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
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
60 #include <vm/vm_param.h>
61 #include <vm/vm_kern.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_pager.h>
66 #include <vm/vm_extern.h>
68 #include <machine/stdarg.h>
69 #include <machine/smp.h>
71 #if !defined(KTR_CTXSW)
72 #define KTR_CTXSW KTR_ALL
74 KTR_INFO_MASTER(ctxsw);
75 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "sw %p > %p", 2 * sizeof(struct thread *));
76 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "pre %p > %p", 2 * sizeof(struct thread *));
78 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
81 static int mplock_countx = 0;
84 static int panic_on_cscount = 0;
86 static __int64_t switch_count = 0;
87 static __int64_t preempt_hit = 0;
88 static __int64_t preempt_miss = 0;
89 static __int64_t preempt_weird = 0;
90 static __int64_t token_contention_count = 0;
91 static __int64_t mplock_contention_count = 0;
92 static int lwkt_use_spin_port;
94 static int chain_mplock = 0;
95 static int bgl_yield = 10;
97 static struct objcache *thread_cache;
99 volatile cpumask_t mp_lock_contention_mask;
101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
103 extern void cpu_heavy_restore(void);
104 extern void cpu_lwkt_restore(void);
105 extern void cpu_kthread_restore(void);
106 extern void cpu_idle_restore(void);
111 jg_tos_ok(struct thread *td)
119 KKASSERT(td->td_sp != NULL);
120 tos = ((void **)td->td_sp)[0];
122 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
123 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
132 * We can make all thread ports use the spin backend instead of the thread
133 * backend. This should only be set to debug the spin backend.
135 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
138 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141 SYSCTL_INT(_lwkt, OID_AUTO, chain_mplock, CTLFLAG_RW, &chain_mplock, 0, "");
142 SYSCTL_INT(_lwkt, OID_AUTO, bgl_yield_delay, CTLFLAG_RW, &bgl_yield, 0, "");
144 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
146 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
151 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
152 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
158 #if !defined(KTR_GIANT_CONTENTION)
159 #define KTR_GIANT_CONTENTION KTR_ALL
162 KTR_INFO_MASTER(giant);
163 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
164 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
166 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
169 * These helper procedures handle the runq, they can only be called from
170 * within a critical section.
172 * WARNING! Prior to SMP being brought up it is possible to enqueue and
173 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
174 * instead of 'mycpu' when referencing the globaldata structure. Once
175 * SMP live enqueuing and dequeueing only occurs on the current cpu.
179 _lwkt_dequeue(thread_t td)
181 if (td->td_flags & TDF_RUNQ) {
182 int nq = td->td_pri & TDPRI_MASK;
183 struct globaldata *gd = td->td_gd;
185 td->td_flags &= ~TDF_RUNQ;
186 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
187 /* runqmask is passively cleaned up by the switcher */
193 _lwkt_enqueue(thread_t td)
195 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
196 int nq = td->td_pri & TDPRI_MASK;
197 struct globaldata *gd = td->td_gd;
199 td->td_flags |= TDF_RUNQ;
200 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
201 gd->gd_runqmask |= 1 << nq;
206 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
208 struct thread *td = (struct thread *)obj;
210 td->td_kstack = NULL;
211 td->td_kstack_size = 0;
212 td->td_flags = TDF_ALLOCATED_THREAD;
217 _lwkt_thread_dtor(void *obj, void *privdata)
219 struct thread *td = (struct thread *)obj;
221 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
222 ("_lwkt_thread_dtor: not allocated from objcache"));
223 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
224 td->td_kstack_size > 0,
225 ("_lwkt_thread_dtor: corrupted stack"));
226 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
230 * Initialize the lwkt s/system.
235 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
236 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
237 NULL, CACHE_NTHREADS/2,
238 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
242 * Schedule a thread to run. As the current thread we can always safely
243 * schedule ourselves, and a shortcut procedure is provided for that
246 * (non-blocking, self contained on a per cpu basis)
249 lwkt_schedule_self(thread_t td)
251 crit_enter_quick(td);
252 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
253 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
259 * Deschedule a thread.
261 * (non-blocking, self contained on a per cpu basis)
264 lwkt_deschedule_self(thread_t td)
266 crit_enter_quick(td);
272 * LWKTs operate on a per-cpu basis
274 * WARNING! Called from early boot, 'mycpu' may not work yet.
