2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.76 2005/07/07 20:28:26 hmp Exp $
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
50 #include <sys/rtprio.h>
51 #include <sys/queue.h>
52 #include <sys/thread2.h>
53 #include <sys/sysctl.h>
54 #include <sys/kthread.h>
55 #include <machine/cpu.h>
60 #include <vm/vm_param.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_pager.h>
66 #include <vm/vm_extern.h>
67 #include <vm/vm_zone.h>
69 #include <machine/stdarg.h>
70 #include <machine/ipl.h>
71 #include <machine/smp.h>
75 #include <sys/stdint.h>
76 #include <libcaps/thread.h>
77 #include <sys/thread.h>
78 #include <sys/msgport.h>
79 #include <sys/errno.h>
80 #include <libcaps/globaldata.h>
81 #include <machine/cpufunc.h>
82 #include <sys/thread2.h>
83 #include <sys/msgport2.h>
87 #include <machine/lock.h>
88 #include <machine/atomic.h>
89 #include <machine/cpu.h>
93 static int untimely_switch = 0;
95 static int panic_on_cscount = 0;
97 static __int64_t switch_count = 0;
98 static __int64_t preempt_hit = 0;
99 static __int64_t preempt_miss = 0;
100 static __int64_t preempt_weird = 0;
101 static __int64_t token_contention_count = 0;
102 static __int64_t mplock_contention_count = 0;
106 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
108 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
112 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
116 &token_contention_count, 0, "spinning due to token contention");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
118 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
133 _lwkt_dequeue(thread_t td)
135 if (td->td_flags & TDF_RUNQ) {
136 int nq = td->td_pri & TDPRI_MASK;
137 struct globaldata *gd = td->td_gd;
139 td->td_flags &= ~TDF_RUNQ;
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
147 _lwkt_enqueue(thread_t td)
149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING)) == 0) {
150 int nq = td->td_pri & TDPRI_MASK;
151 struct globaldata *gd = td->td_gd;
153 td->td_flags |= TDF_RUNQ;
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
160 * Schedule a thread to run. As the current thread we can always safely
161 * schedule ourselves, and a shortcut procedure is provided for that
164 * (non-blocking, self contained on a per cpu basis)
167 lwkt_schedule_self(thread_t td)
169 crit_enter_quick(td);
170 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
171 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
174 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
175 panic("SCHED SELF PANIC");
181 * Deschedule a thread.
183 * (non-blocking, self contained on a per cpu basis)
186 lwkt_deschedule_self(thread_t td)
188 crit_enter_quick(td);
189 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
197 * LWKTs operate on a per-cpu basis
199 * WARNING! Called from early boot, 'mycpu' may not work yet.
202 lwkt_gdinit(struct globaldata *gd)
206 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
207 TAILQ_INIT(&gd->gd_tdrunq[i]);
209 TAILQ_INIT(&gd->gd_tdallq);
215 * Initialize a thread wait structure prior to first use.
217 * NOTE! called from low level boot code, we cannot do anything fancy!
220 lwkt_wait_init(lwkt_wait_t w)
222 lwkt_token_init(&w->wa_token);
223 TAILQ_INIT(&w->wa_waitq);
229 * Create a new thread. The thread must be associated with a process context
230 * or LWKT start address before it can be scheduled. If the target cpu is
231 * -1 the thread will be created on the current cpu.
233 * If you intend to create a thread without a process context this function
234 * does everything except load the startup and switcher function.
237 lwkt_alloc_thread(struct thread *td, int stksize, int cpu)
241 globaldata_t gd = mycpu;
245 if (gd->gd_tdfreecount > 0) {
246 --gd->gd_tdfreecount;
247 td = TAILQ_FIRST(&gd->gd_tdfreeq);
248 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
249 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
250 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
252 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
256 td = zalloc(thread_zone);
258 td = malloc(sizeof(struct thread));
260 td->td_kstack = NULL;
261 td->td_kstack_size = 0;
262 flags |= TDF_ALLOCATED_THREAD;
265 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
266 if (flags & TDF_ALLOCATED_STACK) {
268 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
270 libcaps_free_stack(stack, td->td_kstack_size);
277 stack = (void *)kmem_alloc(kernel_map, stksize);
279 stack = libcaps_alloc_stack(stksize);
281 flags |= TDF_ALLOCATED_STACK;
284 lwkt_init_thread(td, stack, stksize, flags, mycpu);
286 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
293 * Initialize a preexisting thread structure. This function is used by
294 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
296 * All threads start out in a critical section at a priority of
297 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
298 * appropriate. This function may send an IPI message when the
299 * requested cpu is not the current cpu and consequently gd_tdallq may
300 * not be initialized synchronously from the point of view of the originating
303 * NOTE! we have to be careful in regards to creating threads for other cpus
304 * if SMP has not yet been activated.
