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 "Successful preemption events");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
144 "Failed preemption events");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
148 &token_contention_count, 0, "spinning due to token contention");
152 * These helper procedures handle the runq, they can only be called from
153 * within a critical section.
155 * WARNING! Prior to SMP being brought up it is possible to enqueue and
156 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
157 * instead of 'mycpu' when referencing the globaldata structure. Once
158 * SMP live enqueuing and dequeueing only occurs on the current cpu.
162 _lwkt_dequeue(thread_t td)
164 if (td->td_flags & TDF_RUNQ) {
165 int nq = td->td_pri & TDPRI_MASK;
166 struct globaldata *gd = td->td_gd;
168 td->td_flags &= ~TDF_RUNQ;
169 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
170 /* runqmask is passively cleaned up by the switcher */
176 _lwkt_enqueue(thread_t td)
178 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
179 int nq = td->td_pri & TDPRI_MASK;
180 struct globaldata *gd = td->td_gd;
182 td->td_flags |= TDF_RUNQ;
183 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
184 gd->gd_runqmask |= 1 << nq;
189 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
191 struct thread *td = (struct thread *)obj;
193 td->td_kstack = NULL;
194 td->td_kstack_size = 0;
195 td->td_flags = TDF_ALLOCATED_THREAD;
200 _lwkt_thread_dtor(void *obj, void *privdata)
202 struct thread *td = (struct thread *)obj;
204 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
205 ("_lwkt_thread_dtor: not allocated from objcache"));
206 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
207 td->td_kstack_size > 0,
208 ("_lwkt_thread_dtor: corrupted stack"));
209 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
213 * Initialize the lwkt s/system.
218 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
219 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
220 NULL, CACHE_NTHREADS/2,
221 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
225 * Schedule a thread to run. As the current thread we can always safely
226 * schedule ourselves, and a shortcut procedure is provided for that
229 * (non-blocking, self contained on a per cpu basis)
232 lwkt_schedule_self(thread_t td)
234 crit_enter_quick(td);
235 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
236 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
242 * Deschedule a thread.
244 * (non-blocking, self contained on a per cpu basis)
247 lwkt_deschedule_self(thread_t td)
249 crit_enter_quick(td);
255 * LWKTs operate on a per-cpu basis
257 * WARNING! Called from early boot, 'mycpu' may not work yet.
260 lwkt_gdinit(struct globaldata *gd)
264 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
265 TAILQ_INIT(&gd->gd_tdrunq[i]);
267 TAILQ_INIT(&gd->gd_tdallq);
271 * Create a new thread. The thread must be associated with a process context
272 * or LWKT start address before it can be scheduled. If the target cpu is
273 * -1 the thread will be created on the current cpu.
275 * If you intend to create a thread without a process context this function
276 * does everything except load the startup and switcher function.
279 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
281 globaldata_t gd = mycpu;
285 * If static thread storage is not supplied allocate a thread. Reuse
286 * a cached free thread if possible. gd_freetd is used to keep an exiting
287 * thread intact through the exit.
290 if ((td = gd->gd_freetd) != NULL)
291 gd->gd_freetd = NULL;
293 td = objcache_get(thread_cache, M_WAITOK);
294 KASSERT((td->td_flags &
295 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
296 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
297 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
301 * Try to reuse cached stack.
303 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
304 if (flags & TDF_ALLOCATED_STACK) {
305 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
310 stack = (void *)kmem_alloc(&kernel_map, stksize);
311 flags |= TDF_ALLOCATED_STACK;
314 lwkt_init_thread(td, stack, stksize, flags, gd);
316 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
321 * Initialize a preexisting thread structure. This function is used by
322 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
324 * All threads start out in a critical section at a priority of
325 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
326 * appropriate. This function may send an IPI message when the
327 * requested cpu is not the current cpu and consequently gd_tdallq may
328 * not be initialized synchronously from the point of view of the originating
331 * NOTE! we have to be careful in regards to creating threads for other cpus
332 * if SMP has not yet been activated.
