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.
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
48 #include <sys/rtprio.h>
49 #include <sys/queue.h>
50 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
55 #include <sys/spinlock.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
62 #include <vm/vm_param.h>
63 #include <vm/vm_kern.h>
64 #include <vm/vm_object.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_pager.h>
68 #include <vm/vm_extern.h>
70 #include <machine/stdarg.h>
71 #include <machine/smp.h>
74 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
77 static int mplock_countx = 0;
80 static int panic_on_cscount = 0;
82 static __int64_t switch_count = 0;
83 static __int64_t preempt_hit = 0;
84 static __int64_t preempt_miss = 0;
85 static __int64_t preempt_weird = 0;
86 static __int64_t token_contention_count = 0;
87 static __int64_t mplock_contention_count = 0;
88 static int lwkt_use_spin_port;
90 static int chain_mplock = 0;
92 static struct objcache *thread_cache;
94 volatile cpumask_t mp_lock_contention_mask;
96 extern void cpu_heavy_restore(void);
97 extern void cpu_lwkt_restore(void);
98 extern void cpu_kthread_restore(void);
99 extern void cpu_idle_restore(void);
104 jg_tos_ok(struct thread *td)
112 KKASSERT(td->td_sp != NULL);
113 tos = ((void **)td->td_sp)[0];
115 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
116 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
125 * We can make all thread ports use the spin backend instead of the thread
126 * backend. This should only be set to debug the spin backend.
128 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
131 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
134 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
136 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
137 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
138 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
139 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
142 &token_contention_count, 0, "spinning due to token contention");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
144 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
150 #if !defined(KTR_GIANT_CONTENTION)
151 #define KTR_GIANT_CONTENTION KTR_ALL
154 KTR_INFO_MASTER(giant);
155 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
156 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
158 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
161 * These helper procedures handle the runq, they can only be called from
162 * within a critical section.
164 * WARNING! Prior to SMP being brought up it is possible to enqueue and
165 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
166 * instead of 'mycpu' when referencing the globaldata structure. Once
167 * SMP live enqueuing and dequeueing only occurs on the current cpu.
171 _lwkt_dequeue(thread_t td)
173 if (td->td_flags & TDF_RUNQ) {
174 int nq = td->td_pri & TDPRI_MASK;
175 struct globaldata *gd = td->td_gd;
177 td->td_flags &= ~TDF_RUNQ;
178 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
179 /* runqmask is passively cleaned up by the switcher */
185 _lwkt_enqueue(thread_t td)
187 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
188 int nq = td->td_pri & TDPRI_MASK;
189 struct globaldata *gd = td->td_gd;
191 td->td_flags |= TDF_RUNQ;
192 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
193 gd->gd_runqmask |= 1 << nq;
198 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
200 struct thread *td = (struct thread *)obj;
202 td->td_kstack = NULL;
203 td->td_kstack_size = 0;
204 td->td_flags = TDF_ALLOCATED_THREAD;
209 _lwkt_thread_dtor(void *obj, void *privdata)
211 struct thread *td = (struct thread *)obj;
213 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
214 ("_lwkt_thread_dtor: not allocated from objcache"));
215 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
216 td->td_kstack_size > 0,
217 ("_lwkt_thread_dtor: corrupted stack"));
218 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
222 * Initialize the lwkt s/system.
227 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
228 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
229 NULL, CACHE_NTHREADS/2,
230 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
234 * Schedule a thread to run. As the current thread we can always safely
235 * schedule ourselves, and a shortcut procedure is provided for that
238 * (non-blocking, self contained on a per cpu basis)
241 lwkt_schedule_self(thread_t td)
243 crit_enter_quick(td);
244 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
245 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
251 * Deschedule a thread.
253 * (non-blocking, self contained on a per cpu basis)
256 lwkt_deschedule_self(thread_t td)
258 crit_enter_quick(td);
264 * LWKTs operate on a per-cpu basis
266 * WARNING! Called from early boot, 'mycpu' may not work yet.
