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;
92 static int chain_mplock = 0;
94 static struct objcache *thread_cache;
96 volatile cpumask_t mp_lock_contention_mask;
99 * We can make all thread ports use the spin backend instead of the thread
100 * backend. This should only be set to debug the spin backend.
102 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
104 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
106 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
109 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
112 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
117 &token_contention_count, 0, "spinning due to token contention");
118 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
119 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
125 #if !defined(KTR_GIANT_CONTENTION)
126 #define KTR_GIANT_CONTENTION KTR_ALL
129 KTR_INFO_MASTER(giant);
130 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
131 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
133 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
136 * These helper procedures handle the runq, they can only be called from
137 * within a critical section.
139 * WARNING! Prior to SMP being brought up it is possible to enqueue and
140 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
141 * instead of 'mycpu' when referencing the globaldata structure. Once
142 * SMP live enqueuing and dequeueing only occurs on the current cpu.
146 _lwkt_dequeue(thread_t td)
148 if (td->td_flags & TDF_RUNQ) {
149 int nq = td->td_pri & TDPRI_MASK;
150 struct globaldata *gd = td->td_gd;
152 td->td_flags &= ~TDF_RUNQ;
153 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
154 /* runqmask is passively cleaned up by the switcher */
160 _lwkt_enqueue(thread_t td)
162 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
163 int nq = td->td_pri & TDPRI_MASK;
164 struct globaldata *gd = td->td_gd;
166 td->td_flags |= TDF_RUNQ;
167 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
168 gd->gd_runqmask |= 1 << nq;
173 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
175 struct thread *td = (struct thread *)obj;
177 td->td_kstack = NULL;
178 td->td_kstack_size = 0;
179 td->td_flags = TDF_ALLOCATED_THREAD;
184 _lwkt_thread_dtor(void *obj, void *privdata)
186 struct thread *td = (struct thread *)obj;
188 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
189 ("_lwkt_thread_dtor: not allocated from objcache"));
190 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
191 td->td_kstack_size > 0,
192 ("_lwkt_thread_dtor: corrupted stack"));
193 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
197 * Initialize the lwkt s/system.
202 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
203 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
204 NULL, CACHE_NTHREADS/2,
205 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
209 * Schedule a thread to run. As the current thread we can always safely
210 * schedule ourselves, and a shortcut procedure is provided for that
213 * (non-blocking, self contained on a per cpu basis)
216 lwkt_schedule_self(thread_t td)
218 crit_enter_quick(td);
219 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
220 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
226 * Deschedule a thread.
228 * (non-blocking, self contained on a per cpu basis)
231 lwkt_deschedule_self(thread_t td)
233 crit_enter_quick(td);
239 * LWKTs operate on a per-cpu basis
241 * WARNING! Called from early boot, 'mycpu' may not work yet.
244 lwkt_gdinit(struct globaldata *gd)
248 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
249 TAILQ_INIT(&gd->gd_tdrunq[i]);
251 TAILQ_INIT(&gd->gd_tdallq);
255 * Create a new thread. The thread must be associated with a process context
256 * or LWKT start address before it can be scheduled. If the target cpu is
257 * -1 the thread will be created on the current cpu.
259 * If you intend to create a thread without a process context this function
260 * does everything except load the startup and switcher function.
263 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
265 globaldata_t gd = mycpu;
269 * If static thread storage is not supplied allocate a thread. Reuse
270 * a cached free thread if possible. gd_freetd is used to keep an exiting
271 * thread intact through the exit.
274 if ((td = gd->gd_freetd) != NULL)
275 gd->gd_freetd = NULL;
277 td = objcache_get(thread_cache, M_WAITOK);
278 KASSERT((td->td_flags &
279 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
280 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
281 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
285 * Try to reuse cached stack.
287 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
288 if (flags & TDF_ALLOCATED_STACK) {
289 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
294 stack = (void *)kmem_alloc(&kernel_map, stksize);
295 flags |= TDF_ALLOCATED_STACK;
298 lwkt_init_thread(td, stack, stksize, flags, gd);
300 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
305 * Initialize a preexisting thread structure. This function is used by
306 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
308 * All threads start out in a critical section at a priority of
309 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
310 * appropriate. This function may send an IPI message when the
311 * requested cpu is not the current cpu and consequently gd_tdallq may
312 * not be initialized synchronously from the point of view of the originating
315 * NOTE! we have to be careful in regards to creating threads for other cpus
316 * if SMP has not yet been activated.
