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 1.55 2004/02/17 19:38:49 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.
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
42 #include <sys/rtprio.h>
43 #include <sys/queue.h>
44 #include <sys/thread2.h>
45 #include <sys/sysctl.h>
46 #include <sys/kthread.h>
47 #include <machine/cpu.h>
52 #include <vm/vm_param.h>
53 #include <vm/vm_kern.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_pager.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_zone.h>
61 #include <machine/stdarg.h>
62 #include <machine/ipl.h>
63 #include <machine/smp.h>
65 #define THREAD_STACK (UPAGES * PAGE_SIZE)
69 #include <sys/stdint.h>
70 #include <libcaps/thread.h>
71 #include <sys/thread.h>
72 #include <sys/msgport.h>
73 #include <sys/errno.h>
74 #include <libcaps/globaldata.h>
75 #include <sys/thread2.h>
76 #include <sys/msgport2.h>
80 #include <machine/cpufunc.h>
81 #include <machine/lock.h>
85 static int untimely_switch = 0;
87 static int panic_on_cscount = 0;
89 static __int64_t switch_count = 0;
90 static __int64_t preempt_hit = 0;
91 static __int64_t preempt_miss = 0;
92 static __int64_t preempt_weird = 0;
96 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
98 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
100 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
101 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
102 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
103 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
108 * These helper procedures handle the runq, they can only be called from
109 * within a critical section.
111 * WARNING! Prior to SMP being brought up it is possible to enqueue and
112 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
113 * instead of 'mycpu' when referencing the globaldata structure. Once
114 * SMP live enqueuing and dequeueing only occurs on the current cpu.
118 _lwkt_dequeue(thread_t td)
120 if (td->td_flags & TDF_RUNQ) {
121 int nq = td->td_pri & TDPRI_MASK;
122 struct globaldata *gd = td->td_gd;
124 td->td_flags &= ~TDF_RUNQ;
125 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
126 /* runqmask is passively cleaned up by the switcher */
132 _lwkt_enqueue(thread_t td)
134 if ((td->td_flags & TDF_RUNQ) == 0) {
135 int nq = td->td_pri & TDPRI_MASK;
136 struct globaldata *gd = td->td_gd;
138 td->td_flags |= TDF_RUNQ;
139 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
140 gd->gd_runqmask |= 1 << nq;
146 _lwkt_wantresched(thread_t ntd, thread_t cur)
148 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
154 * LWKTs operate on a per-cpu basis
156 * WARNING! Called from early boot, 'mycpu' may not work yet.
159 lwkt_gdinit(struct globaldata *gd)
163 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
164 TAILQ_INIT(&gd->gd_tdrunq[i]);
166 TAILQ_INIT(&gd->gd_tdallq);
172 * Initialize a thread wait structure prior to first use.
174 * NOTE! called from low level boot code, we cannot do anything fancy!
177 lwkt_init_wait(lwkt_wait_t w)
179 TAILQ_INIT(&w->wa_waitq);
183 * Create a new thread. The thread must be associated with a process context
184 * or LWKT start address before it can be scheduled. If the target cpu is
185 * -1 the thread will be created on the current cpu.
187 * If you intend to create a thread without a process context this function
188 * does everything except load the startup and switcher function.
191 lwkt_alloc_thread(struct thread *td, int cpu)
198 if (mycpu->gd_tdfreecount > 0) {
199 --mycpu->gd_tdfreecount;
200 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
201 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
202 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
203 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
205 stack = td->td_kstack;
206 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
210 td = zalloc(thread_zone);
212 td = malloc(sizeof(struct thread));
214 td->td_kstack = NULL;
215 flags |= TDF_ALLOCATED_THREAD;
218 if ((stack = td->td_kstack) == NULL) {
220 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
222 stack = libcaps_alloc_stack(THREAD_STACK);
224 flags |= TDF_ALLOCATED_STACK;
227 lwkt_init_thread(td, stack, flags, mycpu);
229 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
236 * Initialize a preexisting thread structure. This function is used by
237 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
239 * All threads start out in a critical section at a priority of
240 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
241 * appropriate. This function may send an IPI message when the
242 * requested cpu is not the current cpu and consequently gd_tdallq may
243 * not be initialized synchronously from the point of view of the originating
246 * NOTE! we have to be careful in regards to creating threads for other cpus
247 * if SMP has not yet been activated.