309 lwkt_init_thread_remote(void *arg)
313 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
319 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
320 struct globaldata *gd)
322 globaldata_t mygd = mycpu;
324 bzero(td, sizeof(struct thread));
325 td->td_kstack = stack;
326 td->td_kstack_size = stksize;
327 td->td_flags |= flags;
329 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
330 lwkt_initport(&td->td_msgport, td);
331 pmap_init_thread(td);
334 * Normally initializing a thread for a remote cpu requires sending an
335 * IPI. However, the idlethread is setup before the other cpus are
336 * activated so we have to treat it as a special case. XXX manipulation
337 * of gd_tdallq requires the BGL.
339 if (gd == mygd || td == &gd->gd_idlethread) {
341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
344 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
356 lwkt_set_comm(thread_t td, const char *ctl, ...)
361 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
366 lwkt_hold(thread_t td)
372 lwkt_rele(thread_t td)
374 KKASSERT(td->td_refs > 0);
381 lwkt_wait_free(thread_t td)
384 tsleep(td, 0, "tdreap", hz);
390 lwkt_free_thread(thread_t td)
392 struct globaldata *gd = mycpu;
394 KASSERT((td->td_flags & TDF_RUNNING) == 0,
395 ("lwkt_free_thread: did not exit! %p", td));
398 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
399 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
400 (td->td_flags & TDF_ALLOCATED_THREAD)
402 ++gd->gd_tdfreecount;
403 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
407 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
409 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
411 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
414 td->td_kstack = NULL;
415 td->td_kstack_size = 0;
417 if (td->td_flags & TDF_ALLOCATED_THREAD) {
419 zfree(thread_zone, td);
429 * Switch to the next runnable lwkt. If no LWKTs are runnable then
430 * switch to the idlethread. Switching must occur within a critical
431 * section to avoid races with the scheduling queue.
433 * We always have full control over our cpu's run queue. Other cpus
434 * that wish to manipulate our queue must use the cpu_*msg() calls to
435 * talk to our cpu, so a critical section is all that is needed and
436 * the result is very, very fast thread switching.
438 * The LWKT scheduler uses a fixed priority model and round-robins at
439 * each priority level. User process scheduling is a totally
440 * different beast and LWKT priorities should not be confused with
441 * user process priorities.
443 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
444 * cleans it up. Note that the td_switch() function cannot do anything that
445 * requires the MP lock since the MP lock will have already been setup for
446 * the target thread (not the current thread). It's nice to have a scheduler
447 * that does not need the MP lock to work because it allows us to do some
448 * really cool high-performance MP lock optimizations.
454 globaldata_t gd = mycpu;
455 thread_t td = gd->gd_curthread;
462 * Switching from within a 'fast' (non thread switched) interrupt is
465 if (gd->gd_intr_nesting_level && panicstr == NULL) {
466 panic("lwkt_switch: cannot switch from within a fast interrupt, yet, td %p\n", td);
470 * Passive release (used to transition from user to kernel mode
471 * when we block or switch rather then when we enter the kernel).
472 * This function is NOT called if we are switching into a preemption
473 * or returning from a preemption. Typically this causes us to lose
474 * our current process designation (if we have one) and become a true
475 * LWKT thread, and may also hand the current process designation to
476 * another process and schedule thread.
485 * td_mpcount cannot be used to determine if we currently hold the
486 * MP lock because get_mplock() will increment it prior to attempting
487 * to get the lock, and switch out if it can't. Our ownership of
488 * the actual lock will remain stable while we are in a critical section
489 * (but, of course, another cpu may own or release the lock so the
490 * actual value of mp_lock is not stable).