321 lwkt_init_thread_remote(void *arg)
326 * Protected by critical section held by IPI dispatch
328 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
334 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
335 struct globaldata *gd)
337 globaldata_t mygd = mycpu;
339 bzero(td, sizeof(struct thread));
340 td->td_kstack = stack;
341 td->td_kstack_size = stksize;
342 td->td_flags = flags;
344 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
346 if ((flags & TDF_MPSAFE) == 0)
349 if (lwkt_use_spin_port)
350 lwkt_initport_spin(&td->td_msgport);
352 lwkt_initport_thread(&td->td_msgport, td);
353 pmap_init_thread(td);
356 * Normally initializing a thread for a remote cpu requires sending an
357 * IPI. However, the idlethread is setup before the other cpus are
358 * activated so we have to treat it as a special case. XXX manipulation
359 * of gd_tdallq requires the BGL.
361 if (gd == mygd || td == &gd->gd_idlethread) {
363 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
366 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
370 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
376 lwkt_set_comm(thread_t td, const char *ctl, ...)
381 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
386 lwkt_hold(thread_t td)
392 lwkt_rele(thread_t td)
394 KKASSERT(td->td_refs > 0);
399 lwkt_wait_free(thread_t td)
402 tsleep(td, 0, "tdreap", hz);
406 lwkt_free_thread(thread_t td)
408 KASSERT((td->td_flags & TDF_RUNNING) == 0,
409 ("lwkt_free_thread: did not exit! %p", td));
411 if (td->td_flags & TDF_ALLOCATED_THREAD) {
412 objcache_put(thread_cache, td);
413 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
414 /* client-allocated struct with internally allocated stack */
415 KASSERT(td->td_kstack && td->td_kstack_size > 0,
416 ("lwkt_free_thread: corrupted stack"));
417 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
418 td->td_kstack = NULL;
419 td->td_kstack_size = 0;
425 * Switch to the next runnable lwkt. If no LWKTs are runnable then
426 * switch to the idlethread. Switching must occur within a critical
427 * section to avoid races with the scheduling queue.
429 * We always have full control over our cpu's run queue. Other cpus
430 * that wish to manipulate our queue must use the cpu_*msg() calls to
431 * talk to our cpu, so a critical section is all that is needed and
432 * the result is very, very fast thread switching.
434 * The LWKT scheduler uses a fixed priority model and round-robins at
435 * each priority level. User process scheduling is a totally
436 * different beast and LWKT priorities should not be confused with
437 * user process priorities.
439 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
440 * cleans it up. Note that the td_switch() function cannot do anything that
441 * requires the MP lock since the MP lock will have already been setup for
442 * the target thread (not the current thread). It's nice to have a scheduler
443 * that does not need the MP lock to work because it allows us to do some
444 * really cool high-performance MP lock optimizations.
446 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
447 * is not called by the current thread in the preemption case, only when
448 * the preempting thread blocks (in order to return to the original thread).
453 globaldata_t gd = mycpu;
454 thread_t td = gd->gd_curthread;
461 * Switching from within a 'fast' (non thread switched) interrupt or IPI
462 * is illegal. However, we may have to do it anyway if we hit a fatal
463 * kernel trap or we have paniced.
465 * If this case occurs save and restore the interrupt nesting level.
467 if (gd->gd_intr_nesting_level) {
471 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
472 panic("lwkt_switch: cannot switch from within "
473 "a fast interrupt, yet, td %p\n", td);
475 savegdnest = gd->gd_intr_nesting_level;
476 savegdtrap = gd->gd_trap_nesting_level;
477 gd->gd_intr_nesting_level = 0;
478 gd->gd_trap_nesting_level = 0;
479 if ((td->td_flags & TDF_PANICWARN) == 0) {
480 td->td_flags |= TDF_PANICWARN;
481 kprintf("Warning: thread switch from interrupt or IPI, "
482 "thread %p (%s)\n", td, td->td_comm);
484 db_print_backtrace();
488 gd->gd_intr_nesting_level = savegdnest;
489 gd->gd_trap_nesting_level = savegdtrap;
495 * Passive release (used to transition from user to kernel mode
496 * when we block or switch rather then when we enter the kernel).
497 * This function is NOT called if we are switching into a preemption
498 * or returning from a preemption. Typically this causes us to lose
499 * our current process designation (if we have one) and become a true
500 * LWKT thread, and may also hand the current process designation to
501 * another process and schedule thread.