277 lwkt_gdinit(struct globaldata *gd)
281 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
282 TAILQ_INIT(&gd->gd_tdrunq[i]);
284 TAILQ_INIT(&gd->gd_tdallq);
288 * Create a new thread. The thread must be associated with a process context
289 * or LWKT start address before it can be scheduled. If the target cpu is
290 * -1 the thread will be created on the current cpu.
292 * If you intend to create a thread without a process context this function
293 * does everything except load the startup and switcher function.
296 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
298 globaldata_t gd = mycpu;
302 * If static thread storage is not supplied allocate a thread. Reuse
303 * a cached free thread if possible. gd_freetd is used to keep an exiting
304 * thread intact through the exit.
307 if ((td = gd->gd_freetd) != NULL)
308 gd->gd_freetd = NULL;
310 td = objcache_get(thread_cache, M_WAITOK);
311 KASSERT((td->td_flags &
312 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
313 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
314 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
318 * Try to reuse cached stack.
320 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
321 if (flags & TDF_ALLOCATED_STACK) {
322 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
327 stack = (void *)kmem_alloc(&kernel_map, stksize);
328 flags |= TDF_ALLOCATED_STACK;
331 lwkt_init_thread(td, stack, stksize, flags, gd);
333 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
338 * Initialize a preexisting thread structure. This function is used by
339 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
341 * All threads start out in a critical section at a priority of
342 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
343 * appropriate. This function may send an IPI message when the
344 * requested cpu is not the current cpu and consequently gd_tdallq may
345 * not be initialized synchronously from the point of view of the originating
348 * NOTE! we have to be careful in regards to creating threads for other cpus
349 * if SMP has not yet been activated.
354 lwkt_init_thread_remote(void *arg)
359 * Protected by critical section held by IPI dispatch
361 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
367 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
368 struct globaldata *gd)
370 globaldata_t mygd = mycpu;
372 bzero(td, sizeof(struct thread));
373 td->td_kstack = stack;
374 td->td_kstack_size = stksize;
375 td->td_flags = flags;
377 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
379 if ((flags & TDF_MPSAFE) == 0)
382 if (lwkt_use_spin_port)
383 lwkt_initport_spin(&td->td_msgport);
385 lwkt_initport_thread(&td->td_msgport, td);
386 pmap_init_thread(td);
389 * Normally initializing a thread for a remote cpu requires sending an
390 * IPI. However, the idlethread is setup before the other cpus are
391 * activated so we have to treat it as a special case. XXX manipulation
392 * of gd_tdallq requires the BGL.
394 if (gd == mygd || td == &gd->gd_idlethread) {
396 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
399 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
403 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
409 lwkt_set_comm(thread_t td, const char *ctl, ...)
414 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
419 lwkt_hold(thread_t td)
425 lwkt_rele(thread_t td)
427 KKASSERT(td->td_refs > 0);
432 lwkt_wait_free(thread_t td)
435 tsleep(td, 0, "tdreap", hz);
439 lwkt_free_thread(thread_t td)
441 KASSERT((td->td_flags & TDF_RUNNING) == 0,
442 ("lwkt_free_thread: did not exit! %p", td));
444 if (td->td_flags & TDF_ALLOCATED_THREAD) {
445 objcache_put(thread_cache, td);
446 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
447 /* client-allocated struct with internally allocated stack */
448 KASSERT(td->td_kstack && td->td_kstack_size > 0,
449 ("lwkt_free_thread: corrupted stack"));
450 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
451 td->td_kstack = NULL;
452 td->td_kstack_size = 0;
458 * Switch to the next runnable lwkt. If no LWKTs are runnable then
459 * switch to the idlethread. Switching must occur within a critical
460 * section to avoid races with the scheduling queue.
462 * We always have full control over our cpu's run queue. Other cpus
463 * that wish to manipulate our queue must use the cpu_*msg() calls to
464 * talk to our cpu, so a critical section is all that is needed and
465 * the result is very, very fast thread switching.
467 * The LWKT scheduler uses a fixed priority model and round-robins at
468 * each priority level. User process scheduling is a totally
469 * different beast and LWKT priorities should not be confused with
470 * user process priorities.
472 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
473 * cleans it up. Note that the td_switch() function cannot do anything that
474 * requires the MP lock since the MP lock will have already been setup for
475 * the target thread (not the current thread). It's nice to have a scheduler
476 * that does not need the MP lock to work because it allows us to do some
477 * really cool high-performance MP lock optimizations.
