2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.120 2008/10/26 04:29:19 sephe Exp $
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
49 #include <sys/rtprio.h>
50 #include <sys/queue.h>
51 #include <sys/sysctl.h>
52 #include <sys/kthread.h>
53 #include <machine/cpu.h>
56 #include <sys/spinlock.h>
59 #include <sys/thread2.h>
60 #include <sys/spinlock2.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>
78 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
80 static int untimely_switch = 0;
82 static int panic_on_cscount = 0;
84 static __int64_t switch_count = 0;
85 static __int64_t preempt_hit = 0;
86 static __int64_t preempt_miss = 0;
87 static __int64_t preempt_weird = 0;
88 static __int64_t token_contention_count = 0;
89 static __int64_t mplock_contention_count = 0;
90 static int lwkt_use_spin_port;
91 static int chain_mplock = 0;
92 static struct objcache *thread_cache;
94 volatile cpumask_t mp_lock_contention_mask;
97 * We can make all thread ports use the spin backend instead of the thread
98 * backend. This should only be set to debug the spin backend.
100 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
102 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
104 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
107 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
109 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
112 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
115 &token_contention_count, 0, "spinning due to token contention");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
117 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
123 #if !defined(KTR_GIANT_CONTENTION)
124 #define KTR_GIANT_CONTENTION KTR_ALL
127 KTR_INFO_MASTER(giant);
128 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
129 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
131 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
134 * These helper procedures handle the runq, they can only be called from
135 * within a critical section.
137 * WARNING! Prior to SMP being brought up it is possible to enqueue and
138 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
139 * instead of 'mycpu' when referencing the globaldata structure. Once
140 * SMP live enqueuing and dequeueing only occurs on the current cpu.
144 _lwkt_dequeue(thread_t td)
146 if (td->td_flags & TDF_RUNQ) {
147 int nq = td->td_pri & TDPRI_MASK;
148 struct globaldata *gd = td->td_gd;
150 td->td_flags &= ~TDF_RUNQ;
151 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
152 /* runqmask is passively cleaned up by the switcher */
158 _lwkt_enqueue(thread_t td)
160 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
161 int nq = td->td_pri & TDPRI_MASK;
162 struct globaldata *gd = td->td_gd;
164 td->td_flags |= TDF_RUNQ;
165 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
166 gd->gd_runqmask |= 1 << nq;
171 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
173 struct thread *td = (struct thread *)obj;
175 td->td_kstack = NULL;
176 td->td_kstack_size = 0;
177 td->td_flags = TDF_ALLOCATED_THREAD;
182 _lwkt_thread_dtor(void *obj, void *privdata)
184 struct thread *td = (struct thread *)obj;
186 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
187 ("_lwkt_thread_dtor: not allocated from objcache"));
188 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
189 td->td_kstack_size > 0,
190 ("_lwkt_thread_dtor: corrupted stack"));
191 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
195 * Initialize the lwkt s/system.
200 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
201 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
202 NULL, CACHE_NTHREADS/2,
203 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
207 * Schedule a thread to run. As the current thread we can always safely
208 * schedule ourselves, and a shortcut procedure is provided for that
211 * (non-blocking, self contained on a per cpu basis)
214 lwkt_schedule_self(thread_t td)
216 crit_enter_quick(td);
217 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
218 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
224 * Deschedule a thread.
226 * (non-blocking, self contained on a per cpu basis)
229 lwkt_deschedule_self(thread_t td)
231 crit_enter_quick(td);
237 * LWKTs operate on a per-cpu basis
239 * WARNING! Called from early boot, 'mycpu' may not work yet.
242 lwkt_gdinit(struct globaldata *gd)
246 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
247 TAILQ_INIT(&gd->gd_tdrunq[i]);
249 TAILQ_INIT(&gd->gd_tdallq);
253 * Create a new thread. The thread must be associated with a process context
254 * or LWKT start address before it can be scheduled. If the target cpu is
255 * -1 the thread will be created on the current cpu.