337 lwkt_init_thread_remote(void *arg)
342 * Protected by critical section held by IPI dispatch
344 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
350 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
351 struct globaldata *gd)
353 globaldata_t mygd = mycpu;
355 bzero(td, sizeof(struct thread));
356 td->td_kstack = stack;
357 td->td_kstack_size = stksize;
358 td->td_flags = flags;
360 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
361 td->td_toks_stop = &td->td_toks_base;
363 if ((flags & TDF_MPSAFE) == 0)
366 if (lwkt_use_spin_port)
367 lwkt_initport_spin(&td->td_msgport);
369 lwkt_initport_thread(&td->td_msgport, td);
370 pmap_init_thread(td);
373 * Normally initializing a thread for a remote cpu requires sending an
374 * IPI. However, the idlethread is setup before the other cpus are
375 * activated so we have to treat it as a special case. XXX manipulation
376 * of gd_tdallq requires the BGL.
378 if (gd == mygd || td == &gd->gd_idlethread) {
380 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
383 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
387 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
391 dsched_new_thread(td);
395 lwkt_set_comm(thread_t td, const char *ctl, ...)
400 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
402 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
406 lwkt_hold(thread_t td)
412 lwkt_rele(thread_t td)
414 KKASSERT(td->td_refs > 0);
419 lwkt_wait_free(thread_t td)
422 tsleep(td, 0, "tdreap", hz);
426 lwkt_free_thread(thread_t td)
428 KASSERT((td->td_flags & TDF_RUNNING) == 0,
429 ("lwkt_free_thread: did not exit! %p", td));
431 if (td->td_flags & TDF_ALLOCATED_THREAD) {
432 objcache_put(thread_cache, td);
433 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
434 /* client-allocated struct with internally allocated stack */
435 KASSERT(td->td_kstack && td->td_kstack_size > 0,
436 ("lwkt_free_thread: corrupted stack"));
437 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
438 td->td_kstack = NULL;
439 td->td_kstack_size = 0;
441 KTR_LOG(ctxsw_deadtd, td);
446 * Switch to the next runnable lwkt. If no LWKTs are runnable then
447 * switch to the idlethread. Switching must occur within a critical
448 * section to avoid races with the scheduling queue.
450 * We always have full control over our cpu's run queue. Other cpus
451 * that wish to manipulate our queue must use the cpu_*msg() calls to
452 * talk to our cpu, so a critical section is all that is needed and
453 * the result is very, very fast thread switching.
455 * The LWKT scheduler uses a fixed priority model and round-robins at
456 * each priority level. User process scheduling is a totally
457 * different beast and LWKT priorities should not be confused with
458 * user process priorities.
460 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
461 * cleans it up. Note that the td_switch() function cannot do anything that
462 * requires the MP lock since the MP lock will have already been setup for
463 * the target thread (not the current thread). It's nice to have a scheduler
464 * that does not need the MP lock to work because it allows us to do some
465 * really cool high-performance MP lock optimizations.
467 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
468 * is not called by the current thread in the preemption case, only when
469 * the preempting thread blocks (in order to return to the original thread).
474 globaldata_t gd = mycpu;
475 thread_t td = gd->gd_curthread;
482 * Switching from within a 'fast' (non thread switched) interrupt or IPI
483 * is illegal. However, we may have to do it anyway if we hit a fatal
484 * kernel trap or we have paniced.
486 * If this case occurs save and restore the interrupt nesting level.
488 if (gd->gd_intr_nesting_level) {
492 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
493 panic("lwkt_switch: cannot switch from within "
494 "a fast interrupt, yet, td %p\n", td);
496 savegdnest = gd->gd_intr_nesting_level;
497 savegdtrap = gd->gd_trap_nesting_level;
498 gd->gd_intr_nesting_level = 0;
499 gd->gd_trap_nesting_level = 0;
500 if ((td->td_flags & TDF_PANICWARN) == 0) {
501 td->td_flags |= TDF_PANICWARN;
502 kprintf("Warning: thread switch from interrupt or IPI, "
503 "thread %p (%s)\n", td, td->td_comm);
507 gd->gd_intr_nesting_level = savegdnest;
508 gd->gd_trap_nesting_level = savegdtrap;
514 * Passive release (used to transition from user to kernel mode
515 * when we block or switch rather then when we enter the kernel).
516 * This function is NOT called if we are switching into a preemption
517 * or returning from a preemption. Typically this causes us to lose
518 * our current process designation (if we have one) and become a true
519 * LWKT thread, and may also hand the current process designation to
520 * another process and schedule thread.
526 if (TD_TOKS_HELD(td))
527 lwkt_relalltokens(td);
530 * We had better not be holding any spin locks, but don't get into an
531 * endless panic loop.
533 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
534 ("lwkt_switch: still holding a shared spinlock %p!",
535 gd->gd_spinlock_rd));
536 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
537 ("lwkt_switch: still holding %d exclusive spinlocks!",
538 gd->gd_spinlocks_wr));
543 * td_mpcount cannot be used to determine if we currently hold the
544 * MP lock because get_mplock() will increment it prior to attempting
545 * to get the lock, and switch out if it can't. Our ownership of
546 * the actual lock will remain stable while we are in a critical section
547 * (but, of course, another cpu may own or release the lock so the
548 * actual value of mp_lock is not stable).
550 mpheld = MP_LOCK_HELD();
552 if (td->td_cscount) {
553 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
555 if (panic_on_cscount)
556 panic("switching while mastering cpusync");
560 if ((ntd = td->td_preempted) != NULL) {
562 * We had preempted another thread on this cpu, resume the preempted
563 * thread. This occurs transparently, whether the preempted thread
564 * was scheduled or not (it may have been preempted after descheduling
567 * We have to setup the MP lock for the original thread after backing
568 * out the adjustment that was made to curthread when the original
571 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
573 if (ntd->td_mpcount && mpheld == 0) {
574 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
575 td, ntd, td->td_mpcount, ntd->td_mpcount);
577 if (ntd->td_mpcount) {
578 td->td_mpcount -= ntd->td_mpcount;
579 KKASSERT(td->td_mpcount >= 0);
582 ntd->td_flags |= TDF_PREEMPT_DONE;
585 * The interrupt may have woken a thread up, we need to properly
586 * set the reschedule flag if the originally interrupted thread is
587 * at a lower priority.
589 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
591 /* YYY release mp lock on switchback if original doesn't need it */
594 * Priority queue / round-robin at each priority. Note that user
595 * processes run at a fixed, low priority and the user process
596 * scheduler deals with interactions between user processes
597 * by scheduling and descheduling them from the LWKT queue as
600 * We have to adjust the MP lock for the target thread. If we
601 * need the MP lock and cannot obtain it we try to locate a
602 * thread that does not need the MP lock. If we cannot, we spin
605 * A similar issue exists for the tokens held by the target thread.
606 * If we cannot obtain ownership of the tokens we cannot immediately
607 * schedule the thread.
611 * If an LWKT reschedule was requested, well that is what we are
612 * doing now so clear it.
614 clear_lwkt_resched();
616 if (gd->gd_runqmask) {
617 int nq = bsrl(gd->gd_runqmask);
618 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
619 gd->gd_runqmask &= ~(1 << nq);
624 * THREAD SELECTION FOR AN SMP MACHINE BUILD
626 * If the target needs the MP lock and we couldn't get it,
627 * or if the target is holding tokens and we could not
628 * gain ownership of the tokens, continue looking for a
629 * thread to schedule and spin instead of HLT if we can't.
631 * NOTE: the mpheld variable invalid after this conditional, it
632 * can change due to both cpu_try_mplock() returning success
633 * AND interactions in lwkt_getalltokens() due to the fact that
634 * we are trying to check the mpcount of a thread other then
635 * the current thread. Because of this, if the current thread
636 * is not holding td_mpcount, an IPI indirectly run via
637 * lwkt_getalltokens() can obtain and release the MP lock and
638 * cause the core MP lock to be released.
640 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
641 (TD_TOKS_HELD(ntd) && lwkt_getalltokens(ntd) == 0)
643 u_int32_t rqmask = gd->gd_runqmask;
647 mpheld = MP_LOCK_HELD();
650 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
651 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
652 /* spinning due to MP lock being held */
657 * mpheld state invalid after getalltokens call returns
658 * failure, but the variable is only needed for
661 if (TD_TOKS_HELD(ntd) && !lwkt_getalltokens(ntd)) {
662 /* spinning due to token contention */
664 ++token_contention_count;
666 mpheld = MP_LOCK_HELD();
673 rqmask &= ~(1 << nq);
677 * We have two choices. We can either refuse to run a
678 * user thread when a kernel thread needs the MP lock
679 * but could not get it, or we can allow it to run but
680 * then expect an IPI (hopefully) later on to force a
681 * reschedule when the MP lock might become available.
683 if (nq < TDPRI_KERN_LPSCHED) {
684 break; /* for now refuse to run */
686 if (chain_mplock == 0)
688 /* continue loop, allow user threads to be scheduled */
694 * Case where a (kernel) thread needed the MP lock and could
695 * not get one, and we may or may not have found another
696 * thread which does not need the MP lock to run while
700 ntd = &gd->gd_idlethread;
701 ntd->td_flags |= TDF_IDLE_NOHLT;
702 set_mplock_contention_mask(gd);
703 cpu_mplock_contested();
704 goto using_idle_thread;
706 clr_mplock_contention_mask(gd);
707 ++gd->gd_cnt.v_swtch;
708 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
709 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 clr_mplock_contention_mask(gd);
713 ++gd->gd_cnt.v_swtch;
714 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
715 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
719 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
720 * worry about tokens or the BGL. However, we still have
721 * to call lwkt_getalltokens() in order to properly detect
722 * stale tokens. This call cannot fail for a UP build!
724 lwkt_getalltokens(ntd);
725 ++gd->gd_cnt.v_swtch;
726 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
727 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
731 * We have nothing to run but only let the idle loop halt
732 * the cpu if there are no pending interrupts.
734 ntd = &gd->gd_idlethread;
735 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
736 ntd->td_flags |= TDF_IDLE_NOHLT;
740 * The idle thread should not be holding the MP lock unless we
741 * are trapping in the kernel or in a panic. Since we select the
742 * idle thread unconditionally when no other thread is available,
743 * if the MP lock is desired during a panic or kernel trap, we
744 * have to loop in the scheduler until we get it.
746 if (ntd->td_mpcount) {
747 mpheld = MP_LOCK_HELD();
748 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
749 panic("Idle thread %p was holding the BGL!", ntd);
756 KASSERT(ntd->td_pri >= TDPRI_CRIT,
757 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
760 * Do the actual switch. If the new target does not need the MP lock
761 * and we are holding it, release the MP lock. If the new target requires
762 * the MP lock we have already acquired it for the target.
765 if (ntd->td_mpcount == 0 ) {
769 ASSERT_MP_LOCK_HELD(ntd);
776 int tos_ok __debugvar = jg_tos_ok(ntd);
780 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
783 /* NOTE: current cpu may have changed after switch */
788 * Request that the target thread preempt the current thread. Preemption
789 * only works under a specific set of conditions:
791 * - We are not preempting ourselves
792 * - The target thread is owned by the current cpu
793 * - We are not currently being preempted
794 * - The target is not currently being preempted
795 * - We are not holding any spin locks
796 * - The target thread is not holding any tokens
797 * - We are able to satisfy the target's MP lock requirements (if any).
799 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
800 * this is called via lwkt_schedule() through the td_preemptable callback.
