176 lwkt_init_wait(lwkt_wait_t w)
178 TAILQ_INIT(&w->wa_waitq);
182 * Create a new thread. The thread must be associated with a process context
183 * or LWKT start address before it can be scheduled. If the target cpu is
184 * -1 the thread will be created on the current cpu.
186 * If you intend to create a thread without a process context this function
187 * does everything except load the startup and switcher function.
190 lwkt_alloc_thread(struct thread *td, int cpu)
197 if (mycpu->gd_tdfreecount > 0) {
198 --mycpu->gd_tdfreecount;
199 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
200 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
201 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
202 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
204 stack = td->td_kstack;
205 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
209 td = zalloc(thread_zone);
211 td = malloc(sizeof(struct thread));
213 td->td_kstack = NULL;
214 flags |= TDF_ALLOCATED_THREAD;
217 if ((stack = td->td_kstack) == NULL) {
219 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
221 stack = libcaps_alloc_stack(THREAD_STACK);
223 flags |= TDF_ALLOCATED_STACK;
226 lwkt_init_thread(td, stack, flags, mycpu);
228 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
235 * Initialize a preexisting thread structure. This function is used by
236 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
238 * All threads start out in a critical section at a priority of
239 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
240 * appropriate. This function may send an IPI message when the
241 * requested cpu is not the current cpu and consequently gd_tdallq may
242 * not be initialized synchronously from the point of view of the originating
245 * NOTE! we have to be careful in regards to creating threads for other cpus
246 * if SMP has not yet been activated.
249 lwkt_init_thread_remote(void *arg)
253 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
257 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
259 bzero(td, sizeof(struct thread));
260 td->td_kstack = stack;
261 td->td_flags |= flags;
263 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
264 lwkt_initport(&td->td_msgport, td);
265 pmap_init_thread(td);
266 if (smp_active == 0 || gd == mycpu) {
268 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
271 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
278 lwkt_set_comm(thread_t td, const char *ctl, ...)
283 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
288 lwkt_hold(thread_t td)
294 lwkt_rele(thread_t td)
296 KKASSERT(td->td_refs > 0);
303 lwkt_wait_free(thread_t td)
306 tsleep(td, 0, "tdreap", hz);
312 lwkt_free_thread(thread_t td)
314 struct globaldata *gd = mycpu;
316 KASSERT((td->td_flags & TDF_RUNNING) == 0,
317 ("lwkt_free_thread: did not exit! %p", td));
320 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
321 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
322 (td->td_flags & TDF_ALLOCATED_THREAD)
324 ++gd->gd_tdfreecount;
325 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
329 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
331 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
333 libcaps_free_stack(td->td_kstack, THREAD_STACK);
336 td->td_kstack = NULL;
338 if (td->td_flags & TDF_ALLOCATED_THREAD) {
340 zfree(thread_zone, td);
350 * Switch to the next runnable lwkt. If no LWKTs are runnable then
351 * switch to the idlethread. Switching must occur within a critical
352 * section to avoid races with the scheduling queue.
354 * We always have full control over our cpu's run queue. Other cpus
355 * that wish to manipulate our queue must use the cpu_*msg() calls to
356 * talk to our cpu, so a critical section is all that is needed and
357 * the result is very, very fast thread switching.
359 * The LWKT scheduler uses a fixed priority model and round-robins at
360 * each priority level. User process scheduling is a totally
361 * different beast and LWKT priorities should not be confused with
362 * user process priorities.
364 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
365 * cleans it up. Note that the td_switch() function cannot do anything that
366 * requires the MP lock since the MP lock will have already been setup for
367 * the target thread (not the current thread). It's nice to have a scheduler
368 * that does not need the MP lock to work because it allows us to do some
369 * really cool high-performance MP lock optimizations.
375 struct globaldata *gd;
376 thread_t td = curthread;
383 * Switching from within a 'fast' (non thread switched) interrupt is
386 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
387 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
391 * Passive release (used to transition from user to kernel mode
392 * when we block or switch rather then when we enter the kernel).
393 * This function is NOT called if we are switching into a preemption
394 * or returning from a preemption. Typically this causes us to lose
395 * our P_CURPROC designation (if we have one) and become a true LWKT
396 * thread, and may also hand P_CURPROC to another process and schedule
407 * td_mpcount cannot be used to determine if we currently hold the
408 * MP lock because get_mplock() will increment it prior to attempting
409 * to get the lock, and switch out if it can't. Our ownership of
410 * the actual lock will remain stable while we are in a critical section
411 * (but, of course, another cpu may own or release the lock so the
412 * actual value of mp_lock is not stable).
414 mpheld = MP_LOCK_HELD();
416 if ((ntd = td->td_preempted) != NULL) {
418 * We had preempted another thread on this cpu, resume the preempted
419 * thread. This occurs transparently, whether the preempted thread
420 * was scheduled or not (it may have been preempted after descheduling
423 * We have to setup the MP lock for the original thread after backing
424 * out the adjustment that was made to curthread when the original
427 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
429 if (ntd->td_mpcount && mpheld == 0) {
430 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
431 td, ntd, td->td_mpcount, ntd->td_mpcount);
433 if (ntd->td_mpcount) {
434 td->td_mpcount -= ntd->td_mpcount;
435 KKASSERT(td->td_mpcount >= 0);
438 ntd->td_flags |= TDF_PREEMPT_DONE;
439 /* YYY release mp lock on switchback if original doesn't need it */
442 * Priority queue / round-robin at each priority. Note that user
443 * processes run at a fixed, low priority and the user process
444 * scheduler deals with interactions between user processes
445 * by scheduling and descheduling them from the LWKT queue as
448 * We have to adjust the MP lock for the target thread. If we
449 * need the MP lock and cannot obtain it we try to locate a
450 * thread that does not need the MP lock.
454 if (gd->gd_runqmask) {
455 int nq = bsrl(gd->gd_runqmask);
456 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
457 gd->gd_runqmask &= ~(1 << nq);
461 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
463 * Target needs MP lock and we couldn't get it, try
464 * to locate a thread which does not need the MP lock
465 * to run. If we cannot locate a thread spin in idle.
467 u_int32_t rqmask = gd->gd_runqmask;
469 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
470 if (ntd->td_mpcount == 0)
475 rqmask &= ~(1 << nq);
479 ntd = &gd->gd_idlethread;
480 ntd->td_flags |= TDF_IDLE_NOHLT;
482 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
483 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
486 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
487 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
490 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
491 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
495 * We have nothing to run but only let the idle loop halt
496 * the cpu if there are no pending interrupts.
498 ntd = &gd->gd_idlethread;
499 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
500 ntd->td_flags |= TDF_IDLE_NOHLT;
503 KASSERT(ntd->td_pri >= TDPRI_CRIT,
504 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
507 * Do the actual switch. If the new target does not need the MP lock
508 * and we are holding it, release the MP lock. If the new target requires
509 * the MP lock we have already acquired it for the target.
512 if (ntd->td_mpcount == 0 ) {
516 ASSERT_MP_LOCK_HELD();
527 * Switch if another thread has a higher priority. Do not switch to other
528 * threads at the same priority.
533 struct globaldata *gd = mycpu;
534 struct thread *td = gd->gd_curthread;
536 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
542 * Request that the target thread preempt the current thread. Preemption
543 * only works under a specific set of conditions:
545 * - We are not preempting ourselves
546 * - The target thread is owned by the current cpu
547 * - We are not currently being preempted
548 * - The target is not currently being preempted
549 * - We are able to satisfy the target's MP lock requirements (if any).
551 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
552 * this is called via lwkt_schedule() through the td_preemptable callback.
553 * critpri is the managed critical priority that we should ignore in order
554 * to determine whether preemption is possible (aka usually just the crit
555 * priority of lwkt_schedule() itself).
557 * XXX at the moment we run the target thread in a critical section during
558 * the preemption in order to prevent the target from taking interrupts
559 * that *WE* can't. Preemption is strictly limited to interrupt threads
560 * and interrupt-like threads, outside of a critical section, and the
561 * preempted source thread will be resumed the instant the target blocks
562 * whether or not the source is scheduled (i.e. preemption is supposed to
563 * be as transparent as possible).