508 lwkt_relalltokens(td);
511 * We had better not be holding any spin locks, but don't get into an
512 * endless panic loop.
514 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
515 ("lwkt_switch: still holding a shared spinlock %p!",
516 gd->gd_spinlock_rd));
517 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
518 ("lwkt_switch: still holding %d exclusive spinlocks!",
519 gd->gd_spinlocks_wr));
524 * td_mpcount cannot be used to determine if we currently hold the
525 * MP lock because get_mplock() will increment it prior to attempting
526 * to get the lock, and switch out if it can't. Our ownership of
527 * the actual lock will remain stable while we are in a critical section
528 * (but, of course, another cpu may own or release the lock so the
529 * actual value of mp_lock is not stable).
531 mpheld = MP_LOCK_HELD();
533 if (td->td_cscount) {
534 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
536 if (panic_on_cscount)
537 panic("switching while mastering cpusync");
541 if ((ntd = td->td_preempted) != NULL) {
543 * We had preempted another thread on this cpu, resume the preempted
544 * thread. This occurs transparently, whether the preempted thread
545 * was scheduled or not (it may have been preempted after descheduling
548 * We have to setup the MP lock for the original thread after backing
549 * out the adjustment that was made to curthread when the original
552 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
554 if (ntd->td_mpcount && mpheld == 0) {
555 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
556 td, ntd, td->td_mpcount, ntd->td_mpcount);
558 if (ntd->td_mpcount) {
559 td->td_mpcount -= ntd->td_mpcount;
560 KKASSERT(td->td_mpcount >= 0);
563 ntd->td_flags |= TDF_PREEMPT_DONE;
566 * The interrupt may have woken a thread up, we need to properly
567 * set the reschedule flag if the originally interrupted thread is
568 * at a lower priority.
570 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
572 /* YYY release mp lock on switchback if original doesn't need it */
575 * Priority queue / round-robin at each priority. Note that user
576 * processes run at a fixed, low priority and the user process
577 * scheduler deals with interactions between user processes
578 * by scheduling and descheduling them from the LWKT queue as
581 * We have to adjust the MP lock for the target thread. If we
582 * need the MP lock and cannot obtain it we try to locate a
583 * thread that does not need the MP lock. If we cannot, we spin
586 * A similar issue exists for the tokens held by the target thread.
587 * If we cannot obtain ownership of the tokens we cannot immediately
588 * schedule the thread.
592 * If an LWKT reschedule was requested, well that is what we are
593 * doing now so clear it.
595 clear_lwkt_resched();
597 if (gd->gd_runqmask) {
598 int nq = bsrl(gd->gd_runqmask);
599 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
600 gd->gd_runqmask &= ~(1 << nq);
605 * THREAD SELECTION FOR AN SMP MACHINE BUILD
607 * If the target needs the MP lock and we couldn't get it,
608 * or if the target is holding tokens and we could not
609 * gain ownership of the tokens, continue looking for a
610 * thread to schedule and spin instead of HLT if we can't.
612 * NOTE: the mpheld variable invalid after this conditional, it
613 * can change due to both cpu_try_mplock() returning success
614 * AND interactions in lwkt_getalltokens() due to the fact that
615 * we are trying to check the mpcount of a thread other then
616 * the current thread. Because of this, if the current thread
617 * is not holding td_mpcount, an IPI indirectly run via
618 * lwkt_getalltokens() can obtain and release the MP lock and
619 * cause the core MP lock to be released.
621 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
622 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
624 u_int32_t rqmask = gd->gd_runqmask;
626 mpheld = MP_LOCK_HELD();
629 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
630 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
631 /* spinning due to MP lock being held */
633 ++mplock_contention_count;
635 /* mplock still not held, 'mpheld' still valid */
640 * mpheld state invalid after getalltokens call returns
641 * failure, but the variable is only needed for
644 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
645 /* spinning due to token contention */
647 ++token_contention_count;
649 mpheld = MP_LOCK_HELD();
656 rqmask &= ~(1 << nq);
660 * We have two choices. We can either refuse to run a
661 * user thread when a kernel thread needs the MP lock
662 * but could not get it, or we can allow it to run but
663 * then expect an IPI (hopefully) later on to force a
664 * reschedule when the MP lock might become available.
