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;
98 extern void cpu_heavy_restore(void);
99 extern void cpu_lwkt_restore(void);
100 extern void cpu_kthread_restore(void);
101 extern void cpu_idle_restore(void);
104 jg_tos_ok(struct thread *td)
109 KKASSERT(td->td_sp != NULL);
110 unsigned long tos = ((unsigned long *)td->td_sp)[0];
112 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore)
113 || (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
120 * We can make all thread ports use the spin backend instead of the thread
121 * backend. This should only be set to debug the spin backend.
123 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
125 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
127 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
130 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
132 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
133 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
134 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
135 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
137 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
138 &token_contention_count, 0, "spinning due to token contention");
139 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
140 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
146 #if !defined(KTR_GIANT_CONTENTION)
147 #define KTR_GIANT_CONTENTION KTR_ALL
150 KTR_INFO_MASTER(giant);
151 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
152 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
154 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
157 * These helper procedures handle the runq, they can only be called from
158 * within a critical section.
160 * WARNING! Prior to SMP being brought up it is possible to enqueue and
161 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
162 * instead of 'mycpu' when referencing the globaldata structure. Once
163 * SMP live enqueuing and dequeueing only occurs on the current cpu.
167 _lwkt_dequeue(thread_t td)
169 if (td->td_flags & TDF_RUNQ) {
170 int nq = td->td_pri & TDPRI_MASK;
171 struct globaldata *gd = td->td_gd;
173 td->td_flags &= ~TDF_RUNQ;
174 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
175 /* runqmask is passively cleaned up by the switcher */
181 _lwkt_enqueue(thread_t td)
183 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
184 int nq = td->td_pri & TDPRI_MASK;
185 struct globaldata *gd = td->td_gd;
187 td->td_flags |= TDF_RUNQ;
188 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
189 gd->gd_runqmask |= 1 << nq;
194 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
196 struct thread *td = (struct thread *)obj;
198 td->td_kstack = NULL;
199 td->td_kstack_size = 0;
200 td->td_flags = TDF_ALLOCATED_THREAD;
205 _lwkt_thread_dtor(void *obj, void *privdata)
207 struct thread *td = (struct thread *)obj;
209 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
210 ("_lwkt_thread_dtor: not allocated from objcache"));
211 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
212 td->td_kstack_size > 0,
213 ("_lwkt_thread_dtor: corrupted stack"));
214 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
218 * Initialize the lwkt s/system.
223 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
224 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
225 NULL, CACHE_NTHREADS/2,
226 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
230 * Schedule a thread to run. As the current thread we can always safely
231 * schedule ourselves, and a shortcut procedure is provided for that
234 * (non-blocking, self contained on a per cpu basis)
237 lwkt_schedule_self(thread_t td)
239 crit_enter_quick(td);
240 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
241 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
247 * Deschedule a thread.
249 * (non-blocking, self contained on a per cpu basis)
252 lwkt_deschedule_self(thread_t td)
254 crit_enter_quick(td);
260 * LWKTs operate on a per-cpu basis
262 * WARNING! Called from early boot, 'mycpu' may not work yet.
265 lwkt_gdinit(struct globaldata *gd)
269 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
270 TAILQ_INIT(&gd->gd_tdrunq[i]);
272 TAILQ_INIT(&gd->gd_tdallq);
276 * Create a new thread. The thread must be associated with a process context
277 * or LWKT start address before it can be scheduled. If the target cpu is
278 * -1 the thread will be created on the current cpu.
280 * If you intend to create a thread without a process context this function
281 * does everything except load the startup and switcher function.
284 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
286 globaldata_t gd = mycpu;
290 * If static thread storage is not supplied allocate a thread. Reuse
291 * a cached free thread if possible. gd_freetd is used to keep an exiting
292 * thread intact through the exit.
295 if ((td = gd->gd_freetd) != NULL)
296 gd->gd_freetd = NULL;
298 td = objcache_get(thread_cache, M_WAITOK);
299 KASSERT((td->td_flags &
300 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
301 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
302 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
306 * Try to reuse cached stack.