269 lwkt_gdinit(struct globaldata *gd)
273 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
274 TAILQ_INIT(&gd->gd_tdrunq[i]);
276 TAILQ_INIT(&gd->gd_tdallq);
280 * Create a new thread. The thread must be associated with a process context
281 * or LWKT start address before it can be scheduled. If the target cpu is
282 * -1 the thread will be created on the current cpu.
284 * If you intend to create a thread without a process context this function
285 * does everything except load the startup and switcher function.
288 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
290 globaldata_t gd = mycpu;
294 * If static thread storage is not supplied allocate a thread. Reuse
295 * a cached free thread if possible. gd_freetd is used to keep an exiting
296 * thread intact through the exit.
299 if ((td = gd->gd_freetd) != NULL)
300 gd->gd_freetd = NULL;
302 td = objcache_get(thread_cache, M_WAITOK);
303 KASSERT((td->td_flags &
304 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
305 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
306 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
310 * Try to reuse cached stack.
312 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
313 if (flags & TDF_ALLOCATED_STACK) {
314 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
319 stack = (void *)kmem_alloc(&kernel_map, stksize);
320 flags |= TDF_ALLOCATED_STACK;
323 lwkt_init_thread(td, stack, stksize, flags, gd);
325 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
330 * Initialize a preexisting thread structure. This function is used by
331 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
333 * All threads start out in a critical section at a priority of
334 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
335 * appropriate. This function may send an IPI message when the
336 * requested cpu is not the current cpu and consequently gd_tdallq may
337 * not be initialized synchronously from the point of view of the originating
340 * NOTE! we have to be careful in regards to creating threads for other cpus
341 * if SMP has not yet been activated.
346 lwkt_init_thread_remote(void *arg)
351 * Protected by critical section held by IPI dispatch
353 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
359 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
360 struct globaldata *gd)
362 globaldata_t mygd = mycpu;
364 bzero(td, sizeof(struct thread));
365 td->td_kstack = stack;
366 td->td_kstack_size = stksize;
367 td->td_flags = flags;
369 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
371 if ((flags & TDF_MPSAFE) == 0)
374 if (lwkt_use_spin_port)
375 lwkt_initport_spin(&td->td_msgport);
377 lwkt_initport_thread(&td->td_msgport, td);
378 pmap_init_thread(td);
381 * Normally initializing a thread for a remote cpu requires sending an
382 * IPI. However, the idlethread is setup before the other cpus are
383 * activated so we have to treat it as a special case. XXX manipulation
384 * of gd_tdallq requires the BGL.
386 if (gd == mygd || td == &gd->gd_idlethread) {
388 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
391 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
395 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
401 lwkt_set_comm(thread_t td, const char *ctl, ...)
406 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
411 lwkt_hold(thread_t td)
417 lwkt_rele(thread_t td)
419 KKASSERT(td->td_refs > 0);
424 lwkt_wait_free(thread_t td)
427 tsleep(td, 0, "tdreap", hz);
431 lwkt_free_thread(thread_t td)
433 KASSERT((td->td_flags & TDF_RUNNING) == 0,
434 ("lwkt_free_thread: did not exit! %p", td));
436 if (td->td_flags & TDF_ALLOCATED_THREAD) {
437 objcache_put(thread_cache, td);
438 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
439 /* client-allocated struct with internally allocated stack */
440 KASSERT(td->td_kstack && td->td_kstack_size > 0,
441 ("lwkt_free_thread: corrupted stack"));
442 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
443 td->td_kstack = NULL;
444 td->td_kstack_size = 0;
450 * Switch to the next runnable lwkt. If no LWKTs are runnable then
451 * switch to the idlethread. Switching must occur within a critical
452 * section to avoid races with the scheduling queue.
454 * We always have full control over our cpu's run queue. Other cpus
455 * that wish to manipulate our queue must use the cpu_*msg() calls to
456 * talk to our cpu, so a critical section is all that is needed and
457 * the result is very, very fast thread switching.
459 * The LWKT scheduler uses a fixed priority model and round-robins at
460 * each priority level. User process scheduling is a totally
461 * different beast and LWKT priorities should not be confused with
462 * user process priorities.