257 * If you intend to create a thread without a process context this function
258 * does everything except load the startup and switcher function.
261 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
263 globaldata_t gd = mycpu;
267 * If static thread storage is not supplied allocate a thread. Reuse
268 * a cached free thread if possible. gd_freetd is used to keep an exiting
269 * thread intact through the exit.
272 if ((td = gd->gd_freetd) != NULL)
273 gd->gd_freetd = NULL;
275 td = objcache_get(thread_cache, M_WAITOK);
276 KASSERT((td->td_flags &
277 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
278 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
279 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
283 * Try to reuse cached stack.
285 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
286 if (flags & TDF_ALLOCATED_STACK) {
287 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
292 stack = (void *)kmem_alloc(&kernel_map, stksize);
293 flags |= TDF_ALLOCATED_STACK;
296 lwkt_init_thread(td, stack, stksize, flags, gd);
298 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
303 * Initialize a preexisting thread structure. This function is used by
304 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
306 * All threads start out in a critical section at a priority of
307 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
308 * appropriate. This function may send an IPI message when the
309 * requested cpu is not the current cpu and consequently gd_tdallq may
310 * not be initialized synchronously from the point of view of the originating
313 * NOTE! we have to be careful in regards to creating threads for other cpus
314 * if SMP has not yet been activated.
319 lwkt_init_thread_remote(void *arg)
324 * Protected by critical section held by IPI dispatch
326 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
332 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
333 struct globaldata *gd)
335 globaldata_t mygd = mycpu;
337 bzero(td, sizeof(struct thread));
338 td->td_kstack = stack;
339 td->td_kstack_size = stksize;
340 td->td_flags = flags;
342 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
344 if ((flags & TDF_MPSAFE) == 0)
347 if (lwkt_use_spin_port)
348 lwkt_initport_spin(&td->td_msgport);
350 lwkt_initport_thread(&td->td_msgport, td);
351 pmap_init_thread(td);
354 * Normally initializing a thread for a remote cpu requires sending an
355 * IPI. However, the idlethread is setup before the other cpus are
356 * activated so we have to treat it as a special case. XXX manipulation
357 * of gd_tdallq requires the BGL.
359 if (gd == mygd || td == &gd->gd_idlethread) {
361 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
364 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
368 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
374 lwkt_set_comm(thread_t td, const char *ctl, ...)
379 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
384 lwkt_hold(thread_t td)
390 lwkt_rele(thread_t td)
392 KKASSERT(td->td_refs > 0);
397 lwkt_wait_free(thread_t td)
400 tsleep(td, 0, "tdreap", hz);
404 lwkt_free_thread(thread_t td)
406 KASSERT((td->td_flags & TDF_RUNNING) == 0,
407 ("lwkt_free_thread: did not exit! %p", td));
409 if (td->td_flags & TDF_ALLOCATED_THREAD) {
410 objcache_put(thread_cache, td);
411 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
412 /* client-allocated struct with internally allocated stack */
413 KASSERT(td->td_kstack && td->td_kstack_size > 0,
414 ("lwkt_free_thread: corrupted stack"));
415 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
416 td->td_kstack = NULL;
417 td->td_kstack_size = 0;
423 * Switch to the next runnable lwkt. If no LWKTs are runnable then
424 * switch to the idlethread. Switching must occur within a critical
425 * section to avoid races with the scheduling queue.
427 * We always have full control over our cpu's run queue. Other cpus
428 * that wish to manipulate our queue must use the cpu_*msg() calls to
429 * talk to our cpu, so a critical section is all that is needed and
430 * the result is very, very fast thread switching.
432 * The LWKT scheduler uses a fixed priority model and round-robins at
433 * each priority level. User process scheduling is a totally
434 * different beast and LWKT priorities should not be confused with
435 * user process priorities.