666 if (nq < TDPRI_KERN_LPSCHED) {
667 if (chain_mplock == 0)
669 atomic_set_int(&mp_lock_contention_mask,
671 /* continue loop, allow user threads to be scheduled */
675 cpu_mplock_contested();
676 ntd = &gd->gd_idlethread;
677 ntd->td_flags |= TDF_IDLE_NOHLT;
678 goto using_idle_thread;
680 ++gd->gd_cnt.v_swtch;
681 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
682 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
685 ++gd->gd_cnt.v_swtch;
686 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
687 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
691 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
692 * worry about tokens or the BGL. However, we still have
693 * to call lwkt_getalltokens() in order to properly detect
694 * stale tokens. This call cannot fail for a UP build!
696 lwkt_getalltokens(ntd);
697 ++gd->gd_cnt.v_swtch;
698 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
699 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 * We have nothing to run but only let the idle loop halt
704 * the cpu if there are no pending interrupts.
706 ntd = &gd->gd_idlethread;
707 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
708 ntd->td_flags |= TDF_IDLE_NOHLT;
712 * The idle thread should not be holding the MP lock unless we
713 * are trapping in the kernel or in a panic. Since we select the
714 * idle thread unconditionally when no other thread is available,
715 * if the MP lock is desired during a panic or kernel trap, we
716 * have to loop in the scheduler until we get it.
718 if (ntd->td_mpcount) {
719 mpheld = MP_LOCK_HELD();
720 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
721 panic("Idle thread %p was holding the BGL!", ntd);
722 } else if (mpheld == 0) {
723 cpu_mplock_contested();
730 KASSERT(ntd->td_pri >= TDPRI_CRIT,
731 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
734 * Do the actual switch. If the new target does not need the MP lock
735 * and we are holding it, release the MP lock. If the new target requires
736 * the MP lock we have already acquired it for the target.
739 if (ntd->td_mpcount == 0 ) {
743 ASSERT_MP_LOCK_HELD(ntd);
750 /* NOTE: current cpu may have changed after switch */
755 * Request that the target thread preempt the current thread. Preemption
756 * only works under a specific set of conditions:
758 * - We are not preempting ourselves
759 * - The target thread is owned by the current cpu
760 * - We are not currently being preempted
761 * - The target is not currently being preempted
762 * - We are not holding any spin locks
763 * - The target thread is not holding any tokens
764 * - We are able to satisfy the target's MP lock requirements (if any).
766 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
767 * this is called via lwkt_schedule() through the td_preemptable callback.
768 * critpri is the managed critical priority that we should ignore in order
769 * to determine whether preemption is possible (aka usually just the crit
770 * priority of lwkt_schedule() itself).
772 * XXX at the moment we run the target thread in a critical section during
773 * the preemption in order to prevent the target from taking interrupts
774 * that *WE* can't. Preemption is strictly limited to interrupt threads
775 * and interrupt-like threads, outside of a critical section, and the
776 * preempted source thread will be resumed the instant the target blocks
777 * whether or not the source is scheduled (i.e. preemption is supposed to
778 * be as transparent as possible).
780 * The target thread inherits our MP count (added to its own) for the
781 * duration of the preemption in order to preserve the atomicy of the
782 * MP lock during the preemption. Therefore, any preempting targets must be
783 * careful in regards to MP assertions. Note that the MP count may be
784 * out of sync with the physical mp_lock, but we do not have to preserve
785 * the original ownership of the lock if it was out of synch (that is, we
786 * can leave it synchronized on return).
789 lwkt_preempt(thread_t ntd, int critpri)
791 struct globaldata *gd = mycpu;
799 * The caller has put us in a critical section. We can only preempt
800 * if the caller of the caller was not in a critical section (basically
801 * a local interrupt), as determined by the 'critpri' parameter. We
802 * also can't preempt if the caller is holding any spinlocks (even if
803 * he isn't in a critical section). This also handles the tokens test.
805 * YYY The target thread must be in a critical section (else it must
806 * inherit our critical section? I dunno yet).
808 * Set need_lwkt_resched() unconditionally for now YYY.
810 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
812 td = gd->gd_curthread;
813 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
817 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
823 if (ntd->td_gd != gd) {
830 * Take the easy way out and do not preempt if we are holding
831 * any spinlocks. We could test whether the thread(s) being
832 * preempted interlock against the target thread's tokens and whether
833 * we can get all the target thread's tokens, but this situation
834 * should not occur very often so its easier to simply not preempt.