250 lwkt_init_thread_remote(void *arg)
254 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
258 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
260 bzero(td, sizeof(struct thread));
261 td->td_kstack = stack;
262 td->td_flags |= flags;
264 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
265 lwkt_initport(&td->td_msgport, td);
266 pmap_init_thread(td);
270 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
273 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
277 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
285 lwkt_set_comm(thread_t td, const char *ctl, ...)
290 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
295 lwkt_hold(thread_t td)
301 lwkt_rele(thread_t td)
303 KKASSERT(td->td_refs > 0);
310 lwkt_wait_free(thread_t td)
313 tsleep(td, 0, "tdreap", hz);
319 lwkt_free_thread(thread_t td)
321 struct globaldata *gd = mycpu;
323 KASSERT((td->td_flags & TDF_RUNNING) == 0,
324 ("lwkt_free_thread: did not exit! %p", td));
327 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
328 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
329 (td->td_flags & TDF_ALLOCATED_THREAD)
331 ++gd->gd_tdfreecount;
332 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
336 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
338 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
340 libcaps_free_stack(td->td_kstack, THREAD_STACK);
343 td->td_kstack = NULL;
345 if (td->td_flags & TDF_ALLOCATED_THREAD) {
347 zfree(thread_zone, td);
357 * Switch to the next runnable lwkt. If no LWKTs are runnable then
358 * switch to the idlethread. Switching must occur within a critical
359 * section to avoid races with the scheduling queue.
361 * We always have full control over our cpu's run queue. Other cpus
362 * that wish to manipulate our queue must use the cpu_*msg() calls to
363 * talk to our cpu, so a critical section is all that is needed and
364 * the result is very, very fast thread switching.
366 * The LWKT scheduler uses a fixed priority model and round-robins at
367 * each priority level. User process scheduling is a totally
368 * different beast and LWKT priorities should not be confused with
369 * user process priorities.
371 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
372 * cleans it up. Note that the td_switch() function cannot do anything that
373 * requires the MP lock since the MP lock will have already been setup for
374 * the target thread (not the current thread). It's nice to have a scheduler
375 * that does not need the MP lock to work because it allows us to do some
376 * really cool high-performance MP lock optimizations.
382 struct globaldata *gd;
383 thread_t td = curthread;
390 * Switching from within a 'fast' (non thread switched) interrupt is
393 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
394 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
398 * Passive release (used to transition from user to kernel mode
399 * when we block or switch rather then when we enter the kernel).
400 * This function is NOT called if we are switching into a preemption
401 * or returning from a preemption. Typically this causes us to lose
402 * our P_CURPROC designation (if we have one) and become a true LWKT
403 * thread, and may also hand P_CURPROC to another process and schedule
414 * td_mpcount cannot be used to determine if we currently hold the
415 * MP lock because get_mplock() will increment it prior to attempting
416 * to get the lock, and switch out if it can't. Our ownership of
417 * the actual lock will remain stable while we are in a critical section
418 * (but, of course, another cpu may own or release the lock so the
419 * actual value of mp_lock is not stable).