479 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
480 * is not called by the current thread in the preemption case, only when
481 * the preempting thread blocks (in order to return to the original thread).
486 globaldata_t gd = mycpu;
487 thread_t td = gd->gd_curthread;
494 * Switching from within a 'fast' (non thread switched) interrupt or IPI
495 * is illegal. However, we may have to do it anyway if we hit a fatal
496 * kernel trap or we have paniced.
498 * If this case occurs save and restore the interrupt nesting level.
500 if (gd->gd_intr_nesting_level) {
504 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
505 panic("lwkt_switch: cannot switch from within "
506 "a fast interrupt, yet, td %p\n", td);
508 savegdnest = gd->gd_intr_nesting_level;
509 savegdtrap = gd->gd_trap_nesting_level;
510 gd->gd_intr_nesting_level = 0;
511 gd->gd_trap_nesting_level = 0;
512 if ((td->td_flags & TDF_PANICWARN) == 0) {
513 td->td_flags |= TDF_PANICWARN;
514 kprintf("Warning: thread switch from interrupt or IPI, "
515 "thread %p (%s)\n", td, td->td_comm);
519 gd->gd_intr_nesting_level = savegdnest;
520 gd->gd_trap_nesting_level = savegdtrap;
526 * Passive release (used to transition from user to kernel mode
527 * when we block or switch rather then when we enter the kernel).
528 * This function is NOT called if we are switching into a preemption
529 * or returning from a preemption. Typically this causes us to lose
530 * our current process designation (if we have one) and become a true
531 * LWKT thread, and may also hand the current process designation to
532 * another process and schedule thread.
539 lwkt_relalltokens(td);
542 * We had better not be holding any spin locks, but don't get into an
543 * endless panic loop.
545 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
546 ("lwkt_switch: still holding a shared spinlock %p!",
547 gd->gd_spinlock_rd));
548 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
549 ("lwkt_switch: still holding %d exclusive spinlocks!",
550 gd->gd_spinlocks_wr));
555 * td_mpcount cannot be used to determine if we currently hold the
556 * MP lock because get_mplock() will increment it prior to attempting
557 * to get the lock, and switch out if it can't. Our ownership of
558 * the actual lock will remain stable while we are in a critical section
559 * (but, of course, another cpu may own or release the lock so the
560 * actual value of mp_lock is not stable).
562 mpheld = MP_LOCK_HELD();
564 if (td->td_cscount) {
565 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
567 if (panic_on_cscount)
568 panic("switching while mastering cpusync");
572 if ((ntd = td->td_preempted) != NULL) {
574 * We had preempted another thread on this cpu, resume the preempted
575 * thread. This occurs transparently, whether the preempted thread
576 * was scheduled or not (it may have been preempted after descheduling
579 * We have to setup the MP lock for the original thread after backing
580 * out the adjustment that was made to curthread when the original
583 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
585 if (ntd->td_mpcount && mpheld == 0) {
586 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
587 td, ntd, td->td_mpcount, ntd->td_mpcount);
589 if (ntd->td_mpcount) {
590 td->td_mpcount -= ntd->td_mpcount;
591 KKASSERT(td->td_mpcount >= 0);
594 ntd->td_flags |= TDF_PREEMPT_DONE;
597 * The interrupt may have woken a thread up, we need to properly
598 * set the reschedule flag if the originally interrupted thread is
599 * at a lower priority.
601 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
603 /* YYY release mp lock on switchback if original doesn't need it */
606 * Priority queue / round-robin at each priority. Note that user
607 * processes run at a fixed, low priority and the user process
608 * scheduler deals with interactions between user processes
609 * by scheduling and descheduling them from the LWKT queue as
612 * We have to adjust the MP lock for the target thread. If we
613 * need the MP lock and cannot obtain it we try to locate a
614 * thread that does not need the MP lock. If we cannot, we spin
617 * A similar issue exists for the tokens held by the target thread.
618 * If we cannot obtain ownership of the tokens we cannot immediately
619 * schedule the thread.
623 * If an LWKT reschedule was requested, well that is what we are
624 * doing now so clear it.
