2 * Copyright (c) 2003-2010 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
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
58 #include <sys/mplock2.h>
60 #include <sys/dsched.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_extern.h>
71 #include <machine/stdarg.h>
72 #include <machine/smp.h>
74 #if !defined(KTR_CTXSW)
75 #define KTR_CTXSW KTR_ALL
77 KTR_INFO_MASTER(ctxsw);
78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
79 sizeof(int) + sizeof(struct thread *));
80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
81 sizeof(int) + sizeof(struct thread *));
82 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
83 sizeof (struct thread *) + sizeof(char *));
84 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
86 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
89 static int panic_on_cscount = 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;
95 static __int64_t token_contention_count __debugvar = 0;
96 static int lwkt_use_spin_port;
97 static struct objcache *thread_cache;
100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
103 extern void cpu_heavy_restore(void);
104 extern void cpu_lwkt_restore(void);
105 extern void cpu_kthread_restore(void);
106 extern void cpu_idle_restore(void);
111 jg_tos_ok(struct thread *td)
119 KKASSERT(td->td_sp != NULL);
120 tos = ((void **)td->td_sp)[0];
122 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
123 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
132 * We can make all thread ports use the spin backend instead of the thread
133 * backend. This should only be set to debug the spin backend.
135 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
140 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
142 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
146 &token_contention_count, 0, "spinning due to token contention");
150 * These helper procedures handle the runq, they can only be called from
151 * within a critical section.
153 * WARNING! Prior to SMP being brought up it is possible to enqueue and
154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
155 * instead of 'mycpu' when referencing the globaldata structure. Once
156 * SMP live enqueuing and dequeueing only occurs on the current cpu.
160 _lwkt_dequeue(thread_t td)
162 if (td->td_flags & TDF_RUNQ) {
163 int nq = td->td_pri & TDPRI_MASK;
164 struct globaldata *gd = td->td_gd;
166 td->td_flags &= ~TDF_RUNQ;
167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
168 /* runqmask is passively cleaned up by the switcher */
174 _lwkt_enqueue(thread_t td)
176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
177 int nq = td->td_pri & TDPRI_MASK;
178 struct globaldata *gd = td->td_gd;
180 td->td_flags |= TDF_RUNQ;
181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
182 gd->gd_runqmask |= 1 << nq;
187 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
189 struct thread *td = (struct thread *)obj;
191 td->td_kstack = NULL;
192 td->td_kstack_size = 0;
193 td->td_flags = TDF_ALLOCATED_THREAD;
198 _lwkt_thread_dtor(void *obj, void *privdata)
200 struct thread *td = (struct thread *)obj;
202 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
203 ("_lwkt_thread_dtor: not allocated from objcache"));
204 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
205 td->td_kstack_size > 0,
206 ("_lwkt_thread_dtor: corrupted stack"));
207 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
211 * Initialize the lwkt s/system.
216 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
217 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
218 NULL, CACHE_NTHREADS/2,
219 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
223 * Schedule a thread to run. As the current thread we can always safely
224 * schedule ourselves, and a shortcut procedure is provided for that
227 * (non-blocking, self contained on a per cpu basis)
230 lwkt_schedule_self(thread_t td)
232 crit_enter_quick(td);
233 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
234 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
240 * Deschedule a thread.
242 * (non-blocking, self contained on a per cpu basis)
245 lwkt_deschedule_self(thread_t td)
247 crit_enter_quick(td);
253 * LWKTs operate on a per-cpu basis
255 * WARNING! Called from early boot, 'mycpu' may not work yet.
258 lwkt_gdinit(struct globaldata *gd)
262 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
263 TAILQ_INIT(&gd->gd_tdrunq[i]);
265 TAILQ_INIT(&gd->gd_tdallq);
269 * Create a new thread. The thread must be associated with a process context
270 * or LWKT start address before it can be scheduled. If the target cpu is
271 * -1 the thread will be created on the current cpu.
273 * If you intend to create a thread without a process context this function
274 * does everything except load the startup and switcher function.