421 mpheld = MP_LOCK_HELD();
423 if (td->td_cscount) {
424 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
426 if (panic_on_cscount)
427 panic("switching while mastering cpusync");
431 if ((ntd = td->td_preempted) != NULL) {
433 * We had preempted another thread on this cpu, resume the preempted
434 * thread. This occurs transparently, whether the preempted thread
435 * was scheduled or not (it may have been preempted after descheduling
438 * We have to setup the MP lock for the original thread after backing
439 * out the adjustment that was made to curthread when the original
442 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
444 if (ntd->td_mpcount && mpheld == 0) {
445 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
446 td, ntd, td->td_mpcount, ntd->td_mpcount);
448 if (ntd->td_mpcount) {
449 td->td_mpcount -= ntd->td_mpcount;
450 KKASSERT(td->td_mpcount >= 0);
453 ntd->td_flags |= TDF_PREEMPT_DONE;
454 /* YYY release mp lock on switchback if original doesn't need it */
457 * Priority queue / round-robin at each priority. Note that user
458 * processes run at a fixed, low priority and the user process
459 * scheduler deals with interactions between user processes
460 * by scheduling and descheduling them from the LWKT queue as
463 * We have to adjust the MP lock for the target thread. If we
464 * need the MP lock and cannot obtain it we try to locate a
465 * thread that does not need the MP lock.
469 if (gd->gd_runqmask) {
470 int nq = bsrl(gd->gd_runqmask);
471 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
472 gd->gd_runqmask &= ~(1 << nq);
476 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
478 * Target needs MP lock and we couldn't get it, try
479 * to locate a thread which does not need the MP lock
480 * to run. If we cannot locate a thread spin in idle.
482 u_int32_t rqmask = gd->gd_runqmask;
484 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
485 if (ntd->td_mpcount == 0)
490 rqmask &= ~(1 << nq);
494 ntd = &gd->gd_idlethread;
495 ntd->td_flags |= TDF_IDLE_NOHLT;
497 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
498 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
501 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
502 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
505 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
506 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
510 * We have nothing to run but only let the idle loop halt
511 * the cpu if there are no pending interrupts.
513 ntd = &gd->gd_idlethread;
514 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
515 ntd->td_flags |= TDF_IDLE_NOHLT;
518 KASSERT(ntd->td_pri >= TDPRI_CRIT,
519 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
522 * Do the actual switch. If the new target does not need the MP lock
523 * and we are holding it, release the MP lock. If the new target requires
524 * the MP lock we have already acquired it for the target.
527 if (ntd->td_mpcount == 0 ) {
531 ASSERT_MP_LOCK_HELD();
542 * Switch if another thread has a higher priority. Do not switch to other
543 * threads at the same priority.
548 struct globaldata *gd = mycpu;
549 struct thread *td = gd->gd_curthread;
551 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
557 * Request that the target thread preempt the current thread. Preemption
558 * only works under a specific set of conditions:
560 * - We are not preempting ourselves
561 * - The target thread is owned by the current cpu
562 * - We are not currently being preempted
563 * - The target is not currently being preempted
564 * - We are able to satisfy the target's MP lock requirements (if any).
566 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
567 * this is called via lwkt_schedule() through the td_preemptable callback.
568 * critpri is the managed critical priority that we should ignore in order
569 * to determine whether preemption is possible (aka usually just the crit
570 * priority of lwkt_schedule() itself).
572 * XXX at the moment we run the target thread in a critical section during
573 * the preemption in order to prevent the target from taking interrupts
574 * that *WE* can't. Preemption is strictly limited to interrupt threads
575 * and interrupt-like threads, outside of a critical section, and the
576 * preempted source thread will be resumed the instant the target blocks
577 * whether or not the source is scheduled (i.e. preemption is supposed to
578 * be as transparent as possible).
580 * The target thread inherits our MP count (added to its own) for the
581 * duration of the preemption in order to preserve the atomicy of the
582 * MP lock during the preemption. Therefore, any preempting targets must be
583 * careful in regards to MP assertions. Note that the MP count may be
584 * out of sync with the physical mp_lock, but we do not have to preserve
585 * the original ownership of the lock if it was out of synch (that is, we
586 * can leave it synchronized on return).
589 lwkt_preempt(thread_t ntd, int critpri)
591 struct globaldata *gd = mycpu;
592 thread_t td = gd->gd_curthread;
599 * The caller has put us in a critical section. We can only preempt
600 * if the caller of the caller was not in a critical section (basically
601 * a local interrupt), as determined by the 'critpri' parameter. If
602 * we are unable to preempt
604 * YYY The target thread must be in a critical section (else it must
605 * inherit our critical section? I dunno yet).