464 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
465 * cleans it up. Note that the td_switch() function cannot do anything that
466 * requires the MP lock since the MP lock will have already been setup for
467 * the target thread (not the current thread). It's nice to have a scheduler
468 * that does not need the MP lock to work because it allows us to do some
469 * really cool high-performance MP lock optimizations.
471 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
472 * is not called by the current thread in the preemption case, only when
473 * the preempting thread blocks (in order to return to the original thread).
478 globaldata_t gd = mycpu;
479 thread_t td = gd->gd_curthread;
486 * Switching from within a 'fast' (non thread switched) interrupt or IPI
487 * is illegal. However, we may have to do it anyway if we hit a fatal
488 * kernel trap or we have paniced.
490 * If this case occurs save and restore the interrupt nesting level.
492 if (gd->gd_intr_nesting_level) {
496 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
497 panic("lwkt_switch: cannot switch from within "
498 "a fast interrupt, yet, td %p\n", td);
500 savegdnest = gd->gd_intr_nesting_level;
501 savegdtrap = gd->gd_trap_nesting_level;
502 gd->gd_intr_nesting_level = 0;
503 gd->gd_trap_nesting_level = 0;
504 if ((td->td_flags & TDF_PANICWARN) == 0) {
505 td->td_flags |= TDF_PANICWARN;
506 kprintf("Warning: thread switch from interrupt or IPI, "
507 "thread %p (%s)\n", td, td->td_comm);
511 gd->gd_intr_nesting_level = savegdnest;
512 gd->gd_trap_nesting_level = savegdtrap;
518 * Passive release (used to transition from user to kernel mode
519 * when we block or switch rather then when we enter the kernel).
520 * This function is NOT called if we are switching into a preemption
521 * or returning from a preemption. Typically this causes us to lose
522 * our current process designation (if we have one) and become a true
523 * LWKT thread, and may also hand the current process designation to
524 * another process and schedule thread.
531 lwkt_relalltokens(td);
534 * We had better not be holding any spin locks, but don't get into an
535 * endless panic loop.
537 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
538 ("lwkt_switch: still holding a shared spinlock %p!",
539 gd->gd_spinlock_rd));
540 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
541 ("lwkt_switch: still holding %d exclusive spinlocks!",
542 gd->gd_spinlocks_wr));
547 * td_mpcount cannot be used to determine if we currently hold the
548 * MP lock because get_mplock() will increment it prior to attempting
549 * to get the lock, and switch out if it can't. Our ownership of
550 * the actual lock will remain stable while we are in a critical section
551 * (but, of course, another cpu may own or release the lock so the
552 * actual value of mp_lock is not stable).
554 mpheld = MP_LOCK_HELD();
556 if (td->td_cscount) {
557 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
559 if (panic_on_cscount)
560 panic("switching while mastering cpusync");
564 if ((ntd = td->td_preempted) != NULL) {
566 * We had preempted another thread on this cpu, resume the preempted
567 * thread. This occurs transparently, whether the preempted thread
568 * was scheduled or not (it may have been preempted after descheduling
571 * We have to setup the MP lock for the original thread after backing
572 * out the adjustment that was made to curthread when the original
575 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
577 if (ntd->td_mpcount && mpheld == 0) {
578 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
579 td, ntd, td->td_mpcount, ntd->td_mpcount);
581 if (ntd->td_mpcount) {
582 td->td_mpcount -= ntd->td_mpcount;
583 KKASSERT(td->td_mpcount >= 0);
586 ntd->td_flags |= TDF_PREEMPT_DONE;
589 * The interrupt may have woken a thread up, we need to properly
590 * set the reschedule flag if the originally interrupted thread is
591 * at a lower priority.
593 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
595 /* YYY release mp lock on switchback if original doesn't need it */
598 * Priority queue / round-robin at each priority. Note that user
599 * processes run at a fixed, low priority and the user process
600 * scheduler deals with interactions between user processes
601 * by scheduling and descheduling them from the LWKT queue as
604 * We have to adjust the MP lock for the target thread. If we
605 * need the MP lock and cannot obtain it we try to locate a
606 * thread that does not need the MP lock. If we cannot, we spin
609 * A similar issue exists for the tokens held by the target thread.