565 * The target thread inherits our MP count (added to its own) for the
566 * duration of the preemption in order to preserve the atomicy of the
567 * MP lock during the preemption. Therefore, any preempting targets must be
568 * careful in regards to MP assertions. Note that the MP count may be
569 * out of sync with the physical mp_lock, but we do not have to preserve
570 * the original ownership of the lock if it was out of synch (that is, we
571 * can leave it synchronized on return).
574 lwkt_preempt(thread_t ntd, int critpri)
576 struct globaldata *gd = mycpu;
577 thread_t td = gd->gd_curthread;
584 * The caller has put us in a critical section. We can only preempt
585 * if the caller of the caller was not in a critical section (basically
586 * a local interrupt), as determined by the 'critpri' parameter. If
587 * we are unable to preempt
589 * YYY The target thread must be in a critical section (else it must
590 * inherit our critical section? I dunno yet).
592 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
595 if (!_lwkt_wantresched(ntd, td)) {
599 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
604 if (ntd->td_gd != gd) {
609 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
613 if (ntd->td_preempted) {
619 * note: an interrupt might have occured just as we were transitioning
620 * to or from the MP lock. In this case td_mpcount will be pre-disposed
621 * (non-zero) but not actually synchronized with the actual state of the
622 * lock. We can use it to imply an MP lock requirement for the
623 * preemption but we cannot use it to test whether we hold the MP lock
626 savecnt = td->td_mpcount;
627 mpheld = MP_LOCK_HELD();
628 ntd->td_mpcount += td->td_mpcount;
629 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
630 ntd->td_mpcount -= td->td_mpcount;
637 ntd->td_preempted = td;
638 td->td_flags |= TDF_PREEMPT_LOCK;
640 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
642 KKASSERT(savecnt == td->td_mpcount);
643 mpheld = MP_LOCK_HELD();
644 if (mpheld && td->td_mpcount == 0)
646 else if (mpheld == 0 && td->td_mpcount)
647 panic("lwkt_preempt(): MP lock was not held through");
649 ntd->td_preempted = NULL;
650 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
654 * Yield our thread while higher priority threads are pending. This is
655 * typically called when we leave a critical section but it can be safely
656 * called while we are in a critical section.
658 * This function will not generally yield to equal priority threads but it
659 * can occur as a side effect. Note that lwkt_switch() is called from
660 * inside the critical section to prevent its own crit_exit() from reentering
661 * lwkt_yield_quick().
663 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
664 * came along but was blocked and made pending.
666 * (self contained on a per cpu basis)
669 lwkt_yield_quick(void)
671 globaldata_t gd = mycpu;
672 thread_t td = gd->gd_curthread;
675 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
676 * it with a non-zero cpl then we might not wind up calling splz after
677 * a task switch when the critical section is exited even though the
678 * new task could accept the interrupt.
680 * XXX from crit_exit() only called after last crit section is released.
681 * If called directly will run splz() even if in a critical section.
683 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
684 * except for this special case, we MUST call splz() here to handle any
685 * pending ints, particularly after we switch, or we might accidently
686 * halt the cpu with interrupts pending.
688 if (gd->gd_reqflags && td->td_nest_count < 2)
692 * YYY enabling will cause wakeup() to task-switch, which really
693 * confused the old 4.x code. This is a good way to simulate
694 * preemption and MP without actually doing preemption or MP, because a
695 * lot of code assumes that wakeup() does not block.
697 if (untimely_switch && td->td_nest_count == 0 &&
698 gd->gd_intr_nesting_level == 0
702 * YYY temporary hacks until we disassociate the userland scheduler
703 * from the LWKT scheduler.
705 if (td->td_flags & TDF_RUNQ) {
706 lwkt_switch(); /* will not reenter yield function */
708 lwkt_schedule_self(); /* make sure we are scheduled */
709 lwkt_switch(); /* will not reenter yield function */
710 lwkt_deschedule_self(); /* make sure we are descheduled */
712 crit_exit_noyield(td);
717 * This implements a normal yield which, unlike _quick, will yield to equal
718 * priority threads as well. Note that gd_reqflags tests will be handled by
719 * the crit_exit() call in lwkt_switch().