607 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
610 if (!_lwkt_wantresched(ntd, td)) {
614 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
619 if (ntd->td_gd != gd) {
624 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
628 if (ntd->td_preempted) {
634 * note: an interrupt might have occured just as we were transitioning
635 * to or from the MP lock. In this case td_mpcount will be pre-disposed
636 * (non-zero) but not actually synchronized with the actual state of the
637 * lock. We can use it to imply an MP lock requirement for the
638 * preemption but we cannot use it to test whether we hold the MP lock
641 savecnt = td->td_mpcount;
642 mpheld = MP_LOCK_HELD();
643 ntd->td_mpcount += td->td_mpcount;
644 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
645 ntd->td_mpcount -= td->td_mpcount;
652 ntd->td_preempted = td;
653 td->td_flags |= TDF_PREEMPT_LOCK;
655 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
657 KKASSERT(savecnt == td->td_mpcount);
658 mpheld = MP_LOCK_HELD();
659 if (mpheld && td->td_mpcount == 0)
661 else if (mpheld == 0 && td->td_mpcount)
662 panic("lwkt_preempt(): MP lock was not held through");
664 ntd->td_preempted = NULL;
665 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
669 * Yield our thread while higher priority threads are pending. This is
670 * typically called when we leave a critical section but it can be safely
671 * called while we are in a critical section.
673 * This function will not generally yield to equal priority threads but it
674 * can occur as a side effect. Note that lwkt_switch() is called from
675 * inside the critical section to prevent its own crit_exit() from reentering
676 * lwkt_yield_quick().
678 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
679 * came along but was blocked and made pending.
681 * (self contained on a per cpu basis)
684 lwkt_yield_quick(void)
686 globaldata_t gd = mycpu;
687 thread_t td = gd->gd_curthread;
690 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
691 * it with a non-zero cpl then we might not wind up calling splz after
692 * a task switch when the critical section is exited even though the
693 * new task could accept the interrupt.
695 * XXX from crit_exit() only called after last crit section is released.
696 * If called directly will run splz() even if in a critical section.
698 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
699 * except for this special case, we MUST call splz() here to handle any
700 * pending ints, particularly after we switch, or we might accidently
701 * halt the cpu with interrupts pending.
703 if (gd->gd_reqflags && td->td_nest_count < 2)
707 * YYY enabling will cause wakeup() to task-switch, which really
708 * confused the old 4.x code. This is a good way to simulate
709 * preemption and MP without actually doing preemption or MP, because a
710 * lot of code assumes that wakeup() does not block.
712 if (untimely_switch && td->td_nest_count == 0 &&
713 gd->gd_intr_nesting_level == 0
717 * YYY temporary hacks until we disassociate the userland scheduler
718 * from the LWKT scheduler.
720 if (td->td_flags & TDF_RUNQ) {
721 lwkt_switch(); /* will not reenter yield function */
723 lwkt_schedule_self(); /* make sure we are scheduled */
724 lwkt_switch(); /* will not reenter yield function */
725 lwkt_deschedule_self(); /* make sure we are descheduled */
727 crit_exit_noyield(td);
732 * This implements a normal yield which, unlike _quick, will yield to equal
733 * priority threads as well. Note that gd_reqflags tests will be handled by
734 * the crit_exit() call in lwkt_switch().
736 * (self contained on a per cpu basis)
741 lwkt_schedule_self();
746 * Schedule a thread to run. As the current thread we can always safely
747 * schedule ourselves, and a shortcut procedure is provided for that
750 * (non-blocking, self contained on a per cpu basis)
753 lwkt_schedule_self(void)
755 thread_t td = curthread;
758 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
761 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
762 panic("SCHED SELF PANIC");
768 * Generic schedule. Possibly schedule threads belonging to other cpus and
769 * deal with threads that might be blocked on a wait queue.