308 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
309 if (flags & TDF_ALLOCATED_STACK) {
310 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
315 stack = (void *)kmem_alloc(&kernel_map, stksize);
316 flags |= TDF_ALLOCATED_STACK;
319 lwkt_init_thread(td, stack, stksize, flags, gd);
321 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
326 * Initialize a preexisting thread structure. This function is used by
327 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
329 * All threads start out in a critical section at a priority of
330 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
331 * appropriate. This function may send an IPI message when the
332 * requested cpu is not the current cpu and consequently gd_tdallq may
333 * not be initialized synchronously from the point of view of the originating
336 * NOTE! we have to be careful in regards to creating threads for other cpus
337 * if SMP has not yet been activated.
342 lwkt_init_thread_remote(void *arg)
347 * Protected by critical section held by IPI dispatch
349 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
355 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
356 struct globaldata *gd)
358 globaldata_t mygd = mycpu;
360 bzero(td, sizeof(struct thread));
361 td->td_kstack = stack;
362 td->td_kstack_size = stksize;
363 td->td_flags = flags;
365 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
367 if ((flags & TDF_MPSAFE) == 0)
370 if (lwkt_use_spin_port)
371 lwkt_initport_spin(&td->td_msgport);
373 lwkt_initport_thread(&td->td_msgport, td);
374 pmap_init_thread(td);
377 * Normally initializing a thread for a remote cpu requires sending an
378 * IPI. However, the idlethread is setup before the other cpus are
379 * activated so we have to treat it as a special case. XXX manipulation
380 * of gd_tdallq requires the BGL.
382 if (gd == mygd || td == &gd->gd_idlethread) {
384 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
387 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
391 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
397 lwkt_set_comm(thread_t td, const char *ctl, ...)
402 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
407 lwkt_hold(thread_t td)
413 lwkt_rele(thread_t td)
415 KKASSERT(td->td_refs > 0);
420 lwkt_wait_free(thread_t td)
423 tsleep(td, 0, "tdreap", hz);
427 lwkt_free_thread(thread_t td)
429 KASSERT((td->td_flags & TDF_RUNNING) == 0,
430 ("lwkt_free_thread: did not exit! %p", td));
432 if (td->td_flags & TDF_ALLOCATED_THREAD) {
433 objcache_put(thread_cache, td);
434 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
435 /* client-allocated struct with internally allocated stack */
436 KASSERT(td->td_kstack && td->td_kstack_size > 0,
437 ("lwkt_free_thread: corrupted stack"));
438 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
439 td->td_kstack = NULL;
440 td->td_kstack_size = 0;
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);
505 db_print_backtrace();
509 gd->gd_intr_nesting_level = savegdnest;
510 gd->gd_trap_nesting_level = savegdtrap;
516 * Passive release (used to transition from user to kernel mode
517 * when we block or switch rather then when we enter the kernel).
518 * This function is NOT called if we are switching into a preemption
519 * or returning from a preemption. Typically this causes us to lose
520 * our current process designation (if we have one) and become a true
521 * LWKT thread, and may also hand the current process designation to
522 * another process and schedule thread.
529 lwkt_relalltokens(td);
532 * We had better not be holding any spin locks, but don't get into an
533 * endless panic loop.
535 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
536 ("lwkt_switch: still holding a shared spinlock %p!",
537 gd->gd_spinlock_rd));
538 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
539 ("lwkt_switch: still holding %d exclusive spinlocks!",
540 gd->gd_spinlocks_wr));
545 * td_mpcount cannot be used to determine if we currently hold the
546 * MP lock because get_mplock() will increment it prior to attempting
547 * to get the lock, and switch out if it can't. Our ownership of
548 * the actual lock will remain stable while we are in a critical section
549 * (but, of course, another cpu may own or release the lock so the
550 * actual value of mp_lock is not stable).