610 * If we cannot obtain ownership of the tokens we cannot immediately
611 * schedule the thread.
615 * If an LWKT reschedule was requested, well that is what we are
616 * doing now so clear it.
618 clear_lwkt_resched();
620 if (gd->gd_runqmask) {
621 int nq = bsrl(gd->gd_runqmask);
622 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
623 gd->gd_runqmask &= ~(1 << nq);
628 * THREAD SELECTION FOR AN SMP MACHINE BUILD
630 * If the target needs the MP lock and we couldn't get it,
631 * or if the target is holding tokens and we could not
632 * gain ownership of the tokens, continue looking for a
633 * thread to schedule and spin instead of HLT if we can't.
635 * NOTE: the mpheld variable invalid after this conditional, it
636 * can change due to both cpu_try_mplock() returning success
637 * AND interactions in lwkt_getalltokens() due to the fact that
638 * we are trying to check the mpcount of a thread other then
639 * the current thread. Because of this, if the current thread
640 * is not holding td_mpcount, an IPI indirectly run via
641 * lwkt_getalltokens() can obtain and release the MP lock and
642 * cause the core MP lock to be released.
644 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
645 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
647 u_int32_t rqmask = gd->gd_runqmask;
649 mpheld = MP_LOCK_HELD();
652 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
653 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
654 /* spinning due to MP lock being held */
656 ++mplock_contention_count;
658 /* mplock still not held, 'mpheld' still valid */
663 * mpheld state invalid after getalltokens call returns
664 * failure, but the variable is only needed for
667 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
668 /* spinning due to token contention */
670 ++token_contention_count;
672 mpheld = MP_LOCK_HELD();
679 rqmask &= ~(1 << nq);
683 * We have two choices. We can either refuse to run a
684 * user thread when a kernel thread needs the MP lock
685 * but could not get it, or we can allow it to run but
686 * then expect an IPI (hopefully) later on to force a
687 * reschedule when the MP lock might become available.
689 if (nq < TDPRI_KERN_LPSCHED) {
690 if (chain_mplock == 0)
692 atomic_set_int(&mp_lock_contention_mask,
694 /* continue loop, allow user threads to be scheduled */
698 cpu_mplock_contested();
699 ntd = &gd->gd_idlethread;
700 ntd->td_flags |= TDF_IDLE_NOHLT;
701 goto using_idle_thread;
703 ++gd->gd_cnt.v_swtch;
704 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
705 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
710 ++gd->gd_cnt.v_swtch;
711 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
712 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
716 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
717 * worry about tokens or the BGL. However, we still have
718 * to call lwkt_getalltokens() in order to properly detect
719 * stale tokens. This call cannot fail for a UP build!
721 lwkt_getalltokens(ntd);
722 ++gd->gd_cnt.v_swtch;
723 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
724 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
728 * We have nothing to run but only let the idle loop halt
729 * the cpu if there are no pending interrupts.
731 ntd = &gd->gd_idlethread;
732 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
733 ntd->td_flags |= TDF_IDLE_NOHLT;
737 * The idle thread should not be holding the MP lock unless we
738 * are trapping in the kernel or in a panic. Since we select the
739 * idle thread unconditionally when no other thread is available,
740 * if the MP lock is desired during a panic or kernel trap, we
741 * have to loop in the scheduler until we get it.
743 if (ntd->td_mpcount) {
744 mpheld = MP_LOCK_HELD();
745 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
746 panic("Idle thread %p was holding the BGL!", ntd);
747 } else if (mpheld == 0) {
748 cpu_mplock_contested();
755 KASSERT(ntd->td_pri >= TDPRI_CRIT,
756 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
759 * Do the actual switch. If the new target does not need the MP lock
760 * and we are holding it, release the MP lock. If the new target requires
761 * the MP lock we have already acquired it for the target.
764 if (ntd->td_mpcount == 0 ) {
768 ASSERT_MP_LOCK_HELD(ntd);
774 KKASSERT(jg_tos_ok(ntd));
778 /* NOTE: current cpu may have changed after switch */
783 * Request that the target thread preempt the current thread. Preemption
784 * only works under a specific set of conditions:
786 * - We are not preempting ourselves
787 * - The target thread is owned by the current cpu
788 * - We are not currently being preempted
789 * - The target is not currently being preempted
790 * - We are not holding any spin locks
791 * - The target thread is not holding any tokens
792 * - We are able to satisfy the target's MP lock requirements (if any).
794 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
795 * this is called via lwkt_schedule() through the td_preemptable callback.
796 * critpri is the managed critical priority that we should ignore in order
797 * to determine whether preemption is possible (aka usually just the crit
798 * priority of lwkt_schedule() itself).
800 * XXX at the moment we run the target thread in a critical section during
801 * the preemption in order to prevent the target from taking interrupts
802 * that *WE* can't. Preemption is strictly limited to interrupt threads
803 * and interrupt-like threads, outside of a critical section, and the
804 * preempted source thread will be resumed the instant the target blocks
805 * whether or not the source is scheduled (i.e. preemption is supposed to
806 * be as transparent as possible).
808 * The target thread inherits our MP count (added to its own) for the
809 * duration of the preemption in order to preserve the atomicy of the
810 * MP lock during the preemption. Therefore, any preempting targets must be
811 * careful in regards to MP assertions. Note that the MP count may be
812 * out of sync with the physical mp_lock, but we do not have to preserve
813 * the original ownership of the lock if it was out of synch (that is, we
814 * can leave it synchronized on return).
817 lwkt_preempt(thread_t ntd, int critpri)
819 struct globaldata *gd = mycpu;
827 * The caller has put us in a critical section. We can only preempt
828 * if the caller of the caller was not in a critical section (basically
829 * a local interrupt), as determined by the 'critpri' parameter. We
830 * also can't preempt if the caller is holding any spinlocks (even if
831 * he isn't in a critical section). This also handles the tokens test.
833 * YYY The target thread must be in a critical section (else it must
834 * inherit our critical section? I dunno yet).
836 * Set need_lwkt_resched() unconditionally for now YYY.
838 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
840 td = gd->gd_curthread;
841 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
845 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
851 if (ntd->td_gd != gd) {
858 * Take the easy way out and do not preempt if we are holding
859 * any spinlocks. We could test whether the thread(s) being
860 * preempted interlock against the target thread's tokens and whether
861 * we can get all the target thread's tokens, but this situation
862 * should not occur very often so its easier to simply not preempt.
863 * Also, plain spinlocks are impossible to figure out at this point so
864 * just don't preempt.
866 * Do not try to preempt if the target thread is holding any tokens.
867 * We could try to acquire the tokens but this case is so rare there
868 * is no need to support it.
870 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
880 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
885 if (ntd->td_preempted) {
892 * note: an interrupt might have occured just as we were transitioning
893 * to or from the MP lock. In this case td_mpcount will be pre-disposed
894 * (non-zero) but not actually synchronized with the actual state of the
895 * lock. We can use it to imply an MP lock requirement for the
896 * preemption but we cannot use it to test whether we hold the MP lock
899 savecnt = td->td_mpcount;
900 mpheld = MP_LOCK_HELD();
901 ntd->td_mpcount += td->td_mpcount;
902 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
903 ntd->td_mpcount -= td->td_mpcount;
911 * Since we are able to preempt the current thread, there is no need to
912 * call need_lwkt_resched().
915 ntd->td_preempted = td;
916 td->td_flags |= TDF_PREEMPT_LOCK;
919 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
921 KKASSERT(savecnt == td->td_mpcount);
922 mpheld = MP_LOCK_HELD();
923 if (mpheld && td->td_mpcount == 0)
925 else if (mpheld == 0 && td->td_mpcount)
926 panic("lwkt_preempt(): MP lock was not held through");
928 ntd->td_preempted = NULL;
929 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
933 * Conditionally call splz() if gd_reqflags indicates work is pending.
935 * td_nest_count prevents deep nesting via splz() or doreti() which
936 * might otherwise blow out the kernel stack. Note that except for
937 * this special case, we MUST call splz() here to handle any
938 * pending ints, particularly after we switch, or we might accidently
939 * halt the cpu with interrupts pending.