835 * Also, plain spinlocks are impossible to figure out at this point so
836 * just don't preempt.
838 * Do not try to preempt if the target thread is holding any tokens.
839 * We could try to acquire the tokens but this case is so rare there
840 * is no need to support it.
842 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
852 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
857 if (ntd->td_preempted) {
864 * note: an interrupt might have occured just as we were transitioning
865 * to or from the MP lock. In this case td_mpcount will be pre-disposed
866 * (non-zero) but not actually synchronized with the actual state of the
867 * lock. We can use it to imply an MP lock requirement for the
868 * preemption but we cannot use it to test whether we hold the MP lock
871 savecnt = td->td_mpcount;
872 mpheld = MP_LOCK_HELD();
873 ntd->td_mpcount += td->td_mpcount;
874 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
875 ntd->td_mpcount -= td->td_mpcount;
883 * Since we are able to preempt the current thread, there is no need to
884 * call need_lwkt_resched().
887 ntd->td_preempted = td;
888 td->td_flags |= TDF_PREEMPT_LOCK;
891 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
893 KKASSERT(savecnt == td->td_mpcount);
894 mpheld = MP_LOCK_HELD();
895 if (mpheld && td->td_mpcount == 0)
897 else if (mpheld == 0 && td->td_mpcount)
898 panic("lwkt_preempt(): MP lock was not held through");
900 ntd->td_preempted = NULL;
901 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
905 * Yield our thread while higher priority threads are pending. This is
906 * typically called when we leave a critical section but it can be safely
907 * called while we are in a critical section.
909 * This function will not generally yield to equal priority threads but it
910 * can occur as a side effect. Note that lwkt_switch() is called from
911 * inside the critical section to prevent its own crit_exit() from reentering
912 * lwkt_yield_quick().
914 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
915 * came along but was blocked and made pending.
917 * (self contained on a per cpu basis)
920 lwkt_yield_quick(void)
922 globaldata_t gd = mycpu;
923 thread_t td = gd->gd_curthread;
926 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
927 * it with a non-zero cpl then we might not wind up calling splz after
928 * a task switch when the critical section is exited even though the
929 * new task could accept the interrupt.
931 * XXX from crit_exit() only called after last crit section is released.
932 * If called directly will run splz() even if in a critical section.
934 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
935 * except for this special case, we MUST call splz() here to handle any
936 * pending ints, particularly after we switch, or we might accidently
937 * halt the cpu with interrupts pending.
939 if (gd->gd_reqflags && td->td_nest_count < 2)
943 * YYY enabling will cause wakeup() to task-switch, which really
944 * confused the old 4.x code. This is a good way to simulate
945 * preemption and MP without actually doing preemption or MP, because a
946 * lot of code assumes that wakeup() does not block.
948 if (untimely_switch && td->td_nest_count == 0 &&
949 gd->gd_intr_nesting_level == 0
951 crit_enter_quick(td);
953 * YYY temporary hacks until we disassociate the userland scheduler
954 * from the LWKT scheduler.
956 if (td->td_flags & TDF_RUNQ) {
957 lwkt_switch(); /* will not reenter yield function */
959 lwkt_schedule_self(td); /* make sure we are scheduled */
960 lwkt_switch(); /* will not reenter yield function */
961 lwkt_deschedule_self(td); /* make sure we are descheduled */
963 crit_exit_noyield(td);
968 * This implements a normal yield which, unlike _quick, will yield to equal
969 * priority threads as well. Note that gd_reqflags tests will be handled by
970 * the crit_exit() call in lwkt_switch().
972 * (self contained on a per cpu basis)
977 lwkt_schedule_self(curthread);
982 * Return 0 if no runnable threads are pending at the same or higher
983 * priority as the passed thread.
985 * Return 1 if runnable threads are pending at the same priority.
987 * Return 2 if runnable threads are pending at a higher priority.
990 lwkt_check_resched(thread_t td)
992 int pri = td->td_pri & TDPRI_MASK;
994 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
996 if (TAILQ_NEXT(td, td_threadq))
1002 * Generic schedule. Possibly schedule threads belonging to other cpus and
1003 * deal with threads that might be blocked on a wait queue.
1005 * We have a little helper inline function which does additional work after
1006 * the thread has been enqueued, including dealing with preemption and
1007 * setting need_lwkt_resched() (which prevents the kernel from returning
1008 * to userland until it has processed higher priority threads).