552 mpheld = MP_LOCK_HELD();
554 if (td->td_cscount) {
555 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
557 if (panic_on_cscount)
558 panic("switching while mastering cpusync");
562 if ((ntd = td->td_preempted) != NULL) {
564 * We had preempted another thread on this cpu, resume the preempted
565 * thread. This occurs transparently, whether the preempted thread
566 * was scheduled or not (it may have been preempted after descheduling
569 * We have to setup the MP lock for the original thread after backing
570 * out the adjustment that was made to curthread when the original
573 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
575 if (ntd->td_mpcount && mpheld == 0) {
576 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
577 td, ntd, td->td_mpcount, ntd->td_mpcount);
579 if (ntd->td_mpcount) {
580 td->td_mpcount -= ntd->td_mpcount;
581 KKASSERT(td->td_mpcount >= 0);
584 ntd->td_flags |= TDF_PREEMPT_DONE;
587 * The interrupt may have woken a thread up, we need to properly
588 * set the reschedule flag if the originally interrupted thread is
589 * at a lower priority.
591 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
593 /* YYY release mp lock on switchback if original doesn't need it */
596 * Priority queue / round-robin at each priority. Note that user
597 * processes run at a fixed, low priority and the user process
598 * scheduler deals with interactions between user processes
599 * by scheduling and descheduling them from the LWKT queue as
602 * We have to adjust the MP lock for the target thread. If we
603 * need the MP lock and cannot obtain it we try to locate a
604 * thread that does not need the MP lock. If we cannot, we spin
607 * A similar issue exists for the tokens held by the target thread.
608 * If we cannot obtain ownership of the tokens we cannot immediately
609 * schedule the thread.
613 * If an LWKT reschedule was requested, well that is what we are
614 * doing now so clear it.
616 clear_lwkt_resched();
618 if (gd->gd_runqmask) {
619 int nq = bsrl(gd->gd_runqmask);
620 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
621 gd->gd_runqmask &= ~(1 << nq);
626 * THREAD SELECTION FOR AN SMP MACHINE BUILD
628 * If the target needs the MP lock and we couldn't get it,
629 * or if the target is holding tokens and we could not
630 * gain ownership of the tokens, continue looking for a
631 * thread to schedule and spin instead of HLT if we can't.
633 * NOTE: the mpheld variable invalid after this conditional, it
634 * can change due to both cpu_try_mplock() returning success
635 * AND interactions in lwkt_getalltokens() due to the fact that
636 * we are trying to check the mpcount of a thread other then
637 * the current thread. Because of this, if the current thread
638 * is not holding td_mpcount, an IPI indirectly run via
639 * lwkt_getalltokens() can obtain and release the MP lock and
640 * cause the core MP lock to be released.
642 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
643 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
645 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 */
654 ++mplock_contention_count;
656 /* mplock still not held, 'mpheld' still valid */
661 * mpheld state invalid after getalltokens call returns
662 * failure, but the variable is only needed for
665 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
666 /* spinning due to token contention */
668 ++token_contention_count;
670 mpheld = MP_LOCK_HELD();
677 rqmask &= ~(1 << nq);
681 * We have two choices. We can either refuse to run a
682 * user thread when a kernel thread needs the MP lock
683 * but could not get it, or we can allow it to run but
684 * then expect an IPI (hopefully) later on to force a
685 * reschedule when the MP lock might become available.
687 if (nq < TDPRI_KERN_LPSCHED) {
688 if (chain_mplock == 0)
690 atomic_set_int(&mp_lock_contention_mask,
692 /* continue loop, allow user threads to be scheduled */
696 cpu_mplock_contested();
697 ntd = &gd->gd_idlethread;
698 ntd->td_flags |= TDF_IDLE_NOHLT;
699 goto using_idle_thread;
701 ++gd->gd_cnt.v_swtch;
702 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
703 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
706 ++gd->gd_cnt.v_swtch;
707 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
708 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
713 * worry about tokens or the BGL. However, we still have
714 * to call lwkt_getalltokens() in order to properly detect
715 * stale tokens. This call cannot fail for a UP build!
717 lwkt_getalltokens(ntd);
718 ++gd->gd_cnt.v_swtch;
719 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
720 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
724 * We have nothing to run but only let the idle loop halt
725 * the cpu if there are no pending interrupts.
727 ntd = &gd->gd_idlethread;
728 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
729 ntd->td_flags |= TDF_IDLE_NOHLT;
733 * The idle thread should not be holding the MP lock unless we
734 * are trapping in the kernel or in a panic. Since we select the
735 * idle thread unconditionally when no other thread is available,
736 * if the MP lock is desired during a panic or kernel trap, we
737 * have to loop in the scheduler until we get it.
739 if (ntd->td_mpcount) {
740 mpheld = MP_LOCK_HELD();
741 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
742 panic("Idle thread %p was holding the BGL!", ntd);
743 } else if (mpheld == 0) {
744 cpu_mplock_contested();
751 KASSERT(ntd->td_pri >= TDPRI_CRIT,
752 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
755 * Do the actual switch. If the new target does not need the MP lock
756 * and we are holding it, release the MP lock. If the new target requires
757 * the MP lock we have already acquired it for the target.
760 if (ntd->td_mpcount == 0 ) {
764 ASSERT_MP_LOCK_HELD(ntd);
769 KKASSERT(jg_tos_ok(ntd));
772 /* NOTE: current cpu may have changed after switch */
777 * Request that the target thread preempt the current thread. Preemption
778 * only works under a specific set of conditions:
780 * - We are not preempting ourselves
781 * - The target thread is owned by the current cpu
782 * - We are not currently being preempted
783 * - The target is not currently being preempted
784 * - We are not holding any spin locks
785 * - The target thread is not holding any tokens
786 * - We are able to satisfy the target's MP lock requirements (if any).
788 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
789 * this is called via lwkt_schedule() through the td_preemptable callback.
790 * critpri is the managed critical priority that we should ignore in order
791 * to determine whether preemption is possible (aka usually just the crit
792 * priority of lwkt_schedule() itself).
794 * XXX at the moment we run the target thread in a critical section during
795 * the preemption in order to prevent the target from taking interrupts
796 * that *WE* can't. Preemption is strictly limited to interrupt threads
797 * and interrupt-like threads, outside of a critical section, and the
798 * preempted source thread will be resumed the instant the target blocks
799 * whether or not the source is scheduled (i.e. preemption is supposed to
800 * be as transparent as possible).
802 * The target thread inherits our MP count (added to its own) for the
803 * duration of the preemption in order to preserve the atomicy of the
804 * MP lock during the preemption. Therefore, any preempting targets must be
805 * careful in regards to MP assertions. Note that the MP count may be
806 * out of sync with the physical mp_lock, but we do not have to preserve
807 * the original ownership of the lock if it was out of synch (that is, we
808 * can leave it synchronized on return).
811 lwkt_preempt(thread_t ntd, int critpri)
813 struct globaldata *gd = mycpu;
821 * The caller has put us in a critical section. We can only preempt
822 * if the caller of the caller was not in a critical section (basically
823 * a local interrupt), as determined by the 'critpri' parameter. We
824 * also can't preempt if the caller is holding any spinlocks (even if
825 * he isn't in a critical section). This also handles the tokens test.
827 * YYY The target thread must be in a critical section (else it must
828 * inherit our critical section? I dunno yet).
830 * Set need_lwkt_resched() unconditionally for now YYY.
832 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
834 td = gd->gd_curthread;
835 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
839 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
845 if (ntd->td_gd != gd) {
852 * Take the easy way out and do not preempt if we are holding
853 * any spinlocks. We could test whether the thread(s) being
854 * preempted interlock against the target thread's tokens and whether
855 * we can get all the target thread's tokens, but this situation
856 * should not occur very often so its easier to simply not preempt.
857 * Also, plain spinlocks are impossible to figure out at this point so
858 * just don't preempt.
860 * Do not try to preempt if the target thread is holding any tokens.
861 * We could try to acquire the tokens but this case is so rare there
862 * is no need to support it.
864 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
874 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
879 if (ntd->td_preempted) {
886 * note: an interrupt might have occured just as we were transitioning
887 * to or from the MP lock. In this case td_mpcount will be pre-disposed
888 * (non-zero) but not actually synchronized with the actual state of the
889 * lock. We can use it to imply an MP lock requirement for the
890 * preemption but we cannot use it to test whether we hold the MP lock
893 savecnt = td->td_mpcount;
894 mpheld = MP_LOCK_HELD();
895 ntd->td_mpcount += td->td_mpcount;
896 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
897 ntd->td_mpcount -= td->td_mpcount;
905 * Since we are able to preempt the current thread, there is no need to
906 * call need_lwkt_resched().
909 ntd->td_preempted = td;
910 td->td_flags |= TDF_PREEMPT_LOCK;
913 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
915 KKASSERT(savecnt == td->td_mpcount);
916 mpheld = MP_LOCK_HELD();
917 if (mpheld && td->td_mpcount == 0)
919 else if (mpheld == 0 && td->td_mpcount)
920 panic("lwkt_preempt(): MP lock was not held through");
922 ntd->td_preempted = NULL;
923 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
927 * Yield our thread while higher priority threads are pending. This is
928 * typically called when we leave a critical section but it can be safely
929 * called while we are in a critical section.
931 * This function will not generally yield to equal priority threads but it
932 * can occur as a side effect. Note that lwkt_switch() is called from
933 * inside the critical section to prevent its own crit_exit() from reentering
934 * lwkt_yield_quick().
936 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
937 * came along but was blocked and made pending.
939 * (self contained on a per cpu basis)
942 lwkt_yield_quick(void)
944 globaldata_t gd = mycpu;
945 thread_t td = gd->gd_curthread;
948 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
949 * it with a non-zero cpl then we might not wind up calling splz after
950 * a task switch when the critical section is exited even though the
951 * new task could accept the interrupt.
953 * XXX from crit_exit() only called after last crit section is released.
954 * If called directly will run splz() even if in a critical section.
956 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
957 * except for this special case, we MUST call splz() here to handle any
958 * pending ints, particularly after we switch, or we might accidently
959 * halt the cpu with interrupts pending.
961 if (gd->gd_reqflags && td->td_nest_count < 2)
965 * YYY enabling will cause wakeup() to task-switch, which really
966 * confused the old 4.x code. This is a good way to simulate
967 * preemption and MP without actually doing preemption or MP, because a
968 * lot of code assumes that wakeup() does not block.
970 if (untimely_switch && td->td_nest_count == 0 &&
971 gd->gd_intr_nesting_level == 0
973 crit_enter_quick(td);
975 * YYY temporary hacks until we disassociate the userland scheduler
976 * from the LWKT scheduler.
978 if (td->td_flags & TDF_RUNQ) {
979 lwkt_switch(); /* will not reenter yield function */
981 lwkt_schedule_self(td); /* make sure we are scheduled */
982 lwkt_switch(); /* will not reenter yield function */
983 lwkt_deschedule_self(td); /* make sure we are descheduled */
985 crit_exit_noyield(td);
990 * This implements a normal yield which, unlike _quick, will yield to equal
991 * priority threads as well. Note that gd_reqflags tests will be handled by
992 * the crit_exit() call in lwkt_switch().
994 * (self contained on a per cpu basis)
999 lwkt_schedule_self(curthread);
1004 * Return 0 if no runnable threads are pending at the same or higher
1005 * priority as the passed thread.
1007 * Return 1 if runnable threads are pending at the same priority.
1009 * Return 2 if runnable threads are pending at a higher priority.
1012 lwkt_check_resched(thread_t td)
1014 int pri = td->td_pri & TDPRI_MASK;
1016 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1018 if (TAILQ_NEXT(td, td_threadq))
1024 * Generic schedule. Possibly schedule threads belonging to other cpus and
1025 * deal with threads that might be blocked on a wait queue.
1027 * We have a little helper inline function which does additional work after
1028 * the thread has been enqueued, including dealing with preemption and
1029 * setting need_lwkt_resched() (which prevents the kernel from returning
1030 * to userland until it has processed higher priority threads).
1032 * It is possible for this routine to be called after a failed _enqueue
1033 * (due to the target thread migrating, sleeping, or otherwise blocked).
1034 * We have to check that the thread is actually on the run queue!
1036 * reschedok is an optimized constant propagated from lwkt_schedule() or
1037 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1038 * reschedule to be requested if the target thread has a higher priority.
1039 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1040 * be 0, prevented undesired reschedules.