626 clear_lwkt_resched();
628 if (gd->gd_runqmask) {
629 int nq = bsrl(gd->gd_runqmask);
630 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
631 gd->gd_runqmask &= ~(1 << nq);
636 * THREAD SELECTION FOR AN SMP MACHINE BUILD
638 * If the target needs the MP lock and we couldn't get it,
639 * or if the target is holding tokens and we could not
640 * gain ownership of the tokens, continue looking for a
641 * thread to schedule and spin instead of HLT if we can't.
643 * NOTE: the mpheld variable invalid after this conditional, it
644 * can change due to both cpu_try_mplock() returning success
645 * AND interactions in lwkt_getalltokens() due to the fact that
646 * we are trying to check the mpcount of a thread other then
647 * the current thread. Because of this, if the current thread
648 * is not holding td_mpcount, an IPI indirectly run via
649 * lwkt_getalltokens() can obtain and release the MP lock and
650 * cause the core MP lock to be released.
652 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
653 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
655 u_int32_t rqmask = gd->gd_runqmask;
657 mpheld = MP_LOCK_HELD();
660 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
661 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
662 /* spinning due to MP lock being held */
664 ++mplock_contention_count;
666 /* mplock still not held, 'mpheld' still valid */
671 * mpheld state invalid after getalltokens call returns
672 * failure, but the variable is only needed for
675 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
676 /* spinning due to token contention */
678 ++token_contention_count;
680 mpheld = MP_LOCK_HELD();
687 rqmask &= ~(1 << nq);
691 * We have two choices. We can either refuse to run a
692 * user thread when a kernel thread needs the MP lock
693 * but could not get it, or we can allow it to run but
694 * then expect an IPI (hopefully) later on to force a
695 * reschedule when the MP lock might become available.
697 if (nq < TDPRI_KERN_LPSCHED) {
698 if (chain_mplock == 0)
700 atomic_set_int(&mp_lock_contention_mask,
702 /* continue loop, allow user threads to be scheduled */
706 cpu_mplock_contested();
707 ntd = &gd->gd_idlethread;
708 ntd->td_flags |= TDF_IDLE_NOHLT;
709 goto using_idle_thread;
711 ++gd->gd_cnt.v_swtch;
712 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
713 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
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 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
725 * worry about tokens or the BGL. However, we still have
726 * to call lwkt_getalltokens() in order to properly detect
727 * stale tokens. This call cannot fail for a UP build!
729 lwkt_getalltokens(ntd);
730 ++gd->gd_cnt.v_swtch;
731 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
732 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
736 * We have nothing to run but only let the idle loop halt
737 * the cpu if there are no pending interrupts.
739 ntd = &gd->gd_idlethread;
740 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
741 ntd->td_flags |= TDF_IDLE_NOHLT;
745 * The idle thread should not be holding the MP lock unless we
746 * are trapping in the kernel or in a panic. Since we select the
747 * idle thread unconditionally when no other thread is available,
748 * if the MP lock is desired during a panic or kernel trap, we
749 * have to loop in the scheduler until we get it.
751 if (ntd->td_mpcount) {
752 mpheld = MP_LOCK_HELD();
753 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
754 panic("Idle thread %p was holding the BGL!", ntd);
755 } else if (mpheld == 0) {
756 cpu_mplock_contested();
763 KASSERT(ntd->td_pri >= TDPRI_CRIT,
764 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
767 * Do the actual switch. If the new target does not need the MP lock
768 * and we are holding it, release the MP lock. If the new target requires
769 * the MP lock we have already acquired it for the target.
772 if (ntd->td_mpcount == 0 ) {
776 ASSERT_MP_LOCK_HELD(ntd);
782 KKASSERT(jg_tos_ok(ntd));
784 KTR_LOG(ctxsw_sw, td, ntd);
787 /* NOTE: current cpu may have changed after switch */
792 * Request that the target thread preempt the current thread. Preemption
793 * only works under a specific set of conditions:
795 * - We are not preempting ourselves
796 * - The target thread is owned by the current cpu
797 * - We are not currently being preempted
798 * - The target is not currently being preempted
799 * - We are not holding any spin locks
800 * - The target thread is not holding any tokens
801 * - We are able to satisfy the target's MP lock requirements (if any).
803 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
804 * this is called via lwkt_schedule() through the td_preemptable callback.
805 * critpri is the managed critical priority that we should ignore in order
806 * to determine whether preemption is possible (aka usually just the crit
807 * priority of lwkt_schedule() itself).
809 * XXX at the moment we run the target thread in a critical section during
810 * the preemption in order to prevent the target from taking interrupts
811 * that *WE* can't. Preemption is strictly limited to interrupt threads
812 * and interrupt-like threads, outside of a critical section, and the
813 * preempted source thread will be resumed the instant the target blocks
814 * whether or not the source is scheduled (i.e. preemption is supposed to
815 * be as transparent as possible).
817 * The target thread inherits our MP count (added to its own) for the
818 * duration of the preemption in order to preserve the atomicy of the
819 * MP lock during the preemption. Therefore, any preempting targets must be
820 * careful in regards to MP assertions. Note that the MP count may be
821 * out of sync with the physical mp_lock, but we do not have to preserve
822 * the original ownership of the lock if it was out of synch (that is, we
823 * can leave it synchronized on return).
826 lwkt_preempt(thread_t ntd, int critpri)
828 struct globaldata *gd = mycpu;
836 * The caller has put us in a critical section. We can only preempt
837 * if the caller of the caller was not in a critical section (basically
838 * a local interrupt), as determined by the 'critpri' parameter. We
839 * also can't preempt if the caller is holding any spinlocks (even if
840 * he isn't in a critical section). This also handles the tokens test.
842 * YYY The target thread must be in a critical section (else it must
843 * inherit our critical section? I dunno yet).
845 * Set need_lwkt_resched() unconditionally for now YYY.
847 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
849 td = gd->gd_curthread;
850 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
854 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
860 if (ntd->td_gd != gd) {
867 * Take the easy way out and do not preempt if we are holding
868 * any spinlocks. We could test whether the thread(s) being
869 * preempted interlock against the target thread's tokens and whether
870 * we can get all the target thread's tokens, but this situation
871 * should not occur very often so its easier to simply not preempt.
872 * Also, plain spinlocks are impossible to figure out at this point so
873 * just don't preempt.
875 * Do not try to preempt if the target thread is holding any tokens.
876 * We could try to acquire the tokens but this case is so rare there
877 * is no need to support it.
879 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
889 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
894 if (ntd->td_preempted) {
901 * note: an interrupt might have occured just as we were transitioning
902 * to or from the MP lock. In this case td_mpcount will be pre-disposed
903 * (non-zero) but not actually synchronized with the actual state of the
904 * lock. We can use it to imply an MP lock requirement for the
905 * preemption but we cannot use it to test whether we hold the MP lock
908 savecnt = td->td_mpcount;
909 mpheld = MP_LOCK_HELD();
910 ntd->td_mpcount += td->td_mpcount;
911 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
912 ntd->td_mpcount -= td->td_mpcount;
920 * Since we are able to preempt the current thread, there is no need to
921 * call need_lwkt_resched().
924 ntd->td_preempted = td;
925 td->td_flags |= TDF_PREEMPT_LOCK;
926 KTR_LOG(ctxsw_pre, td, ntd);
929 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
931 KKASSERT(savecnt == td->td_mpcount);
932 mpheld = MP_LOCK_HELD();
933 if (mpheld && td->td_mpcount == 0)
935 else if (mpheld == 0 && td->td_mpcount)
936 panic("lwkt_preempt(): MP lock was not held through");
938 ntd->td_preempted = NULL;
939 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
943 * Conditionally call splz() if gd_reqflags indicates work is pending.
945 * td_nest_count prevents deep nesting via splz() or doreti() which
946 * might otherwise blow out the kernel stack. Note that except for
947 * this special case, we MUST call splz() here to handle any
948 * pending ints, particularly after we switch, or we might accidently
949 * halt the cpu with interrupts pending.
951 * (self contained on a per cpu basis)
956 globaldata_t gd = mycpu;
957 thread_t td = gd->gd_curthread;
959 if (gd->gd_reqflags && td->td_nest_count < 2)
964 * This implements a normal yield which will yield to equal priority
965 * threads as well as higher priority threads. Note that gd_reqflags
966 * tests will be handled by the crit_exit() call in lwkt_switch().
968 * (self contained on a per cpu basis)
973 lwkt_schedule_self(curthread);
978 * This function is used along with the lwkt_passive_recover() inline
979 * by the trap code to negotiate a passive release of the current
980 * process/lwp designation with the user scheduler.