941 * (self contained on a per cpu basis)
946 globaldata_t gd = mycpu;
947 thread_t td = gd->gd_curthread;
949 if (gd->gd_reqflags && td->td_nest_count < 2)
954 * This implements a normal yield which will yield to equal priority
955 * threads as well as higher priority threads. Note that gd_reqflags
956 * tests will be handled by the crit_exit() call in lwkt_switch().
958 * (self contained on a per cpu basis)
963 lwkt_schedule_self(curthread);
968 * This function is used along with the lwkt_passive_recover() inline
969 * by the trap code to negotiate a passive release of the current
970 * process/lwp designation with the user scheduler.
973 lwkt_passive_release(struct thread *td)
975 struct lwp *lp = td->td_lwp;
977 td->td_release = NULL;
978 lwkt_setpri_self(TDPRI_KERN_USER);
979 lp->lwp_proc->p_usched->release_curproc(lp);
983 * Make a kernel thread act as if it were in user mode with regards
984 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
985 * loops which may be potentially cpu-bound can call lwkt_user_yield().
987 * The lwkt_user_yield() function is designed to have very low overhead
988 * if no yield is determined to be needed.
991 lwkt_user_yield(void)
993 thread_t td = curthread;
994 struct lwp *lp = td->td_lwp;
998 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
999 * kernel can prevent other cpus from servicing interrupt threads
1000 * which still require the MP lock (which is a lot of them). This
1001 * has a chaining effect since if the interrupt is blocked, so is
1002 * the event, so normal scheduling will not pick up on the problem.
1004 if (mplock_countx && td->td_mpcount) {
1005 int savecnt = td->td_mpcount;
1011 td->td_mpcount = savecnt;
1017 * Another kernel thread wants the cpu
1019 if (lwkt_resched_wanted())
1023 * If the user scheduler has asynchronously determined that the current
1024 * process (when running in user mode) needs to lose the cpu then make
1025 * sure we are released.
1027 if (user_resched_wanted()) {
1033 * If we are released reduce our priority
1035 if (td->td_release == NULL) {
1036 if (lwkt_check_resched(td) > 0)
1038 lp->lwp_proc->p_usched->acquire_curproc(lp);
1039 td->td_release = lwkt_passive_release;
1040 lwkt_setpri_self(TDPRI_USER_NORM);
1045 * Return 0 if no runnable threads are pending at the same or higher
1046 * priority as the passed thread.
1048 * Return 1 if runnable threads are pending at the same priority.
1050 * Return 2 if runnable threads are pending at a higher priority.
1053 lwkt_check_resched(thread_t td)
1055 int pri = td->td_pri & TDPRI_MASK;
1057 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1059 if (TAILQ_NEXT(td, td_threadq))
1065 * Generic schedule. Possibly schedule threads belonging to other cpus and
1066 * deal with threads that might be blocked on a wait queue.
1068 * We have a little helper inline function which does additional work after
1069 * the thread has been enqueued, including dealing with preemption and
1070 * setting need_lwkt_resched() (which prevents the kernel from returning
1071 * to userland until it has processed higher priority threads).
1073 * It is possible for this routine to be called after a failed _enqueue
1074 * (due to the target thread migrating, sleeping, or otherwise blocked).
1075 * We have to check that the thread is actually on the run queue!
1077 * reschedok is an optimized constant propagated from lwkt_schedule() or
1078 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1079 * reschedule to be requested if the target thread has a higher priority.
1080 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1081 * be 0, prevented undesired reschedules.
1085 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1089 if (ntd->td_flags & TDF_RUNQ) {
1090 if (ntd->td_preemptable && reschedok) {
1091 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1092 } else if (reschedok) {
1094 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1095 need_lwkt_resched();
1102 _lwkt_schedule(thread_t td, int reschedok)
1104 globaldata_t mygd = mycpu;
1106 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1107 crit_enter_gd(mygd);
1108 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1109 if (td == mygd->gd_curthread) {
1113 * If we own the thread, there is no race (since we are in a
1114 * critical section). If we do not own the thread there might
1115 * be a race but the target cpu will deal with it.