1044 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1048 if (ntd->td_flags & TDF_RUNQ) {
1049 if (ntd->td_preemptable && reschedok) {
1050 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1051 } else if (reschedok) {
1053 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1054 need_lwkt_resched();
1061 _lwkt_schedule(thread_t td, int reschedok)
1063 globaldata_t mygd = mycpu;
1065 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1066 crit_enter_gd(mygd);
1067 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1068 if (td == mygd->gd_curthread) {
1072 * If we own the thread, there is no race (since we are in a
1073 * critical section). If we do not own the thread there might
1074 * be a race but the target cpu will deal with it.
1077 if (td->td_gd == mygd) {
1079 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1081 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1085 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1092 lwkt_schedule(thread_t td)
1094 _lwkt_schedule(td, 1);
1098 lwkt_schedule_noresched(thread_t td)
1100 _lwkt_schedule(td, 0);
1106 * Thread migration using a 'Pull' method. The thread may or may not be
1107 * the current thread. It MUST be descheduled and in a stable state.
1108 * lwkt_giveaway() must be called on the cpu owning the thread.
1110 * At any point after lwkt_giveaway() is called, the target cpu may
1111 * 'pull' the thread by calling lwkt_acquire().
1113 * MPSAFE - must be called under very specific conditions.
1116 lwkt_giveaway(thread_t td)
1118 globaldata_t gd = mycpu;
1121 KKASSERT(td->td_gd == gd);
1122 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1123 td->td_flags |= TDF_MIGRATING;
1128 lwkt_acquire(thread_t td)
1133 KKASSERT(td->td_flags & TDF_MIGRATING);
1138 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1139 crit_enter_gd(mygd);
1140 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1142 lwkt_process_ipiq();
1147 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1148 td->td_flags &= ~TDF_MIGRATING;
1151 crit_enter_gd(mygd);
1152 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1153 td->td_flags &= ~TDF_MIGRATING;
1161 * Generic deschedule. Descheduling threads other then your own should be
1162 * done only in carefully controlled circumstances. Descheduling is
1165 * This function may block if the cpu has run out of messages.
1168 lwkt_deschedule(thread_t td)
1172 if (td == curthread) {
1175 if (td->td_gd == mycpu) {
1178 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1188 * Set the target thread's priority. This routine does not automatically
1189 * switch to a higher priority thread, LWKT threads are not designed for
1190 * continuous priority changes. Yield if you want to switch.
1192 * We have to retain the critical section count which uses the high bits
1193 * of the td_pri field. The specified priority may also indicate zero or
1194 * more critical sections by adding TDPRI_CRIT*N.
1196 * Note that we requeue the thread whether it winds up on a different runq
1197 * or not. uio_yield() depends on this and the routine is not normally
1198 * called with the same priority otherwise.
1201 lwkt_setpri(thread_t td, int pri)
1204 KKASSERT(td->td_gd == mycpu);
1206 if (td->td_flags & TDF_RUNQ) {
1208 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1211 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1217 lwkt_setpri_self(int pri)
1219 thread_t td = curthread;
1221 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1223 if (td->td_flags & TDF_RUNQ) {
1225 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1228 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1234 * Migrate the current thread to the specified cpu.
1236 * This is accomplished by descheduling ourselves from the current cpu,
1237 * moving our thread to the tdallq of the target cpu, IPI messaging the
1238 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1239 * races while the thread is being migrated.
1242 static void lwkt_setcpu_remote(void *arg);
1246 lwkt_setcpu_self(globaldata_t rgd)
1249 thread_t td = curthread;
1251 if (td->td_gd != rgd) {
1252 crit_enter_quick(td);
1253 td->td_flags |= TDF_MIGRATING;
1254 lwkt_deschedule_self(td);
1255 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1256 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1258 /* we are now on the target cpu */
1259 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1260 crit_exit_quick(td);
1266 lwkt_migratecpu(int cpuid)
1271 rgd = globaldata_find(cpuid);
1272 lwkt_setcpu_self(rgd);
1277 * Remote IPI for cpu migration (called while in a critical section so we
1278 * do not have to enter another one). The thread has already been moved to
1279 * our cpu's allq, but we must wait for the thread to be completely switched
1280 * out on the originating cpu before we schedule it on ours or the stack
1281 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1282 * change to main memory.