277 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
279 globaldata_t gd = mycpu;
283 * If static thread storage is not supplied allocate a thread. Reuse
284 * a cached free thread if possible. gd_freetd is used to keep an exiting
285 * thread intact through the exit.
288 if ((td = gd->gd_freetd) != NULL)
289 gd->gd_freetd = NULL;
291 td = objcache_get(thread_cache, M_WAITOK);
292 KASSERT((td->td_flags &
293 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
294 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
295 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
299 * Try to reuse cached stack.
301 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
302 if (flags & TDF_ALLOCATED_STACK) {
303 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
308 stack = (void *)kmem_alloc(&kernel_map, stksize);
309 flags |= TDF_ALLOCATED_STACK;
312 lwkt_init_thread(td, stack, stksize, flags, gd);
314 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
319 * Initialize a preexisting thread structure. This function is used by
320 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
322 * All threads start out in a critical section at a priority of
323 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
324 * appropriate. This function may send an IPI message when the
325 * requested cpu is not the current cpu and consequently gd_tdallq may
326 * not be initialized synchronously from the point of view of the originating
329 * NOTE! we have to be careful in regards to creating threads for other cpus
330 * if SMP has not yet been activated.
335 lwkt_init_thread_remote(void *arg)
340 * Protected by critical section held by IPI dispatch
342 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
348 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
349 struct globaldata *gd)
351 globaldata_t mygd = mycpu;
353 bzero(td, sizeof(struct thread));
354 td->td_kstack = stack;
355 td->td_kstack_size = stksize;
356 td->td_flags = flags;
358 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
359 td->td_toks_stop = &td->td_toks_base;
361 if ((flags & TDF_MPSAFE) == 0)
364 if (lwkt_use_spin_port)
365 lwkt_initport_spin(&td->td_msgport);
367 lwkt_initport_thread(&td->td_msgport, td);
368 pmap_init_thread(td);
371 * Normally initializing a thread for a remote cpu requires sending an
372 * IPI. However, the idlethread is setup before the other cpus are
373 * activated so we have to treat it as a special case. XXX manipulation
374 * of gd_tdallq requires the BGL.
376 if (gd == mygd || td == &gd->gd_idlethread) {
378 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
381 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
385 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
389 dsched_new_thread(td);
393 lwkt_set_comm(thread_t td, const char *ctl, ...)
398 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
400 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
404 lwkt_hold(thread_t td)
410 lwkt_rele(thread_t td)
412 KKASSERT(td->td_refs > 0);
417 lwkt_wait_free(thread_t td)
420 tsleep(td, 0, "tdreap", hz);
424 lwkt_free_thread(thread_t td)
426 KASSERT((td->td_flags & TDF_RUNNING) == 0,
427 ("lwkt_free_thread: did not exit! %p", td));
429 if (td->td_flags & TDF_ALLOCATED_THREAD) {
430 objcache_put(thread_cache, td);
431 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
432 /* client-allocated struct with internally allocated stack */
433 KASSERT(td->td_kstack && td->td_kstack_size > 0,
434 ("lwkt_free_thread: corrupted stack"));
435 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
436 td->td_kstack = NULL;
437 td->td_kstack_size = 0;
439 KTR_LOG(ctxsw_deadtd, td);
444 * Switch to the next runnable lwkt. If no LWKTs are runnable then
445 * switch to the idlethread. Switching must occur within a critical
446 * section to avoid races with the scheduling queue.
448 * We always have full control over our cpu's run queue. Other cpus
449 * that wish to manipulate our queue must use the cpu_*msg() calls to
450 * talk to our cpu, so a critical section is all that is needed and
451 * the result is very, very fast thread switching.
453 * The LWKT scheduler uses a fixed priority model and round-robins at
454 * each priority level. User process scheduling is a totally
455 * different beast and LWKT priorities should not be confused with
456 * user process priorities.
458 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
459 * cleans it up. Note that the td_switch() function cannot do anything that
460 * requires the MP lock since the MP lock will have already been setup for
461 * the target thread (not the current thread). It's nice to have a scheduler
462 * that does not need the MP lock to work because it allows us to do some
463 * really cool high-performance MP lock optimizations.
465 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
466 * is not called by the current thread in the preemption case, only when
467 * the preempting thread blocks (in order to return to the original thread).
472 globaldata_t gd = mycpu;
473 thread_t td = gd->gd_curthread;
480 * Switching from within a 'fast' (non thread switched) interrupt or IPI
481 * is illegal. However, we may have to do it anyway if we hit a fatal
482 * kernel trap or we have paniced.
484 * If this case occurs save and restore the interrupt nesting level.
486 if (gd->gd_intr_nesting_level) {
490 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
491 panic("lwkt_switch: cannot switch from within "
492 "a fast interrupt, yet, td %p\n", td);
494 savegdnest = gd->gd_intr_nesting_level;
495 savegdtrap = gd->gd_trap_nesting_level;
496 gd->gd_intr_nesting_level = 0;
497 gd->gd_trap_nesting_level = 0;
498 if ((td->td_flags & TDF_PANICWARN) == 0) {
499 td->td_flags |= TDF_PANICWARN;
500 kprintf("Warning: thread switch from interrupt or IPI, "
501 "thread %p (%s)\n", td, td->td_comm);
505 gd->gd_intr_nesting_level = savegdnest;
506 gd->gd_trap_nesting_level = savegdtrap;
512 * Passive release (used to transition from user to kernel mode
513 * when we block or switch rather then when we enter the kernel).
514 * This function is NOT called if we are switching into a preemption
515 * or returning from a preemption. Typically this causes us to lose
516 * our current process designation (if we have one) and become a true
517 * LWKT thread, and may also hand the current process designation to
518 * another process and schedule thread.
524 if (TD_TOKS_HELD(td))
525 lwkt_relalltokens(td);
528 * We had better not be holding any spin locks, but don't get into an
529 * endless panic loop.
531 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
532 ("lwkt_switch: still holding a shared spinlock %p!",
533 gd->gd_spinlock_rd));
534 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
535 ("lwkt_switch: still holding %d exclusive spinlocks!",
536 gd->gd_spinlocks_wr));
541 * td_mpcount cannot be used to determine if we currently hold the
542 * MP lock because get_mplock() will increment it prior to attempting
543 * to get the lock, and switch out if it can't. Our ownership of
544 * the actual lock will remain stable while we are in a critical section
545 * (but, of course, another cpu may own or release the lock so the
546 * actual value of mp_lock is not stable).
548 mpheld = MP_LOCK_HELD();
550 if (td->td_cscount) {
551 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
553 if (panic_on_cscount)
554 panic("switching while mastering cpusync");
558 if ((ntd = td->td_preempted) != NULL) {
560 * We had preempted another thread on this cpu, resume the preempted
561 * thread. This occurs transparently, whether the preempted thread
562 * was scheduled or not (it may have been preempted after descheduling
565 * We have to setup the MP lock for the original thread after backing
566 * out the adjustment that was made to curthread when the original
569 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
571 if (ntd->td_mpcount && mpheld == 0) {
572 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
573 td, ntd, td->td_mpcount, ntd->td_mpcount);
575 if (ntd->td_mpcount) {
576 td->td_mpcount -= ntd->td_mpcount;
577 KKASSERT(td->td_mpcount >= 0);
580 ntd->td_flags |= TDF_PREEMPT_DONE;
583 * The interrupt may have woken a thread up, we need to properly
584 * set the reschedule flag if the originally interrupted thread is
585 * at a lower priority.
587 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
589 /* YYY release mp lock on switchback if original doesn't need it */
592 * Priority queue / round-robin at each priority. Note that user
593 * processes run at a fixed, low priority and the user process
594 * scheduler deals with interactions between user processes
595 * by scheduling and descheduling them from the LWKT queue as
598 * We have to adjust the MP lock for the target thread. If we
599 * need the MP lock and cannot obtain it we try to locate a
600 * thread that does not need the MP lock. If we cannot, we spin
603 * A similar issue exists for the tokens held by the target thread.
604 * If we cannot obtain ownership of the tokens we cannot immediately
605 * schedule the thread.
609 * If an LWKT reschedule was requested, well that is what we are
610 * doing now so clear it.