1118 if (td->td_gd == mygd) {
1120 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1122 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1126 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1133 lwkt_schedule(thread_t td)
1135 _lwkt_schedule(td, 1);
1139 lwkt_schedule_noresched(thread_t td)
1141 _lwkt_schedule(td, 0);
1147 * Thread migration using a 'Pull' method. The thread may or may not be
1148 * the current thread. It MUST be descheduled and in a stable state.
1149 * lwkt_giveaway() must be called on the cpu owning the thread.
1151 * At any point after lwkt_giveaway() is called, the target cpu may
1152 * 'pull' the thread by calling lwkt_acquire().
1154 * We have to make sure the thread is not sitting on a per-cpu tsleep
1155 * queue or it will blow up when it moves to another cpu.
1157 * MPSAFE - must be called under very specific conditions.
1160 lwkt_giveaway(thread_t td)
1162 globaldata_t gd = mycpu;
1165 if (td->td_flags & TDF_TSLEEPQ)
1167 KKASSERT(td->td_gd == gd);
1168 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1169 td->td_flags |= TDF_MIGRATING;
1174 lwkt_acquire(thread_t td)
1179 KKASSERT(td->td_flags & TDF_MIGRATING);
1184 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1185 crit_enter_gd(mygd);
1186 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1188 lwkt_process_ipiq();
1193 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1194 td->td_flags &= ~TDF_MIGRATING;
1197 crit_enter_gd(mygd);
1198 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1199 td->td_flags &= ~TDF_MIGRATING;
1207 * Generic deschedule. Descheduling threads other then your own should be
1208 * done only in carefully controlled circumstances. Descheduling is
1211 * This function may block if the cpu has run out of messages.
1214 lwkt_deschedule(thread_t td)
1218 if (td == curthread) {
1221 if (td->td_gd == mycpu) {
1224 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1234 * Set the target thread's priority. This routine does not automatically
1235 * switch to a higher priority thread, LWKT threads are not designed for
1236 * continuous priority changes. Yield if you want to switch.
1238 * We have to retain the critical section count which uses the high bits
1239 * of the td_pri field. The specified priority may also indicate zero or
1240 * more critical sections by adding TDPRI_CRIT*N.
1242 * Note that we requeue the thread whether it winds up on a different runq
1243 * or not. uio_yield() depends on this and the routine is not normally
1244 * called with the same priority otherwise.
1247 lwkt_setpri(thread_t td, int pri)
1250 KKASSERT(td->td_gd == mycpu);
1252 if (td->td_flags & TDF_RUNQ) {
1254 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1257 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1263 lwkt_setpri_self(int pri)
1265 thread_t td = curthread;
1267 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1269 if (td->td_flags & TDF_RUNQ) {
1271 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1274 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1280 * Migrate the current thread to the specified cpu.
1282 * This is accomplished by descheduling ourselves from the current cpu,
1283 * moving our thread to the tdallq of the target cpu, IPI messaging the
1284 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1285 * races while the thread is being migrated.
1287 * We must be sure to remove ourselves from the current cpu's tsleepq
1288 * before potentially moving to another queue. The thread can be on
1289 * a tsleepq due to a left-over tsleep_interlock().
1292 static void lwkt_setcpu_remote(void *arg);
1296 lwkt_setcpu_self(globaldata_t rgd)
1299 thread_t td = curthread;
1301 if (td->td_gd != rgd) {
1302 crit_enter_quick(td);
1303 if (td->td_flags & TDF_TSLEEPQ)
1305 td->td_flags |= TDF_MIGRATING;
1306 lwkt_deschedule_self(td);
1307 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1308 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1310 /* we are now on the target cpu */
1311 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1312 crit_exit_quick(td);
1318 lwkt_migratecpu(int cpuid)
1323 rgd = globaldata_find(cpuid);
1324 lwkt_setcpu_self(rgd);
1329 * Remote IPI for cpu migration (called while in a critical section so we
1330 * do not have to enter another one). The thread has already been moved to
1331 * our cpu's allq, but we must wait for the thread to be completely switched
1332 * out on the originating cpu before we schedule it on ours or the stack
1333 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1334 * change to main memory.