492 mpheld = MP_LOCK_HELD();
494 if (td->td_cscount) {
495 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
497 if (panic_on_cscount)
498 panic("switching while mastering cpusync");
502 if ((ntd = td->td_preempted) != NULL) {
504 * We had preempted another thread on this cpu, resume the preempted
505 * thread. This occurs transparently, whether the preempted thread
506 * was scheduled or not (it may have been preempted after descheduling
509 * We have to setup the MP lock for the original thread after backing
510 * out the adjustment that was made to curthread when the original
513 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
515 if (ntd->td_mpcount && mpheld == 0) {
516 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
517 td, ntd, td->td_mpcount, ntd->td_mpcount);
519 if (ntd->td_mpcount) {
520 td->td_mpcount -= ntd->td_mpcount;
521 KKASSERT(td->td_mpcount >= 0);
524 ntd->td_flags |= TDF_PREEMPT_DONE;
527 * XXX. The interrupt may have woken a thread up, we need to properly
528 * set the reschedule flag if the originally interrupted thread is at
531 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
533 /* YYY release mp lock on switchback if original doesn't need it */
536 * Priority queue / round-robin at each priority. Note that user
537 * processes run at a fixed, low priority and the user process
538 * scheduler deals with interactions between user processes
539 * by scheduling and descheduling them from the LWKT queue as
542 * We have to adjust the MP lock for the target thread. If we
543 * need the MP lock and cannot obtain it we try to locate a
544 * thread that does not need the MP lock. If we cannot, we spin
547 * A similar issue exists for the tokens held by the target thread.
548 * If we cannot obtain ownership of the tokens we cannot immediately
549 * schedule the thread.
553 * We are switching threads. If there are any pending requests for
554 * tokens we can satisfy all of them here.
557 if (gd->gd_tokreqbase)
558 lwkt_drain_token_requests();
562 * If an LWKT reschedule was requested, well that is what we are
563 * doing now so clear it.
565 clear_lwkt_resched();
567 if (gd->gd_runqmask) {
568 int nq = bsrl(gd->gd_runqmask);
569 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
570 gd->gd_runqmask &= ~(1 << nq);
575 * If the target needs the MP lock and we couldn't get it,
576 * or if the target is holding tokens and we could not
577 * gain ownership of the tokens, continue looking for a
578 * thread to schedule and spin instead of HLT if we can't.
580 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
581 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
583 u_int32_t rqmask = gd->gd_runqmask;
585 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
586 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
587 /* spinning due to MP lock being held */
589 ++mplock_contention_count;
593 mpheld = MP_LOCK_HELD();
594 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
595 /* spinning due to token contention */
597 ++token_contention_count;
605 rqmask &= ~(1 << nq);
609 ntd = &gd->gd_idlethread;
610 ntd->td_flags |= TDF_IDLE_NOHLT;
612 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
613 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
616 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
617 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
620 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
621 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
625 * We have nothing to run but only let the idle loop halt
626 * the cpu if there are no pending interrupts.
628 ntd = &gd->gd_idlethread;
629 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
630 ntd->td_flags |= TDF_IDLE_NOHLT;
633 KASSERT(ntd->td_pri >= TDPRI_CRIT,
634 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
637 * Do the actual switch. If the new target does not need the MP lock
638 * and we are holding it, release the MP lock. If the new target requires
639 * the MP lock we have already acquired it for the target.
642 if (ntd->td_mpcount == 0 ) {
646 ASSERT_MP_LOCK_HELD();
653 /* NOTE: current cpu may have changed after switch */
658 * Request that the target thread preempt the current thread. Preemption
659 * only works under a specific set of conditions:
661 * - We are not preempting ourselves
662 * - The target thread is owned by the current cpu
663 * - We are not currently being preempted
664 * - The target is not currently being preempted
665 * - We are able to satisfy the target's MP lock requirements (if any).
667 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
668 * this is called via lwkt_schedule() through the td_preemptable callback.
669 * critpri is the managed critical priority that we should ignore in order
670 * to determine whether preemption is possible (aka usually just the crit
671 * priority of lwkt_schedule() itself).
673 * XXX at the moment we run the target thread in a critical section during
674 * the preemption in order to prevent the target from taking interrupts
675 * that *WE* can't. Preemption is strictly limited to interrupt threads
676 * and interrupt-like threads, outside of a critical section, and the
677 * preempted source thread will be resumed the instant the target blocks
678 * whether or not the source is scheduled (i.e. preemption is supposed to
679 * be as transparent as possible).
681 * The target thread inherits our MP count (added to its own) for the
682 * duration of the preemption in order to preserve the atomicy of the
683 * MP lock during the preemption. Therefore, any preempting targets must be
684 * careful in regards to MP assertions. Note that the MP count may be
685 * out of sync with the physical mp_lock, but we do not have to preserve
686 * the original ownership of the lock if it was out of synch (that is, we
687 * can leave it synchronized on return).