771 * YYY this is one of the best places to implement load balancing code.
772 * Load balancing can be accomplished by requesting other sorts of actions
773 * for the thread in question.
776 lwkt_schedule(thread_t td)
779 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
780 && td->td_proc->p_stat == SSLEEP
782 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
784 curthread->td_proc ? curthread->td_proc->p_pid : -1,
785 curthread->td_proc ? curthread->td_proc->p_stat : -1,
787 td->td_proc ? curthread->td_proc->p_pid : -1,
788 td->td_proc ? curthread->td_proc->p_stat : -1
790 panic("SCHED PANIC");
794 if (td == curthread) {
800 * If the thread is on a wait list we have to send our scheduling
801 * request to the owner of the wait structure. Otherwise we send
802 * the scheduling request to the cpu owning the thread. Races
803 * are ok, the target will forward the message as necessary (the
804 * message may chase the thread around before it finally gets
807 * (remember, wait structures use stable storage)
809 if ((w = td->td_wait) != NULL) {
810 if (lwkt_trytoken(&w->wa_token)) {
811 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
815 if (td->td_gd == mycpu) {
817 if (td->td_preemptable)
818 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
819 else if (_lwkt_wantresched(td, curthread))
822 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
826 if (td->td_preemptable)
827 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
828 else if (_lwkt_wantresched(td, curthread))
831 lwkt_reltoken(&w->wa_token);
833 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
837 * If the wait structure is NULL and we own the thread, there
838 * is no race (since we are in a critical section). If we
839 * do not own the thread there might be a race but the
840 * target cpu will deal with it.
843 if (td->td_gd == mycpu) {
845 if (td->td_preemptable) {
846 td->td_preemptable(td, TDPRI_CRIT);
847 } else if (_lwkt_wantresched(td, curthread)) {
851 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
855 if (td->td_preemptable) {
856 td->td_preemptable(td, TDPRI_CRIT);
857 } else if (_lwkt_wantresched(td, curthread)) {
867 * Managed acquisition. This code assumes that the MP lock is held for
868 * the tdallq operation and that the thread has been descheduled from its
869 * original cpu. We also have to wait for the thread to be entirely switched
870 * out on its original cpu (this is usually fast enough that we never loop)
871 * since the LWKT system does not have to hold the MP lock while switching
872 * and the target may have released it before switching.
875 lwkt_acquire(thread_t td)
877 struct globaldata *gd;
880 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
881 while (td->td_flags & TDF_RUNNING) /* XXX spin */
885 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
888 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
894 * Deschedule a thread.
896 * (non-blocking, self contained on a per cpu basis)
899 lwkt_deschedule_self(void)
901 thread_t td = curthread;
904 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
910 * Generic deschedule. Descheduling threads other then your own should be
911 * done only in carefully controlled circumstances. Descheduling is
914 * This function may block if the cpu has run out of messages.
917 lwkt_deschedule(thread_t td)
920 if (td == curthread) {
923 if (td->td_gd == mycpu) {
926 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
933 * Set the target thread's priority. This routine does not automatically
934 * switch to a higher priority thread, LWKT threads are not designed for
935 * continuous priority changes. Yield if you want to switch.
937 * We have to retain the critical section count which uses the high bits
938 * of the td_pri field. The specified priority may also indicate zero or
939 * more critical sections by adding TDPRI_CRIT*N.
942 lwkt_setpri(thread_t td, int pri)
945 KKASSERT(td->td_gd == mycpu);
947 if (td->td_flags & TDF_RUNQ) {
949 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
952 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
958 lwkt_setpri_self(int pri)
960 thread_t td = curthread;
962 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
964 if (td->td_flags & TDF_RUNQ) {
966 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
969 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
975 lwkt_preempted_proc(void)
977 thread_t td = curthread;
978 while (td->td_preempted)
979 td = td->td_preempted;
986 * This function deschedules the current thread and blocks on the specified
987 * wait queue. We obtain ownership of the wait queue in order to block
988 * on it. A generation number is used to interlock the wait queue in case
989 * it gets signalled while we are blocked waiting on the token.