983 lwkt_passive_release(struct thread *td)
985 struct lwp *lp = td->td_lwp;
987 td->td_release = NULL;
988 lwkt_setpri_self(TDPRI_KERN_USER);
989 lp->lwp_proc->p_usched->release_curproc(lp);
993 * Make a kernel thread act as if it were in user mode with regards
994 * to scheduling, to avoid becoming cpu-bound in the kernel. Kernel
995 * loops which may be potentially cpu-bound can call lwkt_user_yield().
997 * The lwkt_user_yield() function is designed to have very low overhead
998 * if no yield is determined to be needed.
1001 lwkt_user_yield(void)
1003 thread_t td = curthread;
1004 struct lwp *lp = td->td_lwp;
1008 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1009 * kernel can prevent other cpus from servicing interrupt threads
1010 * which still require the MP lock (which is a lot of them). This
1011 * has a chaining effect since if the interrupt is blocked, so is
1012 * the event, so normal scheduling will not pick up on the problem.
1014 if (mplock_countx && td->td_mpcount) {
1015 int savecnt = td->td_mpcount;
1022 td->td_mpcount = savecnt;
1027 * Another kernel thread wants the cpu
1029 if (lwkt_resched_wanted())
1033 * If the user scheduler has asynchronously determined that the current
1034 * process (when running in user mode) needs to lose the cpu then make
1035 * sure we are released.
1037 if (user_resched_wanted()) {
1043 * If we are released reduce our priority
1045 if (td->td_release == NULL) {
1046 if (lwkt_check_resched(td) > 0)
1049 lp->lwp_proc->p_usched->acquire_curproc(lp);
1050 td->td_release = lwkt_passive_release;
1051 lwkt_setpri_self(TDPRI_USER_NORM);
1057 * Return 0 if no runnable threads are pending at the same or higher
1058 * priority as the passed thread.
1060 * Return 1 if runnable threads are pending at the same priority.
1062 * Return 2 if runnable threads are pending at a higher priority.
1065 lwkt_check_resched(thread_t td)
1067 int pri = td->td_pri & TDPRI_MASK;
1069 if (td->td_gd->gd_runqmask > (2 << pri) - 1)
1071 if (TAILQ_NEXT(td, td_threadq))
1077 * Generic schedule. Possibly schedule threads belonging to other cpus and
1078 * deal with threads that might be blocked on a wait queue.
1080 * We have a little helper inline function which does additional work after
1081 * the thread has been enqueued, including dealing with preemption and
1082 * setting need_lwkt_resched() (which prevents the kernel from returning
1083 * to userland until it has processed higher priority threads).
1085 * It is possible for this routine to be called after a failed _enqueue
1086 * (due to the target thread migrating, sleeping, or otherwise blocked).
1087 * We have to check that the thread is actually on the run queue!
1089 * reschedok is an optimized constant propagated from lwkt_schedule() or
1090 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1091 * reschedule to be requested if the target thread has a higher priority.
1092 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1093 * be 0, prevented undesired reschedules.
1097 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri, int reschedok)
1101 if (ntd->td_flags & TDF_RUNQ) {
1102 if (ntd->td_preemptable && reschedok) {
1103 ntd->td_preemptable(ntd, cpri); /* YYY +token */
1104 } else if (reschedok) {
1106 if ((ntd->td_pri & TDPRI_MASK) > (otd->td_pri & TDPRI_MASK))
1107 need_lwkt_resched();
1114 _lwkt_schedule(thread_t td, int reschedok)
1116 globaldata_t mygd = mycpu;
1118 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1119 crit_enter_gd(mygd);
1120 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1121 if (td == mygd->gd_curthread) {
1125 * If we own the thread, there is no race (since we are in a
1126 * critical section). If we do not own the thread there might
1127 * be a race but the target cpu will deal with it.
1130 if (td->td_gd == mygd) {
1132 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1134 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1138 _lwkt_schedule_post(mygd, td, TDPRI_CRIT, reschedok);
1145 lwkt_schedule(thread_t td)
1147 _lwkt_schedule(td, 1);
1151 lwkt_schedule_noresched(thread_t td)
1153 _lwkt_schedule(td, 0);
1157 * When scheduled remotely if frame != NULL the IPIQ is being
1158 * run via doreti or an interrupt then preemption can be allowed.
1160 * To allow preemption we have to drop the critical section so only
1161 * one is present in _lwkt_schedule_post.