690 lwkt_preempt(thread_t ntd, int critpri)
692 struct globaldata *gd = mycpu;
700 * The caller has put us in a critical section. We can only preempt
701 * if the caller of the caller was not in a critical section (basically
702 * a local interrupt), as determined by the 'critpri' parameter.
704 * YYY The target thread must be in a critical section (else it must
705 * inherit our critical section? I dunno yet).
707 * Any tokens held by the target may not be held by thread(s) being
708 * preempted. We take the easy way out and do not preempt if
709 * the target is holding tokens.
711 * Set need_lwkt_resched() unconditionally for now YYY.
713 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
715 td = gd->gd_curthread;
716 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
720 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
726 if (ntd->td_gd != gd) {
733 * Take the easy way out and do not preempt if the target is holding
734 * one or more tokens. We could test whether the thread(s) being
735 * preempted interlock against the target thread's tokens and whether
736 * we can get all the target thread's tokens, but this situation
737 * should not occur very often so its easier to simply not preempt.
739 if (ntd->td_toks != NULL) {
744 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
749 if (ntd->td_preempted) {
756 * note: an interrupt might have occured just as we were transitioning
757 * to or from the MP lock. In this case td_mpcount will be pre-disposed
758 * (non-zero) but not actually synchronized with the actual state of the
759 * lock. We can use it to imply an MP lock requirement for the
760 * preemption but we cannot use it to test whether we hold the MP lock
763 savecnt = td->td_mpcount;
764 mpheld = MP_LOCK_HELD();
765 ntd->td_mpcount += td->td_mpcount;
766 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
767 ntd->td_mpcount -= td->td_mpcount;
775 * Since we are able to preempt the current thread, there is no need to
776 * call need_lwkt_resched().
779 ntd->td_preempted = td;
780 td->td_flags |= TDF_PREEMPT_LOCK;
782 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
784 KKASSERT(savecnt == td->td_mpcount);
785 mpheld = MP_LOCK_HELD();
786 if (mpheld && td->td_mpcount == 0)
788 else if (mpheld == 0 && td->td_mpcount)
789 panic("lwkt_preempt(): MP lock was not held through");
791 ntd->td_preempted = NULL;
792 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
796 * Yield our thread while higher priority threads are pending. This is
797 * typically called when we leave a critical section but it can be safely
798 * called while we are in a critical section.
800 * This function will not generally yield to equal priority threads but it
801 * can occur as a side effect. Note that lwkt_switch() is called from
802 * inside the critical section to prevent its own crit_exit() from reentering
803 * lwkt_yield_quick().
805 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
806 * came along but was blocked and made pending.
808 * (self contained on a per cpu basis)
811 lwkt_yield_quick(void)
813 globaldata_t gd = mycpu;
814 thread_t td = gd->gd_curthread;
817 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
818 * it with a non-zero cpl then we might not wind up calling splz after
819 * a task switch when the critical section is exited even though the
820 * new task could accept the interrupt.
822 * XXX from crit_exit() only called after last crit section is released.
823 * If called directly will run splz() even if in a critical section.
825 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
826 * except for this special case, we MUST call splz() here to handle any
827 * pending ints, particularly after we switch, or we might accidently
828 * halt the cpu with interrupts pending.
830 if (gd->gd_reqflags && td->td_nest_count < 2)
834 * YYY enabling will cause wakeup() to task-switch, which really
835 * confused the old 4.x code. This is a good way to simulate
836 * preemption and MP without actually doing preemption or MP, because a
837 * lot of code assumes that wakeup() does not block.
839 if (untimely_switch && td->td_nest_count == 0 &&
840 gd->gd_intr_nesting_level == 0
842 crit_enter_quick(td);
844 * YYY temporary hacks until we disassociate the userland scheduler
845 * from the LWKT scheduler.
847 if (td->td_flags & TDF_RUNQ) {
848 lwkt_switch(); /* will not reenter yield function */
850 lwkt_schedule_self(td); /* make sure we are scheduled */
851 lwkt_switch(); /* will not reenter yield function */
852 lwkt_deschedule_self(td); /* make sure we are descheduled */
854 crit_exit_noyield(td);
859 * This implements a normal yield which, unlike _quick, will yield to equal
860 * priority threads as well. Note that gd_reqflags tests will be handled by
861 * the crit_exit() call in lwkt_switch().
863 * (self contained on a per cpu basis)
868 lwkt_schedule_self(curthread);
873 * Generic schedule. Possibly schedule threads belonging to other cpus and
874 * deal with threads that might be blocked on a wait queue.