1336 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1337 * against wakeups. It is best if this interface is used only when there
1338 * are no pending events that might try to schedule the thread.
1342 lwkt_setcpu_remote(void *arg)
1345 globaldata_t gd = mycpu;
1347 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1349 lwkt_process_ipiq();
1355 td->td_flags &= ~TDF_MIGRATING;
1356 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1362 lwkt_preempted_proc(void)
1364 thread_t td = curthread;
1365 while (td->td_preempted)
1366 td = td->td_preempted;
1371 * Create a kernel process/thread/whatever. It shares it's address space
1372 * with proc0 - ie: kernel only.
1374 * NOTE! By default new threads are created with the MP lock held. A
1375 * thread which does not require the MP lock should release it by calling
1376 * rel_mplock() at the start of the new thread.
1379 lwkt_create(void (*func)(void *), void *arg,
1380 struct thread **tdp, thread_t template, int tdflags, int cpu,
1381 const char *fmt, ...)
1386 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1390 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1393 * Set up arg0 for 'ps' etc
1395 __va_start(ap, fmt);
1396 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1400 * Schedule the thread to run
1402 if ((td->td_flags & TDF_STOPREQ) == 0)
1405 td->td_flags &= ~TDF_STOPREQ;
1410 * Destroy an LWKT thread. Warning! This function is not called when
1411 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1412 * uses a different reaping mechanism.
1417 thread_t td = curthread;
1421 if (td->td_flags & TDF_VERBOSE)
1422 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1426 * Get us into a critical section to interlock gd_freetd and loop
1427 * until we can get it freed.
1429 * We have to cache the current td in gd_freetd because objcache_put()ing
1430 * it would rip it out from under us while our thread is still active.
1433 crit_enter_quick(td);
1434 while ((std = gd->gd_freetd) != NULL) {
1435 gd->gd_freetd = NULL;
1436 objcache_put(thread_cache, std);
1440 * Remove thread resources from kernel lists and deschedule us for
1443 if (td->td_flags & TDF_TSLEEPQ)
1445 lwkt_deschedule_self(td);
1446 lwkt_remove_tdallq(td);
1447 if (td->td_flags & TDF_ALLOCATED_THREAD)
1453 lwkt_remove_tdallq(thread_t td)
1455 KKASSERT(td->td_gd == mycpu);
1456 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1462 thread_t td = curthread;
1463 int lpri = td->td_pri;
1466 panic("td_pri is/would-go negative! %p %d", td, lpri);
1472 * Called from debugger/panic on cpus which have been stopped. We must still
1473 * process the IPIQ while stopped, even if we were stopped while in a critical
1476 * If we are dumping also try to process any pending interrupts. This may
1477 * or may not work depending on the state of the cpu at the point it was
1481 lwkt_smp_stopped(void)
1483 globaldata_t gd = mycpu;
1487 lwkt_process_ipiq();
1490 lwkt_process_ipiq();
1496 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1497 * get_mplock() has already incremented td_mpcount. We must block and
1498 * not return until giant is held.
1500 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1501 * reschedule the thread until it can obtain the giant lock for it.
1504 lwkt_mp_lock_contested(void)
1513 * The rel_mplock() code will call this function after releasing the
1514 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1516 * We then chain an IPI to a single other cpu potentially needing the
1517 * lock. This is a bit heuristical and we can wind up with IPIs flying
1518 * all over the place.
1520 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1523 lwkt_mp_lock_uncontested(void)
1533 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1534 mask = mp_lock_contention_mask;
1535 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1539 cpuid = bsfl(mask & tmpmask);
1542 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1543 dgd = globaldata_find(cpuid);
1544 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1550 * The idea is for this IPI to interrupt a potentially lower priority
1551 * thread, such as a user thread, to allow the scheduler to reschedule
1552 * a higher priority kernel thread that needs the MP lock.
1554 * For now we set the LWKT reschedule flag which generates an AST in
1555 * doreti, though theoretically it is also possible to possibly preempt
1556 * here if the underlying thread was operating in user mode. Nah.
1559 lwkt_mp_lock_uncontested_remote(void *arg __unused)
1561 need_lwkt_resched();