1164 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1166 thread_t td = curthread;
1169 if (frame && ntd->td_preemptable) {
1170 crit_exit_noyield(td);
1171 _lwkt_schedule(ntd, 1);
1172 crit_enter_quick(td);
1174 _lwkt_schedule(ntd, 1);
1181 * Thread migration using a 'Pull' method. The thread may or may not be
1182 * the current thread. It MUST be descheduled and in a stable state.
1183 * lwkt_giveaway() must be called on the cpu owning the thread.
1185 * At any point after lwkt_giveaway() is called, the target cpu may
1186 * 'pull' the thread by calling lwkt_acquire().
1188 * We have to make sure the thread is not sitting on a per-cpu tsleep
1189 * queue or it will blow up when it moves to another cpu.
1191 * MPSAFE - must be called under very specific conditions.
1194 lwkt_giveaway(thread_t td)
1196 globaldata_t gd = mycpu;
1199 if (td->td_flags & TDF_TSLEEPQ)
1201 KKASSERT(td->td_gd == gd);
1202 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1203 td->td_flags |= TDF_MIGRATING;
1208 lwkt_acquire(thread_t td)
1213 KKASSERT(td->td_flags & TDF_MIGRATING);
1218 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1219 crit_enter_gd(mygd);
1220 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1222 lwkt_process_ipiq();
1227 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1228 td->td_flags &= ~TDF_MIGRATING;
1231 crit_enter_gd(mygd);
1232 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1233 td->td_flags &= ~TDF_MIGRATING;
1241 * Generic deschedule. Descheduling threads other then your own should be
1242 * done only in carefully controlled circumstances. Descheduling is
1245 * This function may block if the cpu has run out of messages.
1248 lwkt_deschedule(thread_t td)
1252 if (td == curthread) {
1255 if (td->td_gd == mycpu) {
1258 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1268 * Set the target thread's priority. This routine does not automatically
1269 * switch to a higher priority thread, LWKT threads are not designed for
1270 * continuous priority changes. Yield if you want to switch.
1272 * We have to retain the critical section count which uses the high bits
1273 * of the td_pri field. The specified priority may also indicate zero or
1274 * more critical sections by adding TDPRI_CRIT*N.
1276 * Note that we requeue the thread whether it winds up on a different runq
1277 * or not. uio_yield() depends on this and the routine is not normally
1278 * called with the same priority otherwise.
1281 lwkt_setpri(thread_t td, int pri)
1284 KKASSERT(td->td_gd == mycpu);
1286 if (td->td_flags & TDF_RUNQ) {
1288 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1291 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1297 lwkt_setpri_self(int pri)
1299 thread_t td = curthread;
1301 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1303 if (td->td_flags & TDF_RUNQ) {
1305 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1308 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1314 * Migrate the current thread to the specified cpu.
1316 * This is accomplished by descheduling ourselves from the current cpu,
1317 * moving our thread to the tdallq of the target cpu, IPI messaging the
1318 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1319 * races while the thread is being migrated.
1321 * We must be sure to remove ourselves from the current cpu's tsleepq
1322 * before potentially moving to another queue. The thread can be on
1323 * a tsleepq due to a left-over tsleep_interlock().
1326 static void lwkt_setcpu_remote(void *arg);
1330 lwkt_setcpu_self(globaldata_t rgd)
1333 thread_t td = curthread;
1335 if (td->td_gd != rgd) {
1336 crit_enter_quick(td);
1337 if (td->td_flags & TDF_TSLEEPQ)
1339 td->td_flags |= TDF_MIGRATING;
1340 lwkt_deschedule_self(td);
1341 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1342 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1344 /* we are now on the target cpu */
1345 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1346 crit_exit_quick(td);
1352 lwkt_migratecpu(int cpuid)
1357 rgd = globaldata_find(cpuid);
1358 lwkt_setcpu_self(rgd);
1363 * Remote IPI for cpu migration (called while in a critical section so we
1364 * do not have to enter another one). The thread has already been moved to
1365 * our cpu's allq, but we must wait for the thread to be completely switched
1366 * out on the originating cpu before we schedule it on ours or the stack
1367 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1368 * change to main memory.
1370 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1371 * against wakeups. It is best if this interface is used only when there
1372 * are no pending events that might try to schedule the thread.