721 * (self contained on a per cpu basis)
726 lwkt_schedule_self();
731 * Schedule a thread to run. As the current thread we can always safely
732 * schedule ourselves, and a shortcut procedure is provided for that
735 * (non-blocking, self contained on a per cpu basis)
738 lwkt_schedule_self(void)
740 thread_t td = curthread;
743 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
746 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
747 panic("SCHED SELF PANIC");
753 * Generic schedule. Possibly schedule threads belonging to other cpus and
754 * deal with threads that might be blocked on a wait queue.
756 * YYY this is one of the best places to implement load balancing code.
757 * Load balancing can be accomplished by requesting other sorts of actions
758 * for the thread in question.
761 lwkt_schedule(thread_t td)
764 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
765 && td->td_proc->p_stat == SSLEEP
767 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
769 curthread->td_proc ? curthread->td_proc->p_pid : -1,
770 curthread->td_proc ? curthread->td_proc->p_stat : -1,
772 td->td_proc ? curthread->td_proc->p_pid : -1,
773 td->td_proc ? curthread->td_proc->p_stat : -1
775 panic("SCHED PANIC");
779 if (td == curthread) {
785 * If the thread is on a wait list we have to send our scheduling
786 * request to the owner of the wait structure. Otherwise we send
787 * the scheduling request to the cpu owning the thread. Races
788 * are ok, the target will forward the message as necessary (the
789 * message may chase the thread around before it finally gets
792 * (remember, wait structures use stable storage)
794 if ((w = td->td_wait) != NULL) {
795 if (lwkt_trytoken(&w->wa_token)) {
796 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
799 if (smp_active == 0 || td->td_gd == mycpu) {
801 if (td->td_preemptable) {
802 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
803 } else if (_lwkt_wantresched(td, curthread)) {
807 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
809 lwkt_reltoken(&w->wa_token);
811 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
815 * If the wait structure is NULL and we own the thread, there
816 * is no race (since we are in a critical section). If we
817 * do not own the thread there might be a race but the
818 * target cpu will deal with it.
820 if (smp_active == 0 || td->td_gd == mycpu) {
822 if (td->td_preemptable) {
823 td->td_preemptable(td, TDPRI_CRIT);
824 } else if (_lwkt_wantresched(td, curthread)) {
828 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
836 * Managed acquisition. This code assumes that the MP lock is held for
837 * the tdallq operation and that the thread has been descheduled from its
838 * original cpu. We also have to wait for the thread to be entirely switched
839 * out on its original cpu (this is usually fast enough that we never loop)
840 * since the LWKT system does not have to hold the MP lock while switching
841 * and the target may have released it before switching.
844 lwkt_acquire(thread_t td)
846 struct globaldata *gd;
849 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
850 while (td->td_flags & TDF_RUNNING) /* XXX spin */
854 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
857 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
863 * Deschedule a thread.
865 * (non-blocking, self contained on a per cpu basis)
868 lwkt_deschedule_self(void)
870 thread_t td = curthread;
873 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
879 * Generic deschedule. Descheduling threads other then your own should be
880 * done only in carefully controlled circumstances. Descheduling is
883 * This function may block if the cpu has run out of messages.
886 lwkt_deschedule(thread_t td)
889 if (td == curthread) {
892 if (td->td_gd == mycpu) {
895 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
902 * Set the target thread's priority. This routine does not automatically
903 * switch to a higher priority thread, LWKT threads are not designed for
904 * continuous priority changes. Yield if you want to switch.
906 * We have to retain the critical section count which uses the high bits
907 * of the td_pri field. The specified priority may also indicate zero or
908 * more critical sections by adding TDPRI_CRIT*N.