801 * critpri is the managed critical priority that we should ignore in order
802 * to determine whether preemption is possible (aka usually just the crit
803 * priority of lwkt_schedule() itself).
805 * XXX at the moment we run the target thread in a critical section during
806 * the preemption in order to prevent the target from taking interrupts
807 * that *WE* can't. Preemption is strictly limited to interrupt threads
808 * and interrupt-like threads, outside of a critical section, and the
809 * preempted source thread will be resumed the instant the target blocks
810 * whether or not the source is scheduled (i.e. preemption is supposed to
811 * be as transparent as possible).
813 * The target thread inherits our MP count (added to its own) for the
814 * duration of the preemption in order to preserve the atomicy of the
815 * MP lock during the preemption. Therefore, any preempting targets must be
816 * careful in regards to MP assertions. Note that the MP count may be
817 * out of sync with the physical mp_lock, but we do not have to preserve
818 * the original ownership of the lock if it was out of synch (that is, we
819 * can leave it synchronized on return).
822 lwkt_preempt(thread_t ntd, int critpri)
824 struct globaldata *gd = mycpu;
832 * The caller has put us in a critical section. We can only preempt
833 * if the caller of the caller was not in a critical section (basically
834 * a local interrupt), as determined by the 'critpri' parameter. We
835 * also can't preempt if the caller is holding any spinlocks (even if
836 * he isn't in a critical section). This also handles the tokens test.
838 * YYY The target thread must be in a critical section (else it must
839 * inherit our critical section? I dunno yet).
841 * Set need_lwkt_resched() unconditionally for now YYY.
843 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
845 td = gd->gd_curthread;
846 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
850 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
856 if (ntd->td_gd != gd) {
863 * Take the easy way out and do not preempt if we are holding
864 * any spinlocks. We could test whether the thread(s) being
865 * preempted interlock against the target thread's tokens and whether
866 * we can get all the target thread's tokens, but this situation
867 * should not occur very often so its easier to simply not preempt.
868 * Also, plain spinlocks are impossible to figure out at this point so
869 * just don't preempt.
871 * Do not try to preempt if the target thread is holding any tokens.
872 * We could try to acquire the tokens but this case is so rare there
873 * is no need to support it.
875 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
880 if (TD_TOKS_HELD(ntd)) {
885 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
890 if (ntd->td_preempted) {
897 * note: an interrupt might have occured just as we were transitioning
898 * to or from the MP lock. In this case td_mpcount will be pre-disposed
899 * (non-zero) but not actually synchronized with the actual state of the
900 * lock. We can use it to imply an MP lock requirement for the
901 * preemption but we cannot use it to test whether we hold the MP lock
904 savecnt = td->td_mpcount;
905 mpheld = MP_LOCK_HELD();
906 ntd->td_mpcount += td->td_mpcount;
907 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
908 ntd->td_mpcount -= td->td_mpcount;
916 * Since we are able to preempt the current thread, there is no need to
917 * call need_lwkt_resched().
920 ntd->td_preempted = td;
921 td->td_flags |= TDF_PREEMPT_LOCK;
922 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
925 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
927 KKASSERT(savecnt == td->td_mpcount);
928 mpheld = MP_LOCK_HELD();
929 if (mpheld && td->td_mpcount == 0)
931 else if (mpheld == 0 && td->td_mpcount)
932 panic("lwkt_preempt(): MP lock was not held through");
934 ntd->td_preempted = NULL;
935 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
939 * Conditionally call splz() if gd_reqflags indicates work is pending.
941 * td_nest_count prevents deep nesting via splz() or doreti() which
942 * might otherwise blow out the kernel stack. Note that except for
943 * this special case, we MUST call splz() here to handle any
944 * pending ints, particularly after we switch, or we might accidently
945 * halt the cpu with interrupts pending.
947 * (self contained on a per cpu basis)
952 globaldata_t gd = mycpu;
953 thread_t td = gd->gd_curthread;
955 if (gd->gd_reqflags && td->td_nest_count < 2)
960 * This implements a normal yield which will yield to equal priority
961 * threads as well as higher priority threads. Note that gd_reqflags
962 * tests will be handled by the crit_exit() call in lwkt_switch().
964 * (self contained on a per cpu basis)
969 lwkt_schedule_self(curthread);
974 * This function is used along with the lwkt_passive_recover() inline
975 * by the trap code to negotiate a passive release of the current
976 * process/lwp designation with the user scheduler.
979 lwkt_passive_release(struct thread *td)
981 struct lwp *lp = td->td_lwp;
983 td->td_release = NULL;
984 lwkt_setpri_self(TDPRI_KERN_USER);
985 lp->lwp_proc->p_usched->release_curproc(lp);
989 * Make a kernel thread act as if it were in user mode with regards
990 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
991 * loops which may be potentially cpu-bound can call lwkt_user_yield().
993 * The lwkt_user_yield() function is designed to have very low overhead
994 * if no yield is determined to be needed.
997 lwkt_user_yield(void)
999 thread_t td = curthread;
1000 struct lwp *lp = td->td_lwp;
1004 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1005 * kernel can prevent other cpus from servicing interrupt threads
1006 * which still require the MP lock (which is a lot of them). This
1007 * has a chaining effect since if the interrupt is blocked, so is
1008 * the event, so normal scheduling will not pick up on the problem.
1010 if (mp_lock_contention_mask && td->td_mpcount) {
1016 * Another kernel thread wants the cpu
1018 if (lwkt_resched_wanted())
1022 * If the user scheduler has asynchronously determined that the current
1023 * process (when running in user mode) needs to lose the cpu then make
1024 * sure we are released.
1026 if (user_resched_wanted()) {
1032 * If we are released reduce our priority
1034 if (td->td_release == NULL) {
1035 if (lwkt_check_resched(td) > 0)
1038 lp->lwp_proc->p_usched->acquire_curproc(lp);
1039 td->td_release = lwkt_passive_release;
1040 lwkt_setpri_self(TDPRI_USER_NORM);
1046 * Return 0 if no runnable threads are pending at the same or higher
1047 * priority as the passed thread.
1049 * Return 1 if runnable threads are pending at the same priority.
1051 * Return 2 if runnable threads are pending at a higher priority.
1054 lwkt_check_resched(thread_t td)
1056 int pri = td->td_pri & TDPRI_MASK;
1058 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1060 if (TAILQ_NEXT(td, td_threadq))
1066 * Generic schedule. Possibly schedule threads belonging to other cpus and
1067 * deal with threads that might be blocked on a wait queue.
1069 * We have a little helper inline function which does additional work after
1070 * the thread has been enqueued, including dealing with preemption and
1071 * setting need_lwkt_resched() (which prevents the kernel from returning
1072 * to userland until it has processed higher priority threads).
1074 * It is possible for this routine to be called after a failed _enqueue
1075 * (due to the target thread migrating, sleeping, or otherwise blocked).
1076 * We have to check that the thread is actually on the run queue!
1078 * reschedok is an optimized constant propagated from lwkt_schedule() or
1079 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1080 * reschedule to be requested if the target thread has a higher priority.
1081 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1082 * be 0, prevented undesired reschedules.
1086 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1090 if (ntd->td_flags & TDF_RUNQ) {
1091 if (ntd->td_preemptable && reschedok) {
1092 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1093 } else if (reschedok) {
1095 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1096 need_lwkt_resched();
1103 _lwkt_schedule(thread_t td, int reschedok)
1105 globaldata_t mygd = mycpu;
1107 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1108 crit_enter_gd(mygd);
1109 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1110 if (td == mygd->gd_curthread) {
1114 * If we own the thread, there is no race (since we are in a
1115 * critical section). If we do not own the thread there might
1116 * be a race but the target cpu will deal with it.