1010 * It is possible for this routine to be called after a failed _enqueue
1011 * (due to the target thread migrating, sleeping, or otherwise blocked).
1012 * We have to check that the thread is actually on the run queue!
1014 * reschedok is an optimized constant propagated from lwkt_schedule() or
1015 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1016 * reschedule to be requested if the target thread has a higher priority.
1017 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1018 * be 0, prevented undesired reschedules.
1022 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1026 if (ntd->td_flags & TDF_RUNQ) {
1027 if (ntd->td_preemptable && reschedok) {
1028 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1029 } else if (reschedok) {
1031 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1032 need_lwkt_resched();
1039 _lwkt_schedule(thread_t td, int reschedok)
1041 globaldata_t mygd = mycpu;
1043 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1044 crit_enter_gd(mygd);
1045 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1046 if (td == mygd->gd_curthread) {
1050 * If we own the thread, there is no race (since we are in a
1051 * critical section). If we do not own the thread there might
1052 * be a race but the target cpu will deal with it.
1055 if (td->td_gd == mygd) {
1057 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1059 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1063 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1070 lwkt_schedule(thread_t td)
1072 _lwkt_schedule(td, 1);
1076 lwkt_schedule_noresched(thread_t td)
1078 _lwkt_schedule(td, 0);
1084 * Thread migration using a 'Pull' method. The thread may or may not be
1085 * the current thread. It MUST be descheduled and in a stable state.
1086 * lwkt_giveaway() must be called on the cpu owning the thread.
1088 * At any point after lwkt_giveaway() is called, the target cpu may
1089 * 'pull' the thread by calling lwkt_acquire().
1091 * MPSAFE - must be called under very specific conditions.
1094 lwkt_giveaway(thread_t td)
1096 globaldata_t gd = mycpu;
1099 KKASSERT(td->td_gd == gd);
1100 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1101 td->td_flags |= TDF_MIGRATING;
1106 lwkt_acquire(thread_t td)
1111 KKASSERT(td->td_flags & TDF_MIGRATING);
1116 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1117 crit_enter_gd(mygd);
1118 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1120 lwkt_process_ipiq();
1125 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1126 td->td_flags &= ~TDF_MIGRATING;
1129 crit_enter_gd(mygd);
1130 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1131 td->td_flags &= ~TDF_MIGRATING;
1139 * Generic deschedule. Descheduling threads other then your own should be
1140 * done only in carefully controlled circumstances. Descheduling is
1143 * This function may block if the cpu has run out of messages.
1146 lwkt_deschedule(thread_t td)
1150 if (td == curthread) {
1153 if (td->td_gd == mycpu) {
1156 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1166 * Set the target thread's priority. This routine does not automatically
1167 * switch to a higher priority thread, LWKT threads are not designed for
1168 * continuous priority changes. Yield if you want to switch.
1170 * We have to retain the critical section count which uses the high bits
1171 * of the td_pri field. The specified priority may also indicate zero or
1172 * more critical sections by adding TDPRI_CRIT*N.
1174 * Note that we requeue the thread whether it winds up on a different runq
1175 * or not. uio_yield() depends on this and the routine is not normally
1176 * called with the same priority otherwise.
1179 lwkt_setpri(thread_t td, int pri)
1182 KKASSERT(td->td_gd == mycpu);
1184 if (td->td_flags & TDF_RUNQ) {
1186 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1189 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1195 lwkt_setpri_self(int pri)
1197 thread_t td = curthread;
1199 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1201 if (td->td_flags & TDF_RUNQ) {
1203 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1206 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1212 * Migrate the current thread to the specified cpu.
1214 * This is accomplished by descheduling ourselves from the current cpu,
1215 * moving our thread to the tdallq of the target cpu, IPI messaging the
1216 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1217 * races while the thread is being migrated.
1220 static void lwkt_setcpu_remote(void *arg);
1224 lwkt_setcpu_self(globaldata_t rgd)
1227 thread_t td = curthread;
1229 if (td->td_gd != rgd) {
1230 crit_enter_quick(td);
1231 td->td_flags |= TDF_MIGRATING;
1232 lwkt_deschedule_self(td);
1233 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1234 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1236 /* we are now on the target cpu */
1237 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1238 crit_exit_quick(td);
1244 lwkt_migratecpu(int cpuid)
1249 rgd = globaldata_find(cpuid);
1250 lwkt_setcpu_self(rgd);
1255 * Remote IPI for cpu migration (called while in a critical section so we
1256 * do not have to enter another one). The thread has already been moved to
1257 * our cpu's allq, but we must wait for the thread to be completely switched
1258 * out on the originating cpu before we schedule it on ours or the stack
1259 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1260 * change to main memory.
1262 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1263 * against wakeups. It is best if this interface is used only when there
1264 * are no pending events that might try to schedule the thread.
1268 lwkt_setcpu_remote(void *arg)
1271 globaldata_t gd = mycpu;
1273 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1275 lwkt_process_ipiq();
1281 td->td_flags &= ~TDF_MIGRATING;
1282 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1288 lwkt_preempted_proc(void)
1290 thread_t td = curthread;
1291 while (td->td_preempted)
1292 td = td->td_preempted;
1297 * Create a kernel process/thread/whatever. It shares it's address space
1298 * with proc0 - ie: kernel only.
1300 * NOTE! By default new threads are created with the MP lock held. A
1301 * thread which does not require the MP lock should release it by calling
1302 * rel_mplock() at the start of the new thread.
1305 lwkt_create(void (*func)(void *), void *arg,
1306 struct thread **tdp, thread_t template, int tdflags, int cpu,
1307 const char *fmt, ...)
1312 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1316 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1319 * Set up arg0 for 'ps' etc
1321 __va_start(ap, fmt);
1322 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1326 * Schedule the thread to run
1328 if ((td->td_flags & TDF_STOPREQ) == 0)
1331 td->td_flags &= ~TDF_STOPREQ;
1336 * Destroy an LWKT thread. Warning! This function is not called when
1337 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1338 * uses a different reaping mechanism.
1343 thread_t td = curthread;
1347 if (td->td_flags & TDF_VERBOSE)
1348 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1352 * Get us into a critical section to interlock gd_freetd and loop
1353 * until we can get it freed.
1355 * We have to cache the current td in gd_freetd because objcache_put()ing
1356 * it would rip it out from under us while our thread is still active.
1359 crit_enter_quick(td);
1360 while ((std = gd->gd_freetd) != NULL) {
1361 gd->gd_freetd = NULL;
1362 objcache_put(thread_cache, std);
1364 lwkt_deschedule_self(td);
1365 lwkt_remove_tdallq(td);
1366 if (td->td_flags & TDF_ALLOCATED_THREAD)
1372 lwkt_remove_tdallq(thread_t td)
1374 KKASSERT(td->td_gd == mycpu);
1375 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1381 thread_t td = curthread;
1382 int lpri = td->td_pri;
1385 panic("td_pri is/would-go negative! %p %d", td, lpri);
1391 * Called from debugger/panic on cpus which have been stopped. We must still
1392 * process the IPIQ while stopped, even if we were stopped while in a critical
1395 * If we are dumping also try to process any pending interrupts. This may
1396 * or may not work depending on the state of the cpu at the point it was
1400 lwkt_smp_stopped(void)
1402 globaldata_t gd = mycpu;
1406 lwkt_process_ipiq();
1409 lwkt_process_ipiq();
1415 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1416 * get_mplock() has already incremented td_mpcount. We must block and
1417 * not return until giant is held.
1419 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1420 * reschedule the thread until it can obtain the giant lock for it.
1423 lwkt_mp_lock_contested(void)
1431 * The rel_mplock() code will call this function after releasing the
1432 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1434 * We then chain an IPI to a single other cpu potentially needing the
1435 * lock. This is a bit heuristical and we can wind up with IPIs flying
1436 * all over the place.
1438 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1441 lwkt_mp_lock_uncontested(void)
1451 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1452 mask = mp_lock_contention_mask;
1453 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1457 cpuid = bsfl(mask & tmpmask);
1460 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1461 dgd = globaldata_find(cpuid);
1462 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1468 * The idea is for this IPI to interrupt a potentially lower priority
1469 * thread, such as a user thread, to allow the scheduler to reschedule
1470 * a higher priority kernel thread that needs the MP lock.
1472 * For now we set the LWKT reschedule flag which generates an AST in
1473 * doreti, though theoretically it is also possible to possibly preempt
1474 * here if the underlying thread was operating in user mode. Nah.
1477 lwkt_mp_lock_uncontested_remote(void *arg __unused)
1479 need_lwkt_resched();