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 1.109 2007/07/02 01:37:08 dillon 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.
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
50 #include <sys/rtprio.h>
51 #include <sys/queue.h>
52 #include <sys/sysctl.h>
53 #include <sys/kthread.h>
54 #include <machine/cpu.h>
57 #include <sys/spinlock.h>
60 #include <sys/thread2.h>
61 #include <sys/spinlock2.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_zone.h>
73 #include <machine/stdarg.h>
74 #include <machine/smp.h>
78 #include <sys/stdint.h>
79 #include <libcaps/thread.h>
80 #include <sys/thread.h>
81 #include <sys/msgport.h>
82 #include <sys/errno.h>
83 #include <libcaps/globaldata.h>
84 #include <machine/cpufunc.h>
85 #include <sys/thread2.h>
86 #include <sys/msgport2.h>
90 #include <machine/lock.h>
91 #include <machine/atomic.h>
92 #include <machine/cpu.h>
96 static int untimely_switch = 0;
98 static int panic_on_cscount = 0;
100 static __int64_t switch_count = 0;
101 static __int64_t preempt_hit = 0;
102 static __int64_t preempt_miss = 0;
103 static __int64_t preempt_weird = 0;
104 static __int64_t token_contention_count = 0;
105 static __int64_t mplock_contention_count = 0;
106 static int lwkt_use_spin_port;
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
116 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
118 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
121 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
123 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
125 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
126 &token_contention_count, 0, "spinning due to token contention");
127 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
128 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
137 #if !defined(KTR_GIANT_CONTENTION)
138 #define KTR_GIANT_CONTENTION KTR_ALL
141 KTR_INFO_MASTER(giant);
142 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
143 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
145 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
150 * These helper procedures handle the runq, they can only be called from
151 * within a critical section.
153 * WARNING! Prior to SMP being brought up it is possible to enqueue and
154 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
155 * instead of 'mycpu' when referencing the globaldata structure. Once
156 * SMP live enqueuing and dequeueing only occurs on the current cpu.
160 _lwkt_dequeue(thread_t td)
162 if (td->td_flags & TDF_RUNQ) {
163 int nq = td->td_pri & TDPRI_MASK;
164 struct globaldata *gd = td->td_gd;
166 td->td_flags &= ~TDF_RUNQ;
167 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
168 /* runqmask is passively cleaned up by the switcher */
174 _lwkt_enqueue(thread_t td)
176 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
177 int nq = td->td_pri & TDPRI_MASK;
178 struct globaldata *gd = td->td_gd;
180 td->td_flags |= TDF_RUNQ;
181 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
182 gd->gd_runqmask |= 1 << nq;
187 * Schedule a thread to run. As the current thread we can always safely
188 * schedule ourselves, and a shortcut procedure is provided for that
191 * (non-blocking, self contained on a per cpu basis)
194 lwkt_schedule_self(thread_t td)
196 crit_enter_quick(td);
197 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
198 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
204 * Deschedule a thread.
206 * (non-blocking, self contained on a per cpