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
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, "sw %p > %p", 2 * sizeof(struct thread *));
79 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
80 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "new_td %p %s", sizeof (struct thread *) +
82 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "dead_td %p", sizeof (struct thread *));
84 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
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;
93 static __int64_t token_contention_count __debugvar = 0;
94 static int lwkt_use_spin_port;
95 static struct objcache *thread_cache;
98 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
101 extern void cpu_heavy_restore(void);
102 extern void cpu_lwkt_restore(void);
103 extern void cpu_kthread_restore(void);
104 extern void cpu_idle_restore(void);
109 jg_tos_ok(struct thread *td)
117 KKASSERT(td->td_sp != NULL);
118 tos = ((void **)td->td_sp)[0];
120 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
121 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
130 * We can make all thread ports use the spin backend instead of the thread
131 * backend. This should only be set to debug the spin backend.
133 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
136 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
138 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
139 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
140 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
144 &token_contention_count, 0, "spinning due to token contention");
148 * These helper procedures handle the runq, they can only be called from
149 * within a critical section.
151 * WARNING! Prior to SMP being brought up it is possible to enqueue and
152 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
153 * instead of 'mycpu' when referencing the globaldata structure. Once
154 * SMP live enqueuing and dequeueing only occurs on the current cpu.
158 _lwkt_dequeue(thread_t td)
160 if (td->td_flags & TDF_RUNQ) {
161 int nq = td->td_pri & TDPRI_MASK;
162 struct globaldata *gd = td->td_gd;
164 td->td_flags &= ~TDF_RUNQ;
165 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
166 /* runqmask is passively cleaned up by the switcher */
172 _lwkt_enqueue(thread_t td)
174 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
175 int nq = td->td_pri & TDPRI_MASK;
176 struct globaldata *gd = td->td_gd;
178 td->td_flags |= TDF_RUNQ;
179 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
180 gd->gd_runqmask |= 1 << nq;
185 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
187 struct thread *td = (struct thread *)obj;
189 td->td_kstack = NULL;
190 td->td_kstack_size = 0;
191 td->td_flags = TDF_ALLOCATED_THREAD;
196 _lwkt_thread_dtor(void *obj, void *privdata)
198 struct thread *td = (struct thread *)obj;
200 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
201 ("_lwkt_thread_dtor: not allocated from objcache"));
202 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
203 td->td_kstack_size > 0,
204 ("_lwkt_thread_dtor: corrupted stack"));
205 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
209 * Initialize the lwkt s/system.
214 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
215 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
216 NULL, CACHE_NTHREADS/2,
217 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
221 * Schedule a thread to run. As the current thread we can always safely
222 * schedule ourselves, and a shortcut procedure is provided for that
225 * (non-blocking, self contained on a per cpu basis)
228 lwkt_schedule_self(thread_t td)
230 crit_enter_quick(td);
231 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
232 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
238 * Deschedule a thread.
240 * (non-blocking, self contained on a per cpu basis)
243 lwkt_deschedule_self(thread_t td)
245 crit_enter_quick(td);
251 * LWKTs operate on a per-cpu basis
253 * WARNING! Called from early boot, 'mycpu' may not work yet.
256 lwkt_gdinit(struct globaldata *gd)
260 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
261 TAILQ_INIT(&gd->gd_tdrunq[i]);
263 TAILQ_INIT(&gd->gd_tdallq);
267 * Create a new thread. The thread must be associated with a process context
268 * or LWKT start address before it can be scheduled. If the target cpu is
269 * -1 the thread will be created on the current cpu.
271 * If you intend to create a thread without a process context this function
272 * does everything except load the startup and switcher function.
275 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
277 globaldata_t gd = mycpu;
281 * If static thread storage is not supplied allocate a thread. Reuse
282 * a cached free thread if possible. gd_freetd is used to keep an exiting
283 * thread intact through the exit.
286 if ((td = gd->gd_freetd) != NULL)
287 gd->gd_freetd = NULL;
289 td = objcache_get(thread_cache, M_WAITOK);
290 KASSERT((td->td_flags &
291 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
292 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
293 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
297 * Try to reuse cached stack.
299 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
300 if (flags & TDF_ALLOCATED_STACK) {
301 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
306 stack = (void *)kmem_alloc(&kernel_map, stksize);
307 flags |= TDF_ALLOCATED_STACK;
310 lwkt_init_thread(td, stack, stksize, flags, gd);
312 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
317 * Initialize a preexisting thread structure. This function is used by
318 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
320 * All threads start out in a critical section at a priority of
321 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
322 * appropriate. This function may send an IPI message when the
323 * requested cpu is not the current cpu and consequently gd_tdallq may
324 * not be initialized synchronously from the point of view of the originating
327 * NOTE! we have to be careful in regards to creating threads for other cpus
328 * if SMP has not yet been activated.
333 lwkt_init_thread_remote(void *arg)
338 * Protected by critical section held by IPI dispatch
340 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
346 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
347 struct globaldata *gd)
349 globaldata_t mygd = mycpu;
351 bzero(td, sizeof(struct thread));
352 td->td_kstack = stack;
353 td->td_kstack_size = stksize;
354 td->td_flags = flags;
356 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
358 if ((flags & TDF_MPSAFE) == 0)
361 if (lwkt_use_spin_port)
362 lwkt_initport_spin(&td->td_msgport);
364 lwkt_initport_thread(&td->td_msgport, td);
365 pmap_init_thread(td);
368 * Normally initializing a thread for a remote cpu requires sending an
369 * IPI. However, the idlethread is setup before the other cpus are
370 * activated so we have to treat it as a special case. XXX manipulation
371 * of gd_tdallq requires the BGL.
373 if (gd == mygd || td == &gd->gd_idlethread) {
375 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
378 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
382 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
386 dsched_new_thread(td);
390 lwkt_set_comm(thread_t td, const char *ctl, ...)
395 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
397 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
401 lwkt_hold(thread_t td)
407 lwkt_rele(thread_t td)
409 KKASSERT(td->td_refs > 0);
414 lwkt_wait_free(thread_t td)
417 tsleep(td, 0, "tdreap", hz);
421 lwkt_free_thread(thread_t td)
423 KASSERT((td->td_flags & TDF_RUNNING) == 0,
424 ("lwkt_free_thread: did not exit! %p", td));
426 if (td->td_flags & TDF_ALLOCATED_THREAD) {
427 objcache_put(thread_cache, td);
428 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
429 /* client-allocated struct with internally allocated stack */
430 KASSERT(td->td_kstack && td->td_kstack_size > 0,
431 ("lwkt_free_thread: corrupted stack"));
432 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
433 td->td_kstack = NULL;
434 td->td_kstack_size = 0;
436 KTR_LOG(ctxsw_deadtd, td);
441 * Switch to the next runnable lwkt. If no LWKTs are runnable then
442 * switch to the idlethread. Switching must occur within a critical
443 * section to avoid races with the scheduling queue.
445 * We always have full control over our cpu's run queue. Other cpus
446 * that wish to manipulate our queue must use the cpu_*msg() calls to
447 * talk to our cpu, so a critical section is all that is needed and
448 * the result is very, very fast thread switching.
450 * The LWKT scheduler uses a fixed priority model and round-robins at
451 * each priority level. User process scheduling is a totally
452 * different beast and LWKT priorities should not be confused with
453 * user process priorities.
455 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
456 * cleans it up. Note that the td_switch() function cannot do anything that
457 * requires the MP lock since the MP lock will have already been setup for
458 * the target thread (not the current thread). It's nice to have a scheduler
459 * that does not need the MP lock to work because it allows us to do some
460 * really cool high-performance MP lock optimizations.
462 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
463 * is not called by the current thread in the preemption case, only when
464 * the preempting thread blocks (in order to return to the original thread).
469 globaldata_t gd = mycpu;
470 thread_t td = gd->gd_curthread;
477 * Switching from within a 'fast' (non thread switched) interrupt or IPI
478 * is illegal. However, we may have to do it anyway if we hit a fatal
479 * kernel trap or we have paniced.
481 * If this case occurs save and restore the interrupt nesting level.
483 if (gd->gd_intr_nesting_level) {
487 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
488 panic("lwkt_switch: cannot switch from within "
489 "a fast interrupt, yet, td %p\n", td);
491 savegdnest = gd->gd_intr_nesting_level;
492 savegdtrap = gd->gd_trap_nesting_level;
493 gd->gd_intr_nesting_level = 0;
494 gd->gd_trap_nesting_level = 0;
495 if ((td->td_flags & TDF_PANICWARN) == 0) {
496 td->td_flags |= TDF_PANICWARN;
497 kprintf("Warning: thread switch from interrupt or IPI, "
498 "thread %p (%s)\n", td, td->td_comm);
502 gd->gd_intr_nesting_level = savegdnest;
503 gd->gd_trap_nesting_level = savegdtrap;
509 * Passive release (used to transition from user to kernel mode
510 * when we block or switch rather then when we enter the kernel).
511 * This function is NOT called if we are switching into a preemption
512 * or returning from a preemption. Typically this causes us to lose
513 * our current process designation (if we have one) and become a true
514 * LWKT thread, and may also hand the current process designation to
515 * another process and schedule thread.
522 lwkt_relalltokens(td);
525 * We had better not be holding any spin locks, but don't get into an
526 * endless panic loop.
528 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
529 ("lwkt_switch: still holding a shared spinlock %p!",
530 gd->gd_spinlock_rd));
531 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
532 ("lwkt_switch: still holding %d exclusive spinlocks!",
533 gd->gd_spinlocks_wr));
538 * td_mpcount cannot be used to determine if we currently hold the
539 * MP lock because get_mplock() will increment it prior to attempting
540 * to get the lock, and switch out if it can't. Our ownership of
541 * the actual lock will remain stable while we are in a critical section
542 * (but, of course, another cpu may own or release the lock so the
543 * actual value of mp_lock is not stable).
545 mpheld = MP_LOCK_HELD();
547 if (td->td_cscount) {
548 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
550 if (panic_on_cscount)
551 panic("switching while mastering cpusync");
555 if ((ntd = td->td_preempted) != NULL) {
557 * We had preempted another thread on this cpu, resume the preempted
558 * thread. This occurs transparently, whether the preempted thread
559 * was scheduled or not (it may have been preempted after descheduling
562 * We have to setup the MP lock for the original thread after backing
563 * out the adjustment that was made to curthread when the original
566 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
568 if (ntd->td_mpcount && mpheld == 0) {
569 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
570 td, ntd, td->td_mpcount, ntd->td_mpcount);
572 if (ntd->td_mpcount) {
573 td->td_mpcount -= ntd->td_mpcount;
574 KKASSERT(td->td_mpcount >= 0);
577 ntd->td_flags |= TDF_PREEMPT_DONE;
580 * The interrupt may have woken a thread up, we need to properly
581 * set the reschedule flag if the originally interrupted thread is
582 * at a lower priority.
584 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
586 /* YYY release mp lock on switchback if original doesn't need it */
589 * Priority queue / round-robin at each priority. Note that user
590 * processes run at a fixed, low priority and the user process
591 * scheduler deals with interactions between user processes
592 * by scheduling and descheduling them from the LWKT queue as
595 * We have to adjust the MP lock for the target thread. If we
596 * need the MP lock and cannot obtain it we try to locate a
597 * thread that does not need the MP lock. If we cannot, we spin
600 * A similar issue exists for the tokens held by the target thread.
601 * If we cannot obtain ownership of the tokens we cannot immediately
602 * schedule the thread.
606 * If an LWKT reschedule was requested, well that is what we are
607 * doing now so clear it.
609 clear_lwkt_resched();
611 if (gd->gd_runqmask) {
612 int nq = bsrl(gd->gd_runqmask);
613 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
614 gd->gd_runqmask &= ~(1 << nq);
619 * THREAD SELECTION FOR AN SMP MACHINE BUILD
621 * If the target needs the MP lock and we couldn't get it,
622 * or if the target is holding tokens and we could not
623 * gain ownership of the tokens, continue looking for a
624 * thread to schedule and spin instead of HLT if we can't.
626 * NOTE: the mpheld variable invalid after this conditional, it
627 * can change due to both cpu_try_mplock() returning success
628 * AND interactions in lwkt_getalltokens() due to the fact that
629 * we are trying to check the mpcount of a thread other then
630 * the current thread. Because of this, if the current thread
631 * is not holding td_mpcount, an IPI indirectly run via
632 * lwkt_getalltokens() can obtain and release the MP lock and
633 * cause the core MP lock to be released.
635 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
636 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
638 u_int32_t rqmask = gd->gd_runqmask;
640 mpheld = MP_LOCK_HELD();
643 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
644 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
645 /* spinning due to MP lock being held */
650 * mpheld state invalid after getalltokens call returns
651 * failure, but the variable is only needed for
654 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
655 /* spinning due to token contention */
657 ++token_contention_count;
659 mpheld = MP_LOCK_HELD();
666 rqmask &= ~(1 << nq);
670 * We have two choices. We can either refuse to run a
671 * user thread when a kernel thread needs the MP lock
672 * but could not get it, or we can allow it to run but
673 * then expect an IPI (hopefully) later on to force a
674 * reschedule when the MP lock might become available.
676 if (nq < TDPRI_KERN_LPSCHED) {
677 break; /* for now refuse to run */
679 if (chain_mplock == 0)
681 /* continue loop, allow user threads to be scheduled */
687 * Case where a (kernel) thread needed the MP lock and could
688 * not get one, and we may or may not have found another
689 * thread which does not need the MP lock to run while
693 ntd = &gd->gd_idlethread;
694 ntd->td_flags |= TDF_IDLE_NOHLT;
695 set_mplock_contention_mask(gd);
696 cpu_mplock_contested();
697 goto using_idle_thread;
699 clr_mplock_contention_mask(gd);
700 ++gd->gd_cnt.v_swtch;
701 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
702 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 clr_mplock_contention_mask(gd);
706 ++gd->gd_cnt.v_swtch;
707 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
708 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
713 * worry about tokens or the BGL. However, we still have
714 * to call lwkt_getalltokens() in order to properly detect
715 * stale tokens. This call cannot fail for a UP build!
717 lwkt_getalltokens(ntd);
718 ++gd->gd_cnt.v_swtch;
719 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
720 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
724 * We have nothing to run but only let the idle loop halt
725 * the cpu if there are no pending interrupts.
727 ntd = &gd->gd_idlethread;
728 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
729 ntd->td_flags |= TDF_IDLE_NOHLT;
733 * The idle thread should not be holding the MP lock unless we
734 * are trapping in the kernel or in a panic. Since we select the
735 * idle thread unconditionally when no other thread is available,
736 * if the MP lock is desired during a panic or kernel trap, we
737 * have to loop in the scheduler until we get it.
739 if (ntd->td_mpcount) {
740 mpheld = MP_LOCK_HELD();
741 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
742 panic("Idle thread %p was holding the BGL!", ntd);
749 KASSERT(ntd->td_pri >= TDPRI_CRIT,
750 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
753 * Do the actual switch. If the new target does not need the MP lock
754 * and we are holding it, release the MP lock. If the new target requires
755 * the MP lock we have already acquired it for the target.
758 if (ntd->td_mpcount == 0 ) {
762 ASSERT_MP_LOCK_HELD(ntd);
769 int tos_ok __debugvar = jg_tos_ok(ntd);
773 KTR_LOG(ctxsw_sw, td, ntd);
776 /* NOTE: current cpu may have changed after switch */
781 * Request that the target thread preempt the current thread. Preemption
782 * only works under a specific set of conditions:
784 * - We are not preempting ourselves
785 * - The target thread is owned by the current cpu
786 * - We are not currently being preempted
787 * - The target is not currently being preempted
788 * - We are not holding any spin locks
789 * - The target thread is not holding any tokens
790 * - We are able to satisfy the target's MP lock requirements (if any).
792 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
793 * this is called via lwkt_schedule() through the td_preemptable callback.
794 * critpri is the managed critical priority that we should ignore in order
795 * to determine whether preemption is possible (aka usually just the crit
796 * priority of lwkt_schedule() itself).
798 * XXX at the moment we run the target thread in a critical section during
799 * the preemption in order to prevent the target from taking interrupts
800 * that *WE* can't. Preemption is strictly limited to interrupt threads
801 * and interrupt-like threads, outside of a critical section, and the
802 * preempted source thread will be resumed the instant the target blocks
803 * whether or not the source is scheduled (i.e. preemption is supposed to
804 * be as transparent as possible).
806 * The target thread inherits our MP count (added to its own) for the
807 * duration of the preemption in order to preserve the atomicy of the
808 * MP lock during the preemption. Therefore, any preempting targets must be
809 * careful in regards to MP assertions. Note that the MP count may be
810 * out of sync with the physical mp_lock, but we do not have to preserve
811 * the original ownership of the lock if it was out of synch (that is, we
812 * can leave it synchronized on return).
815 lwkt_preempt(thread_t ntd, int critpri)
817 struct globaldata *gd = mycpu;
825 * The caller has put us in a critical section. We can only preempt
826 * if the caller of the caller was not in a critical section (basically
827 * a local interrupt), as determined by the 'critpri' parameter. We
828 * also can't preempt if the caller is holding any spinlocks (even if
829 * he isn't in a critical section). This also handles the tokens test.
831 * YYY The target thread must be in a critical section (else it must
832 * inherit our critical section? I dunno yet).
834 * Set need_lwkt_resched() unconditionally for now YYY.
836 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
838 td = gd->gd_curthread;
839 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
843 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
849 if (ntd->td_gd != gd) {
856 * Take the easy way out and do not preempt if we are holding
857 * any spinlocks. We could test whether the thread(s) being
858 * preempted interlock against the target thread's tokens and whether
859 * we can get all the target thread's tokens, but this situation
860 * should not occur very often so its easier to simply not preempt.
861 * Also, plain spinlocks are impossible to figure out at this point so
862 * just don't preempt.
864 * Do not try to preempt if the target thread is holding any tokens.
865 * We could try to acquire the tokens but this case is so rare there
866 * is no need to support it.
868 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
878 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
883 if (ntd->td_preempted) {
890 * note: an interrupt might have occured just as we were transitioning
891 * to or from the MP lock. In this case td_mpcount will be pre-disposed
892 * (non-zero) but not actually synchronized with the actual state of the
893 * lock. We can use it to imply an MP lock requirement for the
894 * preemption but we cannot use it to test whether we hold the MP lock
897 savecnt = td->td_mpcount;
898 mpheld = MP_LOCK_HELD();
899 ntd->td_mpcount += td->td_mpcount;
900 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
901 ntd->td_mpcount -= td->td_mpcount;
909 * Since we are able to preempt the current thread, there is no need to
910 * call need_lwkt_resched().
913 ntd->td_preempted = td;
914 td->td_flags |= TDF_PREEMPT_LOCK;
915 KTR_LOG(ctxsw_pre, td, ntd);
918 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
920 KKASSERT(savecnt == td->td_mpcount);
921 mpheld = MP_LOCK_HELD();
922 if (mpheld && td->td_mpcount == 0)
924 else if (mpheld == 0 && td->td_mpcount)
925 panic("lwkt_preempt(): MP lock was not held through");
927 ntd->td_preempted = NULL;
928 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
932 * Conditionally call splz() if gd_reqflags indicates work is pending.
934 * td_nest_count prevents deep nesting via splz() or doreti() which
935 * might otherwise blow out the kernel stack. Note that except for
936 * this special case, we MUST call splz() here to handle any
937 * pending ints, particularly after we switch, or we might accidently
938 * halt the cpu with interrupts pending.
940 * (self contained on a per cpu basis)
945 globaldata_t gd = mycpu;
946 thread_t td = gd->gd_curthread;
948 if (gd->gd_reqflags && td->td_nest_count < 2)
953 * This implements a normal yield which will yield to equal priority
954 * threads as well as higher priority threads. Note that gd_reqflags
955 * tests will be handled by the crit_exit() call in lwkt_switch().
957 * (self contained on a per cpu basis)
962 lwkt_schedule_self(curthread);
967 * This function is used along with the lwkt_passive_recover() inline
968 * by the trap code to negotiate a passive release of the current
969 * process/lwp designation with the user scheduler.
972 lwkt_passive_release(struct thread *td)
974 struct lwp *lp = td->td_lwp;
976 td->td_release = NULL;
977 lwkt_setpri_self(TDPRI_KERN_USER);
978 lp->lwp_proc->p_usched->release_curproc(lp);
982 * Make a kernel thread act as if it were in user mode with regards
983 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
984 * loops which may be potentially cpu-bound can call lwkt_user_yield().
986 * The lwkt_user_yield() function is designed to have very low overhead
987 * if no yield is determined to be needed.
990 lwkt_user_yield(void)
992 thread_t td = curthread;
993 struct lwp *lp = td->td_lwp;
997 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
998 * kernel can prevent other cpus from servicing interrupt threads
999 * which still require the MP lock (which is a lot of them). This
1000 * has a chaining effect since if the interrupt is blocked, so is
1001 * the event, so normal scheduling will not pick up on the problem.
1003 if (mp_lock_contention_mask && td->td_mpcount) {
1009 * Another kernel thread wants the cpu
1011 if (lwkt_resched_wanted())
1015 * If the user scheduler has asynchronously determined that the current
1016 * process (when running in user mode) needs to lose the cpu then make
1017 * sure we are released.
1019 if (user_resched_wanted()) {
1025 * If we are released reduce our priority
1027 if (td->td_release == NULL) {
1028 if (lwkt_check_resched(td) > 0)
1031 lp->lwp_proc->p_usched->acquire_curproc(lp);
1032 td->td_release = lwkt_passive_release;
1033 lwkt_setpri_self(TDPRI_USER_NORM);
1039 * Return 0 if no runnable threads are pending at the same or higher
1040 * priority as the passed thread.
1042 * Return 1 if runnable threads are pending at the same priority.
1044 * Return 2 if runnable threads are pending at a higher priority.
1047 lwkt_check_resched(thread_t td)
1049 int pri = td->td_pri & TDPRI_MASK;
1051 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1053 if (TAILQ_NEXT(td, td_threadq))
1059 * Generic schedule. Possibly schedule threads belonging to other cpus and
1060 * deal with threads that might be blocked on a wait queue.
1062 * We have a little helper inline function which does additional work after
1063 * the thread has been enqueued, including dealing with preemption and
1064 * setting need_lwkt_resched() (which prevents the kernel from returning
1065 * to userland until it has processed higher priority threads).
1067 * It is possible for this routine to be called after a failed _enqueue
1068 * (due to the target thread migrating, sleeping, or otherwise blocked).
1069 * We have to check that the thread is actually on the run queue!
1071 * reschedok is an optimized constant propagated from lwkt_schedule() or
1072 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1073 * reschedule to be requested if the target thread has a higher priority.
1074 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1075 * be 0, prevented undesired reschedules.
1079 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1083 if (ntd->td_flags & TDF_RUNQ) {
1084 if (ntd->td_preemptable && reschedok) {
1085 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1086 } else if (reschedok) {
1088 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1089 need_lwkt_resched();
1096 _lwkt_schedule(thread_t td, int reschedok)
1098 globaldata_t mygd = mycpu;
1100 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1101 crit_enter_gd(mygd);
1102 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1103 if (td == mygd->gd_curthread) {
1107 * If we own the thread, there is no race (since we are in a
1108 * critical section). If we do not own the thread there might
1109 * be a race but the target cpu will deal with it.
1112 if (td->td_gd == mygd) {
1114 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1116 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1120 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1127 lwkt_schedule(thread_t td)
1129 _lwkt_schedule(td, 1);
1133 lwkt_schedule_noresched(thread_t td)
1135 _lwkt_schedule(td, 0);
1141 * When scheduled remotely if frame != NULL the IPIQ is being
1142 * run via doreti or an interrupt then preemption can be allowed.
1144 * To allow preemption we have to drop the critical section so only
1145 * one is present in _lwkt_schedule_post.
1148 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1150 thread_t td = curthread;
1153 if (frame && ntd->td_preemptable) {
1154 crit_exit_noyield(td);
1155 _lwkt_schedule(ntd, 1);
1156 crit_enter_quick(td);
1158 _lwkt_schedule(ntd, 1);
1163 * Thread migration using a 'Pull' method. The thread may or may not be
1164 * the current thread. It MUST be descheduled and in a stable state.
1165 * lwkt_giveaway() must be called on the cpu owning the thread.
1167 * At any point after lwkt_giveaway() is called, the target cpu may
1168 * 'pull' the thread by calling lwkt_acquire().
1170 * We have to make sure the thread is not sitting on a per-cpu tsleep
1171 * queue or it will blow up when it moves to another cpu.
1173 * MPSAFE - must be called under very specific conditions.
1176 lwkt_giveaway(thread_t td)
1178 globaldata_t gd = mycpu;
1181 if (td->td_flags & TDF_TSLEEPQ)
1183 KKASSERT(td->td_gd == gd);
1184 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1185 td->td_flags |= TDF_MIGRATING;
1190 lwkt_acquire(thread_t td)
1195 KKASSERT(td->td_flags & TDF_MIGRATING);
1200 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1201 crit_enter_gd(mygd);
1202 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1204 lwkt_process_ipiq();
1209 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1210 td->td_flags &= ~TDF_MIGRATING;
1213 crit_enter_gd(mygd);
1214 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1215 td->td_flags &= ~TDF_MIGRATING;
1223 * Generic deschedule. Descheduling threads other then your own should be
1224 * done only in carefully controlled circumstances. Descheduling is
1227 * This function may block if the cpu has run out of messages.
1230 lwkt_deschedule(thread_t td)
1234 if (td == curthread) {
1237 if (td->td_gd == mycpu) {
1240 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1250 * Set the target thread's priority. This routine does not automatically
1251 * switch to a higher priority thread, LWKT threads are not designed for
1252 * continuous priority changes. Yield if you want to switch.
1254 * We have to retain the critical section count which uses the high bits
1255 * of the td_pri field. The specified priority may also indicate zero or
1256 * more critical sections by adding TDPRI_CRIT*N.
1258 * Note that we requeue the thread whether it winds up on a different runq
1259 * or not. uio_yield() depends on this and the routine is not normally
1260 * called with the same priority otherwise.
1263 lwkt_setpri(thread_t td, int pri)
1266 KKASSERT(td->td_gd == mycpu);
1268 if (td->td_flags & TDF_RUNQ) {
1270 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1273 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1279 * Set the initial priority for a thread prior to it being scheduled for
1280 * the first time. The thread MUST NOT be scheduled before or during
1281 * this call. The thread may be assigned to a cpu other then the current
1284 * Typically used after a thread has been created with TDF_STOPPREQ,
1285 * and before the thread is initially scheduled.
1288 lwkt_setpri_initial(thread_t td, int pri)
1291 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1292 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1296 lwkt_setpri_self(int pri)
1298 thread_t td = curthread;
1300 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1302 if (td->td_flags & TDF_RUNQ) {
1304 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1307 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1313 * Migrate the current thread to the specified cpu.
1315 * This is accomplished by descheduling ourselves from the current cpu,
1316 * moving our thread to the tdallq of the target cpu, IPI messaging the
1317 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1318 * races while the thread is being migrated.
1320 * We must be sure to remove ourselves from the current cpu's tsleepq
1321 * before potentially moving to another queue. The thread can be on
1322 * a tsleepq due to a left-over tsleep_interlock().
1325 static void lwkt_setcpu_remote(void *arg);
1329 lwkt_setcpu_self(globaldata_t rgd)
1332 thread_t td = curthread;
1334 if (td->td_gd != rgd) {
1335 crit_enter_quick(td);
1336 if (td->td_flags & TDF_TSLEEPQ)
1338 td->td_flags |= TDF_MIGRATING;
1339 lwkt_deschedule_self(td);
1340 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1341 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1343 /* we are now on the target cpu */
1344 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1345 crit_exit_quick(td);
1351 lwkt_migratecpu(int cpuid)
1356 rgd = globaldata_find(cpuid);
1357 lwkt_setcpu_self(rgd);
1362 * Remote IPI for cpu migration (called while in a critical section so we
1363 * do not have to enter another one). The thread has already been moved to
1364 * our cpu's allq, but we must wait for the thread to be completely switched
1365 * out on the originating cpu before we schedule it on ours or the stack
1366 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1367 * change to main memory.
1369 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1370 * against wakeups. It is best if this interface is used only when there
1371 * are no pending events that might try to schedule the thread.
1375 lwkt_setcpu_remote(void *arg)
1378 globaldata_t gd = mycpu;
1380 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1382 lwkt_process_ipiq();
1388 td->td_flags &= ~TDF_MIGRATING;
1389 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1395 lwkt_preempted_proc(void)
1397 thread_t td = curthread;
1398 while (td->td_preempted)
1399 td = td->td_preempted;
1404 * Create a kernel process/thread/whatever. It shares it's address space
1405 * with proc0 - ie: kernel only.
1407 * NOTE! By default new threads are created with the MP lock held. A
1408 * thread which does not require the MP lock should release it by calling
1409 * rel_mplock() at the start of the new thread.
1412 lwkt_create(void (*func)(void *), void *arg,
1413 struct thread **tdp, thread_t template, int tdflags, int cpu,
1414 const char *fmt, ...)
1419 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1423 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1426 * Set up arg0 for 'ps' etc
1428 __va_start(ap, fmt);
1429 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1433 * Schedule the thread to run
1435 if ((td->td_flags & TDF_STOPREQ) == 0)
1438 td->td_flags &= ~TDF_STOPREQ;
1443 * Destroy an LWKT thread. Warning! This function is not called when
1444 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1445 * uses a different reaping mechanism.
1450 thread_t td = curthread;
1454 if (td->td_flags & TDF_VERBOSE)
1455 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1459 * Get us into a critical section to interlock gd_freetd and loop
1460 * until we can get it freed.
1462 * We have to cache the current td in gd_freetd because objcache_put()ing
1463 * it would rip it out from under us while our thread is still active.
1466 crit_enter_quick(td);
1467 while ((std = gd->gd_freetd) != NULL) {
1468 gd->gd_freetd = NULL;
1469 objcache_put(thread_cache, std);
1473 * Remove thread resources from kernel lists and deschedule us for
1476 if (td->td_flags & TDF_TSLEEPQ)
1479 lwkt_deschedule_self(td);
1480 lwkt_remove_tdallq(td);
1481 if (td->td_flags & TDF_ALLOCATED_THREAD)
1487 lwkt_remove_tdallq(thread_t td)
1489 KKASSERT(td->td_gd == mycpu);
1490 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1491 dsched_exit_thread(td);
1497 thread_t td = curthread;
1498 int lpri = td->td_pri;
1501 panic("td_pri is/would-go negative! %p %d", td, lpri);
1507 * Called from debugger/panic on cpus which have been stopped. We must still
1508 * process the IPIQ while stopped, even if we were stopped while in a critical
1511 * If we are dumping also try to process any pending interrupts. This may
1512 * or may not work depending on the state of the cpu at the point it was
1516 lwkt_smp_stopped(void)
1518 globaldata_t gd = mycpu;
1522 lwkt_process_ipiq();
1525 lwkt_process_ipiq();