1119 if (td->td_gd == mygd) {
1121 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1123 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1127 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1134 lwkt_schedule(thread_t td)
1136 _lwkt_schedule(td, 1);
1140 lwkt_schedule_noresched(thread_t td)
1142 _lwkt_schedule(td, 0);
1148 * When scheduled remotely if frame != NULL the IPIQ is being
1149 * run via doreti or an interrupt then preemption can be allowed.
1151 * To allow preemption we have to drop the critical section so only
1152 * one is present in _lwkt_schedule_post.
1155 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1157 thread_t td = curthread;
1160 if (frame && ntd->td_preemptable) {
1161 crit_exit_noyield(td);
1162 _lwkt_schedule(ntd, 1);
1163 crit_enter_quick(td);
1165 _lwkt_schedule(ntd, 1);
1170 * Thread migration using a 'Pull' method. The thread may or may not be
1171 * the current thread. It MUST be descheduled and in a stable state.
1172 * lwkt_giveaway() must be called on the cpu owning the thread.
1174 * At any point after lwkt_giveaway() is called, the target cpu may
1175 * 'pull' the thread by calling lwkt_acquire().
1177 * We have to make sure the thread is not sitting on a per-cpu tsleep
1178 * queue or it will blow up when it moves to another cpu.
1180 * MPSAFE - must be called under very specific conditions.
1183 lwkt_giveaway(thread_t td)
1185 globaldata_t gd = mycpu;
1188 if (td->td_flags & TDF_TSLEEPQ)
1190 KKASSERT(td->td_gd == gd);
1191 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1192 td->td_flags |= TDF_MIGRATING;
1197 lwkt_acquire(thread_t td)
1202 KKASSERT(td->td_flags & TDF_MIGRATING);
1207 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1208 crit_enter_gd(mygd);
1209 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1211 lwkt_process_ipiq();
1216 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1217 td->td_flags &= ~TDF_MIGRATING;
1220 crit_enter_gd(mygd);
1221 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1222 td->td_flags &= ~TDF_MIGRATING;
1230 * Generic deschedule. Descheduling threads other then your own should be
1231 * done only in carefully controlled circumstances. Descheduling is
1234 * This function may block if the cpu has run out of messages.
1237 lwkt_deschedule(thread_t td)
1241 if (td == curthread) {
1244 if (td->td_gd == mycpu) {
1247 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1257 * Set the target thread's priority. This routine does not automatically
1258 * switch to a higher priority thread, LWKT threads are not designed for
1259 * continuous priority changes. Yield if you want to switch.
1261 * We have to retain the critical section count which uses the high bits
1262 * of the td_pri field. The specified priority may also indicate zero or
1263 * more critical sections by adding TDPRI_CRIT*N.
1265 * Note that we requeue the thread whether it winds up on a different runq
1266 * or not. uio_yield() depends on this and the routine is not normally
1267 * called with the same priority otherwise.
1270 lwkt_setpri(thread_t td, int pri)
1273 KKASSERT(td->td_gd == mycpu);
1275 if (td->td_flags & TDF_RUNQ) {
1277 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1280 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1286 * Set the initial priority for a thread prior to it being scheduled for
1287 * the first time. The thread MUST NOT be scheduled before or during
1288 * this call. The thread may be assigned to a cpu other then the current
1291 * Typically used after a thread has been created with TDF_STOPPREQ,
1292 * and before the thread is initially scheduled.
1295 lwkt_setpri_initial(thread_t td, int pri)
1298 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1299 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1303 lwkt_setpri_self(int pri)
1305 thread_t td = curthread;
1307 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1309 if (td->td_flags & TDF_RUNQ) {
1311 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1314 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1320 * Migrate the current thread to the specified cpu.
1322 * This is accomplished by descheduling ourselves from the current cpu,
1323 * moving our thread to the tdallq of the target cpu, IPI messaging the
1324 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1325 * races while the thread is being migrated.
1327 * We must be sure to remove ourselves from the current cpu's tsleepq
1328 * before potentially moving to another queue. The thread can be on
1329 * a tsleepq due to a left-over tsleep_interlock().