612 clear_lwkt_resched();
614 if (gd->gd_runqmask) {
615 int nq = bsrl(gd->gd_runqmask);
616 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
617 gd->gd_runqmask &= ~(1 << nq);
622 * THREAD SELECTION FOR AN SMP MACHINE BUILD
624 * If the target needs the MP lock and we couldn't get it,
625 * or if the target is holding tokens and we could not
626 * gain ownership of the tokens, continue looking for a
627 * thread to schedule and spin instead of HLT if we can't.
629 * NOTE: the mpheld variable invalid after this conditional, it
630 * can change due to both cpu_try_mplock() returning success
631 * AND interactions in lwkt_getalltokens() due to the fact that
632 * we are trying to check the mpcount of a thread other then
633 * the current thread. Because of this, if the current thread
634 * is not holding td_mpcount, an IPI indirectly run via
635 * lwkt_getalltokens() can obtain and release the MP lock and
636 * cause the core MP lock to be released.
638 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
639 (TD_TOKS_HELD(ntd) && lwkt_getalltokens(ntd) == 0)
641 u_int32_t rqmask = gd->gd_runqmask;
645 mpheld = MP_LOCK_HELD();
648 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
649 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
650 /* spinning due to MP lock being held */
655 * mpheld state invalid after getalltokens call returns
656 * failure, but the variable is only needed for
659 if (TD_TOKS_HELD(ntd) && !lwkt_getalltokens(ntd)) {
660 /* spinning due to token contention */
662 ++token_contention_count;
664 mpheld = MP_LOCK_HELD();
671 rqmask &= ~(1 << nq);
675 * We have two choices. We can either refuse to run a
676 * user thread when a kernel thread needs the MP lock
677 * but could not get it, or we can allow it to run but
678 * then expect an IPI (hopefully) later on to force a
679 * reschedule when the MP lock might become available.
681 if (nq < TDPRI_KERN_LPSCHED) {
682 break; /* for now refuse to run */
684 if (chain_mplock == 0)
686 /* continue loop, allow user threads to be scheduled */
692 * Case where a (kernel) thread needed the MP lock and could
693 * not get one, and we may or may not have found another
694 * thread which does not need the MP lock to run while
698 ntd = &gd->gd_idlethread;
699 ntd->td_flags |= TDF_IDLE_NOHLT;
700 set_mplock_contention_mask(gd);
701 cpu_mplock_contested();
702 goto using_idle_thread;
704 clr_mplock_contention_mask(gd);
705 ++gd->gd_cnt.v_swtch;
706 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
707 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 clr_mplock_contention_mask(gd);
711 ++gd->gd_cnt.v_swtch;
712 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
713 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
717 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
718 * worry about tokens or the BGL. However, we still have
719 * to call lwkt_getalltokens() in order to properly detect
720 * stale tokens. This call cannot fail for a UP build!
722 lwkt_getalltokens(ntd);
723 ++gd->gd_cnt.v_swtch;
724 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
725 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
729 * We have nothing to run but only let the idle loop halt
730 * the cpu if there are no pending interrupts.
732 ntd = &gd->gd_idlethread;
733 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
734 ntd->td_flags |= TDF_IDLE_NOHLT;
738 * The idle thread should not be holding the MP lock unless we
739 * are trapping in the kernel or in a panic. Since we select the
740 * idle thread unconditionally when no other thread is available,
741 * if the MP lock is desired during a panic or kernel trap, we
742 * have to loop in the scheduler until we get it.
744 if (ntd->td_mpcount) {
745 mpheld = MP_LOCK_HELD();
746 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
747 panic("Idle thread %p was holding the BGL!", ntd);
754 KASSERT(ntd->td_pri >= TDPRI_CRIT,
755 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
758 * Do the actual switch. If the new target does not need the MP lock
759 * and we are holding it, release the MP lock. If the new target requires
760 * the MP lock we have already acquired it for the target.