1376 lwkt_setcpu_remote(void *arg)
1379 globaldata_t gd = mycpu;
1381 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1383 lwkt_process_ipiq();
1389 td->td_flags &= ~TDF_MIGRATING;
1390 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1396 lwkt_preempted_proc(void)
1398 thread_t td = curthread;
1399 while (td->td_preempted)
1400 td = td->td_preempted;
1405 * Create a kernel process/thread/whatever. It shares it's address space
1406 * with proc0 - ie: kernel only.
1408 * NOTE! By default new threads are created with the MP lock held. A
1409 * thread which does not require the MP lock should release it by calling
1410 * rel_mplock() at the start of the new thread.
1413 lwkt_create(void (*func)(void *), void *arg,
1414 struct thread **tdp, thread_t template, int tdflags, int cpu,
1415 const char *fmt, ...)
1420 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1424 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1427 * Set up arg0 for 'ps' etc
1429 __va_start(ap, fmt);
1430 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1434 * Schedule the thread to run
1436 if ((td->td_flags & TDF_STOPREQ) == 0)
1439 td->td_flags &= ~TDF_STOPREQ;
1444 * Destroy an LWKT thread. Warning! This function is not called when
1445 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1446 * uses a different reaping mechanism.
1451 thread_t td = curthread;
1455 if (td->td_flags & TDF_VERBOSE)
1456 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1460 * Get us into a critical section to interlock gd_freetd and loop
1461 * until we can get it freed.
1463 * We have to cache the current td in gd_freetd because objcache_put()ing
1464 * it would rip it out from under us while our thread is still active.
1467 crit_enter_quick(td);
1468 while ((std = gd->gd_freetd) != NULL) {
1469 gd->gd_freetd = NULL;
1470 objcache_put(thread_cache, std);
1474 * Remove thread resources from kernel lists and deschedule us for
1477 if (td->td_flags & TDF_TSLEEPQ)
1480 lwkt_deschedule_self(td);
1481 lwkt_remove_tdallq(td);
1482 if (td->td_flags & TDF_ALLOCATED_THREAD)
1488 lwkt_remove_tdallq(thread_t td)
1490 KKASSERT(td->td_gd == mycpu);
1491 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1497 thread_t td = curthread;
1498 int lpri = td->td_pri;
1501 panic("td_pri is/would-go negative! %p %d", td, lpri);
1507 * Called from debugger/panic on cpus which have been stopped. We must still
1508 * process the IPIQ while stopped, even if we were stopped while in a critical
1511 * If we are dumping also try to process any pending interrupts. This may
1512 * or may not work depending on the state of the cpu at the point it was
1516 lwkt_smp_stopped(void)
1518 globaldata_t gd = mycpu;
1522 lwkt_process_ipiq();
1525 lwkt_process_ipiq();
1531 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1532 * get_mplock() has already incremented td_mpcount. We must block and
1533 * not return until giant is held.
1535 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1536 * reschedule the thread until it can obtain the giant lock for it.
1539 lwkt_mp_lock_contested(void)
1548 * The rel_mplock() code will call this function after releasing the
1549 * last reference on the MP lock if mp_lock_contention_mask is non-zero.
1551 * We then chain an IPI to a single other cpu potentially needing the
1552 * lock. This is a bit heuristical and we can wind up with IPIs flying
1553 * all over the place.
1555 static void lwkt_mp_lock_uncontested_remote(void *arg __unused);
1558 lwkt_mp_lock_uncontested(void)
1568 atomic_clear_int(&mp_lock_contention_mask, gd->gd_cpumask);
1569 mask = mp_lock_contention_mask;
1570 tmpmask = ~((1 << gd->gd_cpuid) - 1);
1574 cpuid = bsfl(mask & tmpmask);
1577 atomic_clear_int(&mp_lock_contention_mask, 1 << cpuid);
1578 dgd = globaldata_find(cpuid);
1579 lwkt_send_ipiq(dgd, lwkt_mp_lock_uncontested_remote, NULL);
1585 * The idea is for this IPI to interrupt a potentially lower priority
1586 * thread, such as a user thread, to allow the scheduler to reschedule
1587 * a higher priority kernel thread that needs the MP lock.
1589 * For now we set the LWKT reschedule flag which generates an AST in
1590 * doreti, though theoretically it is also possible to possibly preempt
1591 * here if the underlying thread was operating in user mode. Nah.
1594 lwkt_mp_lock_uncontested_remote(void *arg __unused)
1596 need_lwkt_resched();