1284 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1285 * against wakeups. It is best if this interface is used only when there
1286 * are no pending events that might try to schedule the thread.
1290 lwkt_setcpu_remote(void *arg)
1293 globaldata_t gd = mycpu;
1295 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1297 lwkt_process_ipiq();
1303 td->td_flags &= ~TDF_MIGRATING;
1304 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1310 lwkt_preempted_proc(void)
1312 thread_t td = curthread;
1313 while (td->td_preempted)
1314 td = td->td_preempted;
1319 * Create a kernel process/thread/whatever. It shares it's address space
1320 * with proc0 - ie: kernel only.
1322 * NOTE! By default new threads are created with the MP lock held. A
1323 * thread which does not require the MP lock should release it by calling
1324 * rel_mplock() at the start of the new thread.
1327 lwkt_create(void (*func)(void *), void *arg,
1328 struct thread **tdp, thread_t template, int tdflags, int cpu,
1329 const char *fmt, ...)
1334 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1338 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1341 * Set up arg0 for 'ps' etc
1343 __va_start(ap, fmt);
1344 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1348 * Schedule the thread to run
1350 if ((td->td_flags & TDF_STOPREQ) == 0)
1353 td->td_flags &= ~TDF_STOPREQ;
1358 * Destroy an LWKT thread. Warning! This function is not called when
1359 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1360 * uses a different reaping mechanism.
1365 thread_t td = curthread;
1369 if (td->td_flags & TDF_VERBOSE)
1370 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1374 * Get us into a critical section to interlock gd_freetd and loop
1375 * until we can get it freed.
1377 * We have to cache the current td in gd_freetd because objcache_put()ing
1378 * it would rip it out from under us while our thread is still active.
1381 crit_enter_quick(td);
1382 while ((std = gd->gd_freetd) != NULL) {
1383 gd->gd_freetd = NULL;
1384 objcache_put(thread_cache, std);
1386 lwkt_deschedule_self(td);
1387 lwkt_remove_tdallq(td);
1388 if (td->td_flags & TDF_ALLOCATED_THREAD)
1394 lwkt_remove_tdallq(thread_t td)
1396 KKASSERT(td->td_gd == mycpu);
1397 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1403 thread_t td = curthread;
1404 int lpri = td->td_pri;
1407 panic("td_pri is/would-go negative! %p %d", td, lpri);
1413 * Called from debugger/panic on cpus which have been stopped. We must still
1414 * process the IPIQ while stopped, even if we were stopped while in a critical
1417 * If we are dumping also try to process any pending interrupts. This may
1418 * or may not work depending on the state of the cpu at the point it was
1422 lwkt_smp_stopped(void)
1424 globaldata_t gd = mycpu;
1428 lwkt_process_ipiq();
1431 lwkt_process_ipiq();
1437 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1438 * get_mplock() has already incremented td_mpcount. We must block and
1439 * not return until giant is held.
1441 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1442 * reschedule the thread until it can obtain the giant lock for it.
1445 lwkt_mp_lock_contested(void)
1453 * The rel_mplock() code will call this function after releasing the
1454 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1456 * We then chain an IPI to a single other cpu potentially needing the
1457 * lock. This is a bit heuristical and we can wind up with IPIs flying
1458 * all over the place.
1460 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1463 lwkt_mp_lock_uncontested(void)
1473 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1474 mask = mp_lock_contention_mask;
1475 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1479 cpuid = bsfl(mask & tmpmask);
1482 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1483 dgd = globaldata_find(cpuid);
1484 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1490 * The idea is for this IPI to interrupt a potentially lower priority
1491 * thread, such as a user thread, to allow the scheduler to reschedule
1492 * a higher priority kernel thread that needs the MP lock.
1494 * For now we set the LWKT reschedule flag which generates an AST in
1495 * doreti, though theoretically it is also possible to possibly preempt
1496 * here if the underlying thread was operating in user mode. Nah.
1499 lwkt_mp_lock_uncontested_remote(void *arg __unused)
1501 need_lwkt_resched();