763 if (ntd->td_mpcount == 0 ) {
767 ASSERT_MP_LOCK_HELD(ntd);
774 int tos_ok __debugvar = jg_tos_ok(ntd);
778 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
781 /* NOTE: current cpu may have changed after switch */
786 * Request that the target thread preempt the current thread. Preemption
787 * only works under a specific set of conditions:
789 * - We are not preempting ourselves
790 * - The target thread is owned by the current cpu
791 * - We are not currently being preempted
792 * - The target is not currently being preempted
793 * - We are not holding any spin locks
794 * - The target thread is not holding any tokens
795 * - We are able to satisfy the target's MP lock requirements (if any).
797 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
798 * this is called via lwkt_schedule() through the td_preemptable callback.
799 * critpri is the managed critical priority that we should ignore in order
800 * to determine whether preemption is possible (aka usually just the crit
801 * priority of lwkt_schedule() itself).
803 * XXX at the moment we run the target thread in a critical section during
804 * the preemption in order to prevent the target from taking interrupts
805 * that *WE* can't. Preemption is strictly limited to interrupt threads
806 * and interrupt-like threads, outside of a critical section, and the
807 * preempted source thread will be resumed the instant the target blocks
808 * whether or not the source is scheduled (i.e. preemption is supposed to
809 * be as transparent as possible).
811 * The target thread inherits our MP count (added to its own) for the
812 * duration of the preemption in order to preserve the atomicy of the
813 * MP lock during the preemption. Therefore, any preempting targets must be
814 * careful in regards to MP assertions. Note that the MP count may be
815 * out of sync with the physical mp_lock, but we do not have to preserve
816 * the original ownership of the lock if it was out of synch (that is, we
817 * can leave it synchronized on return).
820 lwkt_preempt(thread_t ntd, int critpri)
822 struct globaldata *gd = mycpu;
830 * The caller has put us in a critical section. We can only preempt
831 * if the caller of the caller was not in a critical section (basically
832 * a local interrupt), as determined by the 'critpri' parameter. We
833 * also can't preempt if the caller is holding any spinlocks (even if
834 * he isn't in a critical section). This also handles the tokens test.
836 * YYY The target thread must be in a critical section (else it must
837 * inherit our critical section? I dunno yet).
839 * Set need_lwkt_resched() unconditionally for now YYY.
841 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
843 td = gd->gd_curthread;
844 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
848 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
854 if (ntd->td_gd != gd) {
861 * Take the easy way out and do not preempt if we are holding
862 * any spinlocks. We could test whether the thread(s) being
863 * preempted interlock against the target thread's tokens and whether
864 * we can get all the target thread's tokens, but this situation
865 * should not occur very often so its easier to simply not preempt.
866 * Also, plain spinlocks are impossible to figure out at this point so
867 * just don't preempt.
869 * Do not try to preempt if the target thread is holding any tokens.
870 * We could try to acquire the tokens but this case is so rare there
871 * is no need to support it.
873 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
878 if (TD_TOKS_HELD(ntd)) {
883 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
888 if (ntd->td_preempted) {
895 * note: an interrupt might have occured just as we were transitioning
896 * to or from the MP lock. In this case td_mpcount will be pre-disposed
897 * (non-zero) but not actually synchronized with the actual state of the
898 * lock. We can use it to imply an MP lock requirement for the
899 * preemption but we cannot use it to test whether we hold the MP lock
902 savecnt = td->td_mpcount;
903 mpheld = MP_LOCK_HELD();
904 ntd->td_mpcount += td->td_mpcount;
905 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
906 ntd->td_mpcount -= td->td_mpcount;
914 * Since we are able to preempt the current thread, there is no need to
915 * call need_lwkt_resched().
918 ntd->td_preempted = td;
919 td->td_flags |= TDF_PREEMPT_LOCK;
920 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
923 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
925 KKASSERT(savecnt == td->td_mpcount);
926 mpheld = MP_LOCK_HELD();
927 if (mpheld && td->td_mpcount == 0)
929 else if (mpheld == 0 && td->td_mpcount)
930 panic("lwkt_preempt(): MP lock was not held through");
932 ntd->td_preempted = NULL;
933 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
937 * Conditionally call splz() if gd_reqflags indicates work is pending.