911 lwkt_setpri(thread_t td, int pri)
914 KKASSERT(td->td_gd == mycpu);
916 if (td->td_flags & TDF_RUNQ) {
918 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
921 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
927 lwkt_setpri_self(int pri)
929 thread_t td = curthread;
931 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
933 if (td->td_flags & TDF_RUNQ) {
935 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
938 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
944 lwkt_preempted_proc(void)
946 thread_t td = curthread;
947 while (td->td_preempted)
948 td = td->td_preempted;
955 * This function deschedules the current thread and blocks on the specified
956 * wait queue. We obtain ownership of the wait queue in order to block
957 * on it. A generation number is used to interlock the wait queue in case
958 * it gets signalled while we are blocked waiting on the token.
960 * Note: alternatively we could dequeue our thread and then message the
961 * target cpu owning the wait queue. YYY implement as sysctl.
963 * Note: wait queue signals normally ping-pong the cpu as an optimization.
967 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
969 thread_t td = curthread;
971 lwkt_gettoken(&w->wa_token);
972 if (w->wa_gen == *gen) {
974 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
977 td->td_wmesg = wmesg;
980 lwkt_regettoken(&w->wa_token);
981 if (td->td_wmesg != NULL) {
986 /* token might be lost, doesn't matter for gen update */
988 lwkt_reltoken(&w->wa_token);
992 * Signal a wait queue. We gain ownership of the wait queue in order to
993 * signal it. Once a thread is removed from the wait queue we have to
994 * deal with the cpu owning the thread.
996 * Note: alternatively we could message the target cpu owning the wait
997 * queue. YYY implement as sysctl.
1000 lwkt_signal(lwkt_wait_t w, int count)
1005 lwkt_gettoken(&w->wa_token);
1008 count = w->wa_count;
1009 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1012 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1014 td->td_wmesg = NULL;
1015 if (td->td_gd == mycpu) {
1018 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1020 lwkt_regettoken(&w->wa_token);
1022 lwkt_reltoken(&w->wa_token);
1028 * Create a kernel process/thread/whatever. It shares it's address space
1029 * with proc0 - ie: kernel only.
1031 * NOTE! By default new threads are created with the MP lock held. A
1032 * thread which does not require the MP lock should release it by calling
1033 * rel_mplock() at the start of the new thread.
1036 lwkt_create(void (*func)(void *), void *arg,
1037 struct thread **tdp, thread_t template, int tdflags, int cpu,
1038 const char *fmt, ...)
1043 td = lwkt_alloc_thread(template, cpu);
1046 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1047 td->td_flags |= TDF_VERBOSE | tdflags;
1053 * Set up arg0 for 'ps' etc
1055 __va_start(ap, fmt);
1056 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1060 * Schedule the thread to run
1062 if ((td->td_flags & TDF_STOPREQ) == 0)
1065 td->td_flags &= ~TDF_STOPREQ;
1070 * kthread_* is specific to the kernel and is not needed by userland.
1075 * Destroy an LWKT thread. Warning! This function is not called when
1076 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1077 * uses a different reaping mechanism.
1082 thread_t td = curthread;
1084 if (td->td_flags & TDF_VERBOSE)
1085 printf("kthread %p %s has exited\n", td, td->td_comm);
1088 lwkt_deschedule_self();
1089 ++mycpu->gd_tdfreecount;
1090 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1095 * Create a kernel process/thread/whatever. It shares it's address space
1096 * with proc0 - ie: kernel only. 5.x compatible.
1098 * NOTE! By default kthreads are created with the MP lock held. A
1099 * thread which does not require the MP lock should release it by calling
1100 * rel_mplock() at the start of the new thread.
1103 kthread_create(void (*func)(void *), void *arg,
1104 struct thread **tdp, const char *fmt, ...)
1109 td = lwkt_alloc_thread(NULL, -1);
1112 cpu_set_thread_handler(td, kthread_exit, func, arg);
1113 td->td_flags |= TDF_VERBOSE;
1119 * Set up arg0 for 'ps' etc
1121 __va_start(ap, fmt);
1122 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1126 * Schedule the thread to run
1133 * Destroy an LWKT thread. Warning! This function is not called when
1134 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1135 * uses a different reaping mechanism.
1137 * XXX duplicates lwkt_exit()
1145 #endif /* _KERNEL */
1150 thread_t td = curthread;
1151 int lpri = td->td_pri;
1154 panic("td_pri is/would-go negative! %p %d", td, lpri);