1332 static void lwkt_setcpu_remote(void *arg);
1336 lwkt_setcpu_self(globaldata_t rgd)
1339 thread_t td = curthread;
1341 if (td->td_gd != rgd) {
1342 crit_enter_quick(td);
1343 if (td->td_flags & TDF_TSLEEPQ)
1345 td->td_flags |= TDF_MIGRATING;
1346 lwkt_deschedule_self(td);
1347 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1348 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1350 /* we are now on the target cpu */
1351 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1352 crit_exit_quick(td);
1358 lwkt_migratecpu(int cpuid)
1363 rgd = globaldata_find(cpuid);
1364 lwkt_setcpu_self(rgd);
1369 * Remote IPI for cpu migration (called while in a critical section so we
1370 * do not have to enter another one). The thread has already been moved to
1371 * our cpu's allq, but we must wait for the thread to be completely switched
1372 * out on the originating cpu before we schedule it on ours or the stack
1373 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1374 * change to main memory.
1376 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1377 * against wakeups. It is best if this interface is used only when there
1378 * are no pending events that might try to schedule the thread.
1382 lwkt_setcpu_remote(void *arg)
1385 globaldata_t gd = mycpu;
1387 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1389 lwkt_process_ipiq();
1395 td->td_flags &= ~TDF_MIGRATING;
1396 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1402 lwkt_preempted_proc(void)
1404 thread_t td = curthread;
1405 while (td->td_preempted)
1406 td = td->td_preempted;
1411 * Create a kernel process/thread/whatever. It shares it's address space
1412 * with proc0 - ie: kernel only.
1414 * NOTE! By default new threads are created with the MP lock held. A
1415 * thread which does not require the MP lock should release it by calling
1416 * rel_mplock() at the start of the new thread.
1419 lwkt_create(void (*func)(void *), void *arg,
1420 struct thread **tdp, thread_t template, int tdflags, int cpu,
1421 const char *fmt, ...)
1426 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1430 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1433 * Set up arg0 for 'ps' etc
1435 __va_start(ap, fmt);
1436 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1440 * Schedule the thread to run
1442 if ((td->td_flags & TDF_STOPREQ) == 0)
1445 td->td_flags &= ~TDF_STOPREQ;
1450 * Destroy an LWKT thread. Warning! This function is not called when
1451 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1452 * uses a different reaping mechanism.
1457 thread_t td = curthread;
1461 if (td->td_flags & TDF_VERBOSE)
1462 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1466 * Get us into a critical section to interlock gd_freetd and loop
1467 * until we can get it freed.
1469 * We have to cache the current td in gd_freetd because objcache_put()ing
1470 * it would rip it out from under us while our thread is still active.
1473 crit_enter_quick(td);
1474 while ((std = gd->gd_freetd) != NULL) {
1475 gd->gd_freetd = NULL;
1476 objcache_put(thread_cache, std);
1480 * Remove thread resources from kernel lists and deschedule us for
1483 if (td->td_flags & TDF_TSLEEPQ)
1486 dsched_exit_thread(td);
1487 lwkt_deschedule_self(td);
1488 lwkt_remove_tdallq(td);
1489 if (td->td_flags & TDF_ALLOCATED_THREAD)
1495 lwkt_remove_tdallq(thread_t td)
1497 KKASSERT(td->td_gd == mycpu);
1498 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1504 thread_t td = curthread;
1505 int lpri = td->td_pri;
1508 panic("td_pri is/would-go negative! %p %d", td, lpri);
1514 * Called from debugger/panic on cpus which have been stopped. We must still
1515 * process the IPIQ while stopped, even if we were stopped while in a critical
1518 * If we are dumping also try to process any pending interrupts. This may
1519 * or may not work depending on the state of the cpu at the point it was
1523 lwkt_smp_stopped(void)
1525 globaldata_t gd = mycpu;
1529 lwkt_process_ipiq();
1532 lwkt_process_ipiq();