437 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
438 * cleans it up. Note that the td_switch() function cannot do anything that
439 * requires the MP lock since the MP lock will have already been setup for
440 * the target thread (not the current thread). It's nice to have a scheduler
441 * that does not need the MP lock to work because it allows us to do some
442 * really cool high-performance MP lock optimizations.
444 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
445 * is not called by the current thread in the preemption case, only when
446 * the preempting thread blocks (in order to return to the original thread).
451 globaldata_t gd = mycpu;
452 thread_t td = gd->gd_curthread;
459 * Switching from within a 'fast' (non thread switched) interrupt or IPI
460 * is illegal. However, we may have to do it anyway if we hit a fatal
461 * kernel trap or we have paniced.
463 * If this case occurs save and restore the interrupt nesting level.
465 if (gd->gd_intr_nesting_level) {
469 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
470 panic("lwkt_switch: cannot switch from within "
471 "a fast interrupt, yet, td %p\n", td);
473 savegdnest = gd->gd_intr_nesting_level;
474 savegdtrap = gd->gd_trap_nesting_level;
475 gd->gd_intr_nesting_level = 0;
476 gd->gd_trap_nesting_level = 0;
477 if ((td->td_flags & TDF_PANICWARN) == 0) {
478 td->td_flags |= TDF_PANICWARN;
479 kprintf("Warning: thread switch from interrupt or IPI, "
480 "thread %p (%s)\n", td, td->td_comm);
482 db_print_backtrace();
486 gd->gd_intr_nesting_level = savegdnest;
487 gd->gd_trap_nesting_level = savegdtrap;
493 * Passive release (used to transition from user to kernel mode
494 * when we block or switch rather then when we enter the kernel).
495 * This function is NOT called if we are switching into a preemption
496 * or returning from a preemption. Typically this causes us to lose
497 * our current process designation (if we have one) and become a true
498 * LWKT thread, and may also hand the current process designation to
499 * another process and schedule thread.
506 lwkt_relalltokens(td);
509 * We had better not be holding any spin locks, but don't get into an
510 * endless panic loop.
512 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
513 ("lwkt_switch: still holding a shared spinlock %p!",
514 gd->gd_spinlock_rd));
515 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
516 ("lwkt_switch: still holding %d exclusive spinlocks!",
517 gd->gd_spinlocks_wr));
522 * td_mpcount cannot be used to determine if we currently hold the
523 * MP lock because get_mplock() will increment it prior to attempting
524 * to get the lock, and switch out if it can't. Our ownership of
525 * the actual lock will remain stable while we are in a critical section
526 * (but, of course, another cpu may own or release the lock so the
527 * actual value of mp_lock is not stable).
529 mpheld = MP_LOCK_HELD();
531 if (td->td_cscount) {
532 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
534 if (panic_on_cscount)
535 panic("switching while mastering cpusync");
539 if ((ntd = td->td_preempted) != NULL) {
541 * We had preempted another thread on this cpu, resume the preempted
542 * thread. This occurs transparently, whether the preempted thread
543 * was scheduled or not (it may have been preempted after descheduling
546 * We have to setup the MP lock for the original thread after backing
547 * out the adjustment that was made to curthread when the original
550 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
552 if (ntd->td_mpcount && mpheld == 0) {
553 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
554 td, ntd, td->td_mpcount, ntd->td_mpcount);
556 if (ntd->td_mpcount) {
557 td->td_mpcount -= ntd->td_mpcount;
558 KKASSERT(td->td_mpcount >= 0);
561 ntd->td_flags |= TDF_PREEMPT_DONE;
564 * The interrupt may have woken a thread up, we need to properly
565 * set the reschedule flag if the originally interrupted thread is
566 * at a lower priority.
568 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
570 /* YYY release mp lock on switchback if original doesn't need it */
573 * Priority queue / round-robin at each priority. Note that user
574 * processes run at a fixed, low priority and the user process
575 * scheduler deals with interactions between user processes
576 * by scheduling and descheduling them from the LWKT queue as
579 * We have to adjust the MP lock for the target thread. If we
580 * need the MP lock and cannot obtain it we try to locate a
581 * thread that does not need the MP lock. If we cannot, we spin
584 * A similar issue exists for the tokens held by the target thread.
585 * If we cannot obtain ownership of the tokens we cannot immediately
586 * schedule the thread.
590 * If an LWKT reschedule was requested, well that is what we are
591 * doing now so clear it.
593 clear_lwkt_resched();
595 if (gd->gd_runqmask) {
596 int nq = bsrl(gd->gd_runqmask);
597 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
598 gd->gd_runqmask &= ~(1 << nq);
603 * THREAD SELECTION FOR AN SMP MACHINE BUILD
605 * If the target needs the MP lock and we couldn't get it,
606 * or if the target is holding tokens and we could not
607 * gain ownership of the tokens, continue looking for a
608 * thread to schedule and spin instead of HLT if we can't.
610 * NOTE: the mpheld variable invalid after this conditional, it
611 * can change due to both cpu_try_mplock() returning success
612 * AND interactions in lwkt_getalltokens() due to the fact that
613 * we are trying to check the mpcount of a thread other then
614 * the current thread. Because of this, if the current thread
615 * is not holding td_mpcount, an IPI indirectly run via
616 * lwkt_getalltokens() can obtain and release the MP lock and
617 * cause the core MP lock to be released.
619 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
620 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
622 u_int32_t rqmask = gd->gd_runqmask;
624 mpheld = MP_LOCK_HELD();
627 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
628 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
629 /* spinning due to MP lock being held */
631 ++mplock_contention_count;
633 /* mplock still not held, 'mpheld' still valid */
638 * mpheld state invalid after getalltokens call returns
639 * failure, but the variable is only needed for
642 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
643 /* spinning due to token contention */
645 ++token_contention_count;
647 mpheld = MP_LOCK_HELD();
654 rqmask &= ~(1 << nq);
658 * We have two choices. We can either refuse to run a
659 * user thread when a kernel thread needs the MP lock
660 * but could not get it, or we can allow it to run but
661 * then expect an IPI (hopefully) later on to force a
662 * reschedule when the MP lock might become available.
664 if (nq < TDPRI_KERN_LPSCHED) {
665 if (chain_mplock == 0)
667 atomic_set_int(&mp_lock_contention_mask,
669 /* continue loop, allow user threads to be scheduled */
673 cpu_mplock_contested();
674 ntd = &gd->gd_idlethread;
675 ntd->td_flags |= TDF_IDLE_NOHLT;
676 goto using_idle_thread;
678 ++gd->gd_cnt.v_swtch;
679 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
680 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
683 ++gd->gd_cnt.v_swtch;
684 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
685 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
689 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
690 * worry about tokens or the BGL. However, we still have
691 * to call lwkt_getalltokens() in order to properly detect
692 * stale tokens. This call cannot fail for a UP build!
694 lwkt_getalltokens(ntd);
695 ++gd->gd_cnt.v_swtch;
696 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
697 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
701 * We have nothing to run but only let the idle loop halt
702 * the cpu if there are no pending interrupts.
704 ntd = &gd->gd_idlethread;
705 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
706 ntd->td_flags |= TDF_IDLE_NOHLT;
710 * The idle thread should not be holding the MP lock unless we
711 * are trapping in the kernel or in a panic. Since we select the
712 * idle thread unconditionally when no other thread is available,
713 * if the MP lock is desired during a panic or kernel trap, we
714 * have to loop in the scheduler until we get it.
716 if (ntd->td_mpcount) {
717 mpheld = MP_LOCK_HELD();
718 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
719 panic("Idle thread %p was holding the BGL!", ntd);
720 } else if (mpheld == 0) {
721 cpu_mplock_contested();
728 KASSERT(ntd->td_pri >= TDPRI_CRIT,
729 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
732 * Do the actual switch. If the new target does not need the MP lock
733 * and we are holding it, release the MP lock. If the new target requires
734 * the MP lock we have already acquired it for the target.
737 if (ntd->td_mpcount == 0 ) {
741 ASSERT_MP_LOCK_HELD(ntd);
748 /* NOTE: current cpu may have changed after switch */
753 * Request that the target thread preempt the current thread. Preemption
754 * only works under a specific set of conditions:
756 * - We are not preempting ourselves
757 * - The target thread is owned by the current cpu
758 * - We are not currently being preempted
759 * - The target is not currently being preempted
760 * - We are not holding any spin locks
761 * - The target thread is not holding any tokens
762 * - We are able to satisfy the target's MP lock requirements (if any).
764 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
765 * this is called via lwkt_schedule() through the td_preemptable callback.
766 * critpri is the managed critical priority that we should ignore in order
767 * to determine whether preemption is possible (aka usually just the crit
768 * priority of lwkt_schedule() itself).
770 * XXX at the moment we run the target thread in a critical section during
771 * the preemption in order to prevent the target from taking interrupts
772 * that *WE* can't. Preemption is strictly limited to interrupt threads
773 * and interrupt-like threads, outside of a critical section, and the
774 * preempted source thread will be resumed the instant the target blocks
775 * whether or not the source is scheduled (i.e. preemption is supposed to
776 * be as transparent as possible).
778 * The target thread inherits our MP count (added to its own) for the
779 * duration of the preemption in order to preserve the atomicy of the
780 * MP lock during the preemption. Therefore, any preempting targets must be
781 * careful in regards to MP assertions. Note that the MP count may be
782 * out of sync with the physical mp_lock, but we do not have to preserve
783 * the original ownership of the lock if it was out of synch (that is, we
784 * can leave it synchronized on return).
787 lwkt_preempt(thread_t ntd, int critpri)
789 struct globaldata *gd = mycpu;
797 * The caller has put us in a critical section. We can only preempt
798 * if the caller of the caller was not in a critical section (basically
799 * a local interrupt), as determined by the 'critpri' parameter. We
800 * also can't preempt if the caller is holding any spinlocks (even if
801 * he isn't in a critical section). This also handles the tokens test.
803 * YYY The target thread must be in a critical section (else it must
804 * inherit our critical section? I dunno yet).
806 * Set need_lwkt_resched() unconditionally for now YYY.
808 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
810 td = gd->gd_curthread;
811 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
815 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
821 if (ntd->td_gd != gd) {
828 * Take the easy way out and do not preempt if we are holding
829 * any spinlocks. We could test whether the thread(s) being
830 * preempted interlock against the target thread's tokens and whether
831 * we can get all the target thread's tokens, but this situation
832 * should not occur very often so its easier to simply not preempt.
833 * Also, plain spinlocks are impossible to figure out at this point so
834 * just don't preempt.
836 * Do not try to preempt if the target thread is holding any tokens.
837 * We could try to acquire the tokens but this case is so rare there
838 * is no need to support it.
840 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
850 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
855 if (ntd->td_preempted) {
862 * note: an interrupt might have occured just as we were transitioning
863 * to or from the MP lock. In this case td_mpcount will be pre-disposed
864 * (non-zero) but not actually synchronized with the actual state of the
865 * lock. We can use it to imply an MP lock requirement for the
866 * preemption but we cannot use it to test whether we hold the MP lock
869 savecnt = td->td_mpcount;
870 mpheld = MP_LOCK_HELD();
871 ntd->td_mpcount += td->td_mpcount;
872 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
873 ntd->td_mpcount -= td->td_mpcount;
881 * Since we are able to preempt the current thread, there is no need to
882 * call need_lwkt_resched().
885 ntd->td_preempted = td;
886 td->td_flags |= TDF_PREEMPT_LOCK;
889 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
891 KKASSERT(savecnt == td->td_mpcount);
892 mpheld = MP_LOCK_HELD();
893 if (mpheld && td->td_mpcount == 0)
895 else if (mpheld == 0 && td->td_mpcount)
896 panic("lwkt_preempt(): MP lock was not held through");
898 ntd->td_preempted = NULL;
899 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
903 * Yield our thread while higher priority threads are pending. This is
904 * typically called when we leave a critical section but it can be safely
905 * called while we are in a critical section.
907 * This function will not generally yield to equal priority threads but it
908 * can occur as a side effect. Note that lwkt_switch() is called from
909 * inside the critical section to prevent its own crit_exit() from reentering
910 * lwkt_yield_quick().
912 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
913 * came along but was blocked and made pending.
915 * (self contained on a per cpu basis)
918 lwkt_yield_quick(void)
920 globaldata_t gd = mycpu;
921 thread_t td = gd->gd_curthread;
924 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
925 * it with a non-zero cpl then we might not wind up calling splz after
926 * a task switch when the critical section is exited even though the
927 * new task could accept the interrupt.
929 * XXX from crit_exit() only called after last crit section is released.
930 * If called directly will run splz() even if in a critical section.
932 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
933 * except for this special case, we MUST call splz() here to handle any
934 * pending ints, particularly after we switch, or we might accidently
935 * halt the cpu with interrupts pending.
937 if (gd->gd_reqflags && td->td_nest_count < 2)
941 * YYY enabling will cause wakeup() to task-switch, which really
942 * confused the old 4.x code. This is a good way to simulate
943 * preemption and MP without actually doing preemption or MP, because a
944 * lot of code assumes that wakeup() does not block.
946 if (untimely_switch && td->td_nest_count == 0 &&
947 gd->gd_intr_nesting_level == 0
949 crit_enter_quick(td);
951 * YYY temporary hacks until we disassociate the userland scheduler
952 * from the LWKT scheduler.
954 if (td->td_flags & TDF_RUNQ) {
955 lwkt_switch(); /* will not reenter yield function */
957 lwkt_schedule_self(td); /* make sure we are scheduled */
958 lwkt_switch(); /* will not reenter yield function */
959 lwkt_deschedule_self(td); /* make sure we are descheduled */
961 crit_exit_noyield(td);
966 * This implements a normal yield which, unlike _quick, will yield to equal
967 * priority threads as well. Note that gd_reqflags tests will be handled by
968 * the crit_exit() call in lwkt_switch().
970 * (self contained on a per cpu basis)
975 lwkt_schedule_self(curthread);
980 * Return 0 if no runnable threads are pending at the same or higher
981 * priority as the passed thread.
983 * Return 1 if runnable threads are pending at the same priority.
985 * Return 2 if runnable threads are pending at a higher priority.
988 lwkt_check_resched(thread_t td)
990 int pri = td->td_pri & TDPRI_MASK;
992 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
994 if (TAILQ_NEXT(td, td_threadq))
1000 * Generic schedule. Possibly schedule threads belonging to other cpus and
1001 * deal with threads that might be blocked on a wait queue.
1003 * We have a little helper inline function which does additional work after
1004 * the thread has been enqueued, including dealing with preemption and
1005 * setting need_lwkt_resched() (which prevents the kernel from returning
1006 * to userland until it has processed higher priority threads).
1008 * It is possible for this routine to be called after a failed _enqueue
1009 * (due to the target thread migrating, sleeping, or otherwise blocked).
1010 * We have to check that the thread is actually on the run queue!
1012 * reschedok is an optimized constant propagated from lwkt_schedule() or
1013 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1014 * reschedule to be requested if the target thread has a higher priority.
1015 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1016 * be 0, prevented undesired reschedules.
1020 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1024 if (ntd->td_flags & TDF_RUNQ) {
1025 if (ntd->td_preemptable && reschedok) {
1026 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1027 } else if (reschedok) {
1029 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1030 need_lwkt_resched();
1037 _lwkt_schedule(thread_t td, int reschedok)
1039 globaldata_t mygd = mycpu;
1041 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1042 crit_enter_gd(mygd);
1043 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1044 if (td == mygd->gd_curthread) {
1048 * If we own the thread, there is no race (since we are in a
1049 * critical section). If we do not own the thread there might
1050 * be a race but the target cpu will deal with it.
1053 if (td->td_gd == mygd) {
1055 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1057 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1061 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1068 lwkt_schedule(thread_t td)
1070 _lwkt_schedule(td, 1);
1074 lwkt_schedule_noresched(thread_t td)
1076 _lwkt_schedule(td, 0);
1082 * Thread migration using a 'Pull' method. The thread may or may not be
1083 * the current thread. It MUST be descheduled and in a stable state.
1084 * lwkt_giveaway() must be called on the cpu owning the thread.
1086 * At any point after lwkt_giveaway() is called, the target cpu may
1087 * 'pull' the thread by calling lwkt_acquire().
1089 * MPSAFE - must be called under very specific conditions.
1092 lwkt_giveaway(thread_t td)
1094 globaldata_t gd = mycpu;
1097 KKASSERT(td->td_gd == gd);
1098 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1099 td->td_flags |= TDF_MIGRATING;
1104 lwkt_acquire(thread_t td)
1109 KKASSERT(td->td_flags & TDF_MIGRATING);
1114 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1115 crit_enter_gd(mygd);
1116 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1118 lwkt_process_ipiq();
1123 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1124 td->td_flags &= ~TDF_MIGRATING;
1127 crit_enter_gd(mygd);
1128 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1129 td->td_flags &= ~TDF_MIGRATING;
1137 * Generic deschedule. Descheduling threads other then your own should be
1138 * done only in carefully controlled circumstances. Descheduling is
1141 * This function may block if the cpu has run out of messages.
1144 lwkt_deschedule(thread_t td)
1148 if (td == curthread) {
1151 if (td->td_gd == mycpu) {
1154 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1164 * Set the target thread's priority. This routine does not automatically
1165 * switch to a higher priority thread, LWKT threads are not designed for
1166 * continuous priority changes. Yield if you want to switch.
1168 * We have to retain the critical section count which uses the high bits
1169 * of the td_pri field. The specified priority may also indicate zero or
1170 * more critical sections by adding TDPRI_CRIT*N.
1172 * Note that we requeue the thread whether it winds up on a different runq
1173 * or not. uio_yield() depends on this and the routine is not normally
1174 * called with the same priority otherwise.
1177 lwkt_setpri(thread_t td, int pri)
1180 KKASSERT(td->td_gd == mycpu);
1182 if (td->td_flags & TDF_RUNQ) {
1184 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1187 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1193 lwkt_setpri_self(int pri)
1195 thread_t td = curthread;
1197 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1199 if (td->td_flags & TDF_RUNQ) {
1201 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1204 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1210 * Migrate the current thread to the specified cpu.
1212 * This is accomplished by descheduling ourselves from the current cpu,
1213 * moving our thread to the tdallq of the target cpu, IPI messaging the
1214 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1215 * races while the thread is being migrated.
1218 static void lwkt_setcpu_remote(void *arg);
1222 lwkt_setcpu_self(globaldata_t rgd)
1225 thread_t td = curthread;
1227 if (td->td_gd != rgd) {
1228 crit_enter_quick(td);
1229 td->td_flags |= TDF_MIGRATING;
1230 lwkt_deschedule_self(td);
1231 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1232 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1234 /* we are now on the target cpu */
1235 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1236 crit_exit_quick(td);
1242 lwkt_migratecpu(int cpuid)
1247 rgd = globaldata_find(cpuid);
1248 lwkt_setcpu_self(rgd);
1253 * Remote IPI for cpu migration (called while in a critical section so we
1254 * do not have to enter another one). The thread has already been moved to
1255 * our cpu's allq, but we must wait for the thread to be completely switched
1256 * out on the originating cpu before we schedule it on ours or the stack
1257 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1258 * change to main memory.
1260 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1261 * against wakeups. It is best if this interface is used only when there
1262 * are no pending events that might try to schedule the thread.
1266 lwkt_setcpu_remote(void *arg)
1269 globaldata_t gd = mycpu;
1271 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1273 lwkt_process_ipiq();
1279 td->td_flags &= ~TDF_MIGRATING;
1280 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1286 lwkt_preempted_proc(void)
1288 thread_t td = curthread;
1289 while (td->td_preempted)
1290 td = td->td_preempted;
1295 * Create a kernel process/thread/whatever. It shares it's address space
1296 * with proc0 - ie: kernel only.
1298 * NOTE! By default new threads are created with the MP lock held. A
1299 * thread which does not require the MP lock should release it by calling
1300 * rel_mplock() at the start of the new thread.
1303 lwkt_create(void (*func)(void *), void *arg,
1304 struct thread **tdp, thread_t template, int tdflags, int cpu,
1305 const char *fmt, ...)
1310 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1314 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1317 * Set up arg0 for 'ps' etc
1319 __va_start(ap, fmt);
1320 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1324 * Schedule the thread to run
1326 if ((td->td_flags & TDF_STOPREQ) == 0)
1329 td->td_flags &= ~TDF_STOPREQ;
1334 * Destroy an LWKT thread. Warning! This function is not called when
1335 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1336 * uses a different reaping mechanism.
1341 thread_t td = curthread;
1345 if (td->td_flags & TDF_VERBOSE)
1346 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1350 * Get us into a critical section to interlock gd_freetd and loop
1351 * until we can get it freed.
1353 * We have to cache the current td in gd_freetd because objcache_put()ing
1354 * it would rip it out from under us while our thread is still active.
1357 crit_enter_quick(td);
1358 while ((std = gd->gd_freetd) != NULL) {
1359 gd->gd_freetd = NULL;
1360 objcache_put(thread_cache, std);
1362 lwkt_deschedule_self(td);
1363 lwkt_remove_tdallq(td);
1364 if (td->td_flags & TDF_ALLOCATED_THREAD)
1370 lwkt_remove_tdallq(thread_t td)
1372 KKASSERT(td->td_gd == mycpu);
1373 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1379 thread_t td = curthread;
1380 int lpri = td->td_pri;
1383 panic("td_pri is/would-go negative! %p %d", td, lpri);
1389 * Called from debugger/panic on cpus which have been stopped. We must still
1390 * process the IPIQ while stopped, even if we were stopped while in a critical
1393 * If we are dumping also try to process any pending interrupts. This may
1394 * or may not work depending on the state of the cpu at the point it was
1398 lwkt_smp_stopped(void)
1400 globaldata_t gd = mycpu;
1404 lwkt_process_ipiq();
1407 lwkt_process_ipiq();
1413 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1414 * get_mplock() has already incremented td_mpcount. We must block and
1415 * not return until giant is held.
1417 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1418 * reschedule the thread until it can obtain the giant lock for it.
1421 lwkt_mp_lock_contested(void)
1429 * The rel_mplock() code will call this function after releasing the
1430 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1432 * We then chain an IPI to a single other cpu potentially needing the
1433 * lock. This is a bit heuristical and we can wind up with IPIs flying
1434 * all over the place.
1436 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1439 lwkt_mp_lock_uncontested(void)
1449 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1450 mask = mp_lock_contention_mask;
1451 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1455 cpuid = bsfl(mask & tmpmask);
1458 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1459 dgd = globaldata_find(cpuid);
1460 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1466 * The idea is for this IPI to interrupt a potentially lower priority
1467 * thread, such as a user thread, to allow the scheduler to reschedule
1468 * a higher priority kernel thread that needs the MP lock.
1470 * For now we set the LWKT reschedule flag which generates an AST in
1471 * doreti, though theoretically it is also possible to possibly preempt
1472 * here if the underlying thread was operating in user mode. Nah.
1475 lwkt_mp_lock_uncontested_remote(void *arg __unused)
1477 need_lwkt_resched();