876 * We have a little helper inline function which does additional work after
877 * the thread has been enqueued, including dealing with preemption and
878 * setting need_lwkt_resched() (which prevents the kernel from returning
879 * to userland until it has processed higher priority threads).
883 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
885 if (ntd->td_preemptable) {
886 ntd->td_preemptable(ntd, cpri); /* YYY +token */
887 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
888 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
895 lwkt_schedule(thread_t td)
897 globaldata_t mygd = mycpu;
900 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
901 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
902 && td->td_proc->p_stat == SSLEEP
904 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
906 curthread->td_proc ? curthread->td_proc->p_pid : -1,
907 curthread->td_proc ? curthread->td_proc->p_stat : -1,
909 td->td_proc ? curthread->td_proc->p_pid : -1,
910 td->td_proc ? curthread->td_proc->p_stat : -1
912 panic("SCHED PANIC");
916 if (td == mygd->gd_curthread) {
922 * If the thread is on a wait list we have to send our scheduling
923 * request to the owner of the wait structure. Otherwise we send
924 * the scheduling request to the cpu owning the thread. Races
925 * are ok, the target will forward the message as necessary (the
926 * message may chase the thread around before it finally gets
929 * (remember, wait structures use stable storage)
931 * NOTE: we have to account for the number of critical sections
932 * under our control when calling _lwkt_schedule_post() so it
933 * can figure out whether preemption is allowed.
935 * NOTE: The wait structure algorithms are a mess and need to be
938 * NOTE: We cannot safely acquire or release a token, even
939 * non-blocking, because this routine may be called in the context
940 * of a thread already holding the token and thus not provide any
941 * interlock protection. We cannot safely manipulate the td_toks
942 * list for the same reason. Instead we depend on our critical
943 * section if the token is owned by our cpu.
945 if ((w = td->td_wait) != NULL) {
946 if (w->wa_token.t_cpu == mygd) {
947 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
951 if (td->td_gd == mygd) {
953 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
955 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
959 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
962 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
966 * If the wait structure is NULL and we own the thread, there
967 * is no race (since we are in a critical section). If we
968 * do not own the thread there might be a race but the
969 * target cpu will deal with it.
972 if (td->td_gd == mygd) {
974 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
976 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
980 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
988 * Managed acquisition. This code assumes that the MP lock is held for
989 * the tdallq operation and that the thread has been descheduled from its
990 * original cpu. We also have to wait for the thread to be entirely switched
991 * out on its original cpu (this is usually fast enough that we never loop)
992 * since the LWKT system does not have to hold the MP lock while switching
993 * and the target may have released it before switching.
996 lwkt_acquire(thread_t td)
1004 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1005 while (td->td_flags & TDF_RUNNING) /* XXX spin */
1008 crit_enter_gd(mygd);
1009 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
1011 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1017 * Generic deschedule. Descheduling threads other then your own should be
1018 * done only in carefully controlled circumstances. Descheduling is
1021 * This function may block if the cpu has run out of messages.
1024 lwkt_deschedule(thread_t td)
1027 if (td == curthread) {
1030 if (td->td_gd == mycpu) {
1033 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_deschedule, td);
1040 * Set the target thread's priority. This routine does not automatically
1041 * switch to a higher priority thread, LWKT threads are not designed for
1042 * continuous priority changes. Yield if you want to switch.
1044 * We have to retain the critical section count which uses the high bits
1045 * of the td_pri field. The specified priority may also indicate zero or
1046 * more critical sections by adding TDPRI_CRIT*N.
1048 * Note that we requeue the thread whether it winds up on a different runq
1049 * or not. uio_yield() depends on this and the routine is not normally
1050 * called with the same priority otherwise.
1053 lwkt_setpri(thread_t td, int pri)
1056 KKASSERT(td->td_gd == mycpu);
1058 if (td->td_flags & TDF_RUNQ) {
1060 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1063 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1069 lwkt_setpri_self(int pri)
1071 thread_t td = curthread;
1073 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1075 if (td->td_flags & TDF_RUNQ) {
1077 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1080 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1086 * Determine if there is a runnable thread at a higher priority then
1087 * the current thread. lwkt_setpri() does not check this automatically.
1088 * Return 1 if there is, 0 if there isn't.
1090 * Example: if bit 31 of runqmask is set and the current thread is priority
1091 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1093 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1094 * up comparing against 0xffffffff, a comparison that will always be false.