939 * td_nest_count prevents deep nesting via splz() or doreti() which
940 * might otherwise blow out the kernel stack. Note that except for
941 * this special case, we MUST call splz() here to handle any
942 * pending ints, particularly after we switch, or we might accidently
943 * halt the cpu with interrupts pending.
945 * (self contained on a per cpu basis)
950 globaldata_t gd = mycpu;
951 thread_t td = gd->gd_curthread;
953 if (gd->gd_reqflags && td->td_nest_count < 2)
958 * This implements a normal yield which will yield to equal priority
959 * threads as well as higher priority threads. Note that gd_reqflags
960 * tests will be handled by the crit_exit() call in lwkt_switch().
962 * (self contained on a per cpu basis)
967 lwkt_schedule_self(curthread);
972 * This function is used along with the lwkt_passive_recover() inline
973 * by the trap code to negotiate a passive release of the current
974 * process/lwp designation with the user scheduler.
977 lwkt_passive_release(struct thread *td)
979 struct lwp *lp = td->td_lwp;
981 td->td_release = NULL;
982 lwkt_setpri_self(TDPRI_KERN_USER);
983 lp->lwp_proc->p_usched->release_curproc(lp);
987 * Make a kernel thread act as if it were in user mode with regards
988 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
989 * loops which may be potentially cpu-bound can call lwkt_user_yield().
991 * The lwkt_user_yield() function is designed to have very low overhead
992 * if no yield is determined to be needed.
995 lwkt_user_yield(void)
997 thread_t td = curthread;
998 struct lwp *lp = td->td_lwp;
1002 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1003 * kernel can prevent other cpus from servicing interrupt threads
1004 * which still require the MP lock (which is a lot of them). This
1005 * has a chaining effect since if the interrupt is blocked, so is
1006 * the event, so normal scheduling will not pick up on the problem.
1008 if (mp_lock_contention_mask && td->td_mpcount) {
1014 * Another kernel thread wants the cpu
1016 if (lwkt_resched_wanted())
1020 * If the user scheduler has asynchronously determined that the current
1021 * process (when running in user mode) needs to lose the cpu then make
1022 * sure we are released.
1024 if (user_resched_wanted()) {
1030 * If we are released reduce our priority
1032 if (td->td_release == NULL) {
1033 if (lwkt_check_resched(td) > 0)
1036 lp->lwp_proc->p_usched->acquire_curproc(lp);
1037 td->td_release = lwkt_passive_release;
1038 lwkt_setpri_self(TDPRI_USER_NORM);
1044 * Return 0 if no runnable threads are pending at the same or higher
1045 * priority as the passed thread.
1047 * Return 1 if runnable threads are pending at the same priority.
1049 * Return 2 if runnable threads are pending at a higher priority.
1052 lwkt_check_resched(thread_t td)
1054 int pri = td->td_pri & TDPRI_MASK;
1056 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1058 if (TAILQ_NEXT(td, td_threadq))
1064 * Generic schedule. Possibly schedule threads belonging to other cpus and
1065 * deal with threads that might be blocked on a wait queue.
1067 * We have a little helper inline function which does additional work after
1068 * the thread has been enqueued, including dealing with preemption and
1069 * setting need_lwkt_resched() (which prevents the kernel from returning
1070 * to userland until it has processed higher priority threads).
1072 * It is possible for this routine to be called after a failed _enqueue
1073 * (due to the target thread migrating, sleeping, or otherwise blocked).
1074 * We have to check that the thread is actually on the run queue!
1076 * reschedok is an optimized constant propagated from lwkt_schedule() or
1077 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1078 * reschedule to be requested if the target thread has a higher priority.
1079 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1080 * be 0, prevented undesired reschedules.
1084 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1088 if (ntd->td_flags & TDF_RUNQ) {
1089 if (ntd->td_preemptable && reschedok) {
1090 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1091 } else if (reschedok) {
1093 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1094 need_lwkt_resched();
1101 _lwkt_schedule(thread_t td, int reschedok)
1103 globaldata_t mygd = mycpu;
1105 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1106 crit_enter_gd(mygd);
1107 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1108 if (td == mygd->gd_curthread) {
1112 * If we own the thread, there is no race (since we are in a
1113 * critical section). If we do not own the thread there might
1114 * be a race but the target cpu will deal with it.
1117 if (td->td_gd == mygd) {
1119 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1121 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1125 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1132 lwkt_schedule(thread_t td)
1134 _lwkt_schedule(td, 1);
1138 lwkt_schedule_noresched(thread_t td)
1140 _lwkt_schedule(td, 0);
1146 * When scheduled remotely if frame != NULL the IPIQ is being
1147 * run via doreti or an interrupt then preemption can be allowed.
1149 * To allow preemption we have to drop the critical section so only
1150 * one is present in _lwkt_schedule_post.
1153 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1155 thread_t td = curthread;
1158 if (frame && ntd->td_preemptable) {
1159 crit_exit_noyield(td);
1160 _lwkt_schedule(ntd, 1);
1161 crit_enter_quick(td);
1163 _lwkt_schedule(ntd, 1);
1168 * Thread migration using a 'Pull' method. The thread may or may not be
1169 * the current thread. It MUST be descheduled and in a stable state.
1170 * lwkt_giveaway() must be called on the cpu owning the thread.
1172 * At any point after lwkt_giveaway() is called, the target cpu may
1173 * 'pull' the thread by calling lwkt_acquire().
1175 * We have to make sure the thread is not sitting on a per-cpu tsleep
1176 * queue or it will blow up when it moves to another cpu.
1178 * MPSAFE - must be called under very specific conditions.
1181 lwkt_giveaway(thread_t td)
1183 globaldata_t gd = mycpu;
1186 if (td->td_flags & TDF_TSLEEPQ)
1188 KKASSERT(td->td_gd == gd);
1189 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1190 td->td_flags |= TDF_MIGRATING;
1195 lwkt_acquire(thread_t td)
1200 KKASSERT(td->td_flags & TDF_MIGRATING);
1205 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1206 crit_enter_gd(mygd);
1207 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1209 lwkt_process_ipiq();
1214 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1215 td->td_flags &= ~TDF_MIGRATING;
1218 crit_enter_gd(mygd);
1219 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1220 td->td_flags &= ~TDF_MIGRATING;
1228 * Generic deschedule. Descheduling threads other then your own should be
1229 * done only in carefully controlled circumstances. Descheduling is
1232 * This function may block if the cpu has run out of messages.
1235 lwkt_deschedule(thread_t td)
1239 if (td == curthread) {
1242 if (td->td_gd == mycpu) {
1245 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1255 * Set the target thread's priority. This routine does not automatically
1256 * switch to a higher priority thread, LWKT threads are not designed for
1257 * continuous priority changes. Yield if you want to switch.
1259 * We have to retain the critical section count which uses the high bits
1260 * of the td_pri field. The specified priority may also indicate zero or
1261 * more critical sections by adding TDPRI_CRIT*N.
1263 * Note that we requeue the thread whether it winds up on a different runq
1264 * or not. uio_yield() depends on this and the routine is not normally
1265 * called with the same priority otherwise.
1268 lwkt_setpri(thread_t td, int pri)
1271 KKASSERT(td->td_gd == mycpu);
1273 if (td->td_flags & TDF_RUNQ) {
1275 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1278 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1284 * Set the initial priority for a thread prior to it being scheduled for
1285 * the first time. The thread MUST NOT be scheduled before or during
1286 * this call. The thread may be assigned to a cpu other then the current
1289 * Typically used after a thread has been created with TDF_STOPPREQ,
1290 * and before the thread is initially scheduled.
1293 lwkt_setpri_initial(thread_t td, int pri)
1296 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1297 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1301 lwkt_setpri_self(int pri)
1303 thread_t td = curthread;
1305 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1307 if (td->td_flags & TDF_RUNQ) {
1309 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1312 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1318 * Migrate the current thread to the specified cpu.
1320 * This is accomplished by descheduling ourselves from the current cpu,
1321 * moving our thread to the tdallq of the target cpu, IPI messaging the
1322 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1323 * races while the thread is being migrated.
1325 * We must be sure to remove ourselves from the current cpu's tsleepq
1326 * before potentially moving to another queue. The thread can be on
1327 * a tsleepq due to a left-over tsleep_interlock().
1330 static void lwkt_setcpu_remote(void *arg);
1334 lwkt_setcpu_self(globaldata_t rgd)
1337 thread_t td = curthread;
1339 if (td->td_gd != rgd) {
1340 crit_enter_quick(td);
1341 if (td->td_flags & TDF_TSLEEPQ)
1343 td->td_flags |= TDF_MIGRATING;
1344 lwkt_deschedule_self(td);
1345 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1346 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1348 /* we are now on the target cpu */
1349 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1350 crit_exit_quick(td);
1356 lwkt_migratecpu(int cpuid)
1361 rgd = globaldata_find(cpuid);
1362 lwkt_setcpu_self(rgd);
1367 * Remote IPI for cpu migration (called while in a critical section so we
1368 * do not have to enter another one). The thread has already been moved to
1369 * our cpu's allq, but we must wait for the thread to be completely switched
1370 * out on the originating cpu before we schedule it on ours or the stack
1371 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1372 * change to main memory.
1374 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1375 * against wakeups. It is best if this interface is used only when there
1376 * are no pending events that might try to schedule the thread.
1380 lwkt_setcpu_remote(void *arg)
1383 globaldata_t gd = mycpu;
1385 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1387 lwkt_process_ipiq();
1393 td->td_flags &= ~TDF_MIGRATING;
1394 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1400 lwkt_preempted_proc(void)
1402 thread_t td = curthread;
1403 while (td->td_preempted)
1404 td = td->td_preempted;
1409 * Create a kernel process/thread/whatever. It shares it's address space
1410 * with proc0 - ie: kernel only.
1412 * NOTE! By default new threads are created with the MP lock held. A
1413 * thread which does not require the MP lock should release it by calling
1414 * rel_mplock() at the start of the new thread.
1417 lwkt_create(void (*func)(void *), void *arg,
1418 struct thread **tdp, thread_t template, int tdflags, int cpu,
1419 const char *fmt, ...)
1424 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1428 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1431 * Set up arg0 for 'ps' etc
1433 __va_start(ap, fmt);
1434 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1438 * Schedule the thread to run
1440 if ((td->td_flags & TDF_STOPREQ) == 0)
1443 td->td_flags &= ~TDF_STOPREQ;
1448 * Destroy an LWKT thread. Warning! This function is not called when
1449 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1450 * uses a different reaping mechanism.
1455 thread_t td = curthread;
1459 if (td->td_flags & TDF_VERBOSE)
1460 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1464 * Get us into a critical section to interlock gd_freetd and loop
1465 * until we can get it freed.
1467 * We have to cache the current td in gd_freetd because objcache_put()ing
1468 * it would rip it out from under us while our thread is still active.
1471 crit_enter_quick(td);
1472 while ((std = gd->gd_freetd) != NULL) {
1473 gd->gd_freetd = NULL;
1474 objcache_put(thread_cache, std);
1478 * Remove thread resources from kernel lists and deschedule us for
1481 if (td->td_flags & TDF_TSLEEPQ)
1484 dsched_exit_thread(td);
1485 lwkt_deschedule_self(td);
1486 lwkt_remove_tdallq(td);
1487 if (td->td_flags & TDF_ALLOCATED_THREAD)
1493 lwkt_remove_tdallq(thread_t td)
1495 KKASSERT(td->td_gd == mycpu);
1496 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1502 thread_t td = curthread;
1503 int lpri = td->td_pri;
1506 panic("td_pri is/would-go negative! %p %d", td, lpri);
1512 * Called from debugger/panic on cpus which have been stopped. We must still
1513 * process the IPIQ while stopped, even if we were stopped while in a critical
1516 * If we are dumping also try to process any pending interrupts. This may
1517 * or may not work depending on the state of the cpu at the point it was
1521 lwkt_smp_stopped(void)
1523 globaldata_t gd = mycpu;
1527 lwkt_process_ipiq();
1530 lwkt_process_ipiq();