991 * Note: alternatively we could dequeue our thread and then message the
992 * target cpu owning the wait queue. YYY implement as sysctl.
994 * Note: wait queue signals normally ping-pong the cpu as an optimization.
998 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1000 thread_t td = curthread;
1002 lwkt_gettoken(&w->wa_token);
1003 if (w->wa_gen == *gen) {
1005 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1008 td->td_wmesg = wmesg;
1011 lwkt_regettoken(&w->wa_token);
1012 if (td->td_wmesg != NULL) {
1017 /* token might be lost, doesn't matter for gen update */
1019 lwkt_reltoken(&w->wa_token);
1023 * Signal a wait queue. We gain ownership of the wait queue in order to
1024 * signal it. Once a thread is removed from the wait queue we have to
1025 * deal with the cpu owning the thread.
1027 * Note: alternatively we could message the target cpu owning the wait
1028 * queue. YYY implement as sysctl.
1031 lwkt_signal(lwkt_wait_t w, int count)
1036 lwkt_gettoken(&w->wa_token);
1039 count = w->wa_count;
1040 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1043 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1045 td->td_wmesg = NULL;
1046 if (td->td_gd == mycpu) {
1049 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1051 lwkt_regettoken(&w->wa_token);
1053 lwkt_reltoken(&w->wa_token);
1059 * Create a kernel process/thread/whatever. It shares it's address space
1060 * with proc0 - ie: kernel only.
1062 * NOTE! By default new threads are created with the MP lock held. A
1063 * thread which does not require the MP lock should release it by calling
1064 * rel_mplock() at the start of the new thread.
1067 lwkt_create(void (*func)(void *), void *arg,
1068 struct thread **tdp, thread_t template, int tdflags, int cpu,
1069 const char *fmt, ...)
1074 td = lwkt_alloc_thread(template, cpu);
1077 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1078 td->td_flags |= TDF_VERBOSE | tdflags;
1084 * Set up arg0 for 'ps' etc
1086 __va_start(ap, fmt);
1087 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1091 * Schedule the thread to run
1093 if ((td->td_flags & TDF_STOPREQ) == 0)
1096 td->td_flags &= ~TDF_STOPREQ;
1101 * kthread_* is specific to the kernel and is not needed by userland.
1106 * Destroy an LWKT thread. Warning! This function is not called when
1107 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1108 * uses a different reaping mechanism.
1113 thread_t td = curthread;
1115 if (td->td_flags & TDF_VERBOSE)
1116 printf("kthread %p %s has exited\n", td, td->td_comm);
1119 lwkt_deschedule_self();
1120 ++mycpu->gd_tdfreecount;
1121 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1126 * Create a kernel process/thread/whatever. It shares it's address space
1127 * with proc0 - ie: kernel only. 5.x compatible.
1129 * NOTE! By default kthreads are created with the MP lock held. A
1130 * thread which does not require the MP lock should release it by calling
1131 * rel_mplock() at the start of the new thread.
1134 kthread_create(void (*func)(void *), void *arg,
1135 struct thread **tdp, const char *fmt, ...)
1140 td = lwkt_alloc_thread(NULL, -1);
1143 cpu_set_thread_handler(td, kthread_exit, func, arg);
1144 td->td_flags |= TDF_VERBOSE;
1150 * Set up arg0 for 'ps' etc
1152 __va_start(ap, fmt);
1153 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1157 * Schedule the thread to run
1164 * Destroy an LWKT thread. Warning! This function is not called when
1165 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1166 * uses a different reaping mechanism.
1168 * XXX duplicates lwkt_exit()
1176 #endif /* _KERNEL */
1181 thread_t td = curthread;
1182 int lpri = td->td_pri;
1185 panic("td_pri is/would-go negative! %p %d", td, lpri);