1097 lwkt_checkpri_self(void)
1099 globaldata_t gd = mycpu;
1100 thread_t td = gd->gd_curthread;
1101 int nq = td->td_pri & TDPRI_MASK;
1103 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1104 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1112 * Migrate the current thread to the specified cpu. The BGL must be held
1113 * (for the gd_tdallq manipulation XXX). This is accomplished by
1114 * descheduling ourselves from the current cpu, moving our thread to the
1115 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1116 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1119 static void lwkt_setcpu_remote(void *arg);
1123 lwkt_setcpu_self(globaldata_t rgd)
1126 thread_t td = curthread;
1128 if (td->td_gd != rgd) {
1129 crit_enter_quick(td);
1130 td->td_flags |= TDF_MIGRATING;
1131 lwkt_deschedule_self(td);
1132 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1133 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1134 lwkt_send_ipiq(rgd, (ipifunc_t)lwkt_setcpu_remote, td);
1136 /* we are now on the target cpu */
1137 crit_exit_quick(td);
1143 * Remote IPI for cpu migration (called while in a critical section so we
1144 * do not have to enter another one). The thread has already been moved to
1145 * our cpu's allq, but we must wait for the thread to be completely switched
1146 * out on the originating cpu before we schedule it on ours or the stack
1147 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1148 * change to main memory.
1150 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1151 * against wakeups. It is best if this interface is used only when there
1152 * are no pending events that might try to schedule the thread.
1156 lwkt_setcpu_remote(void *arg)
1159 globaldata_t gd = mycpu;
1161 while (td->td_flags & TDF_RUNNING)
1165 td->td_flags &= ~TDF_MIGRATING;
1171 lwkt_preempted_proc(void)
1173 thread_t td = curthread;
1174 while (td->td_preempted)
1175 td = td->td_preempted;
1176 return(td->td_proc);
1180 * Block on the specified wait queue until signaled. A generation number
1181 * must be supplied to interlock the wait queue. The function will
1182 * return immediately if the generation number does not match the wait
1183 * structure's generation number.
1186 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1188 thread_t td = curthread;
1191 lwkt_gettoken(&ilock, &w->wa_token);
1193 if (w->wa_gen == *gen) {
1195 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1198 td->td_wmesg = wmesg;
1201 if (td->td_wmesg != NULL) {
1208 lwkt_reltoken(&ilock);
1212 * Signal a wait queue. We gain ownership of the wait queue in order to
1213 * signal it. Once a thread is removed from the wait queue we have to
1214 * deal with the cpu owning the thread.
1216 * Note: alternatively we could message the target cpu owning the wait
1217 * queue. YYY implement as sysctl.
1220 lwkt_signal(lwkt_wait_t w, int count)
1225 lwkt_gettoken(&ilock, &w->wa_token);
1229 count = w->wa_count;
1230 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1233 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1235 td->td_wmesg = NULL;
1236 if (td->td_gd == mycpu) {
1239 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_schedule, td);
1243 lwkt_reltoken(&ilock);
1247 * Create a kernel process/thread/whatever. It shares it's address space
1248 * with proc0 - ie: kernel only.
1250 * NOTE! By default new threads are created with the MP lock held. A
1251 * thread which does not require the MP lock should release it by calling
1252 * rel_mplock() at the start of the new thread.
1255 lwkt_create(void (*func)(void *), void *arg,
1256 struct thread **tdp, thread_t template, int tdflags, int cpu,
1257 const char *fmt, ...)
1262 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu);
1265 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1266 td->td_flags |= TDF_VERBOSE | tdflags;
1272 * Set up arg0 for 'ps' etc
1274 __va_start(ap, fmt);
1275 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1279 * Schedule the thread to run
1281 if ((td->td_flags & TDF_STOPREQ) == 0)
1284 td->td_flags &= ~TDF_STOPREQ;
1289 * kthread_* is specific to the kernel and is not needed by userland.
1294 * Destroy an LWKT thread. Warning! This function is not called when
1295 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1296 * uses a different reaping mechanism.
1301 thread_t td = curthread;
1304 if (td->td_flags & TDF_VERBOSE)
1305 printf("kthread %p %s has exited\n", td, td->td_comm);
1307 crit_enter_quick(td);
1308 lwkt_deschedule_self(td);
1310 KKASSERT(gd == td->td_gd);
1311 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1312 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1313 ++gd->gd_tdfreecount;
1314 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1319 #endif /* _KERNEL */
1324 thread_t td = curthread;
1325 int lpri = td->td_pri;
1328 panic("td_pri is/would-go negative! %p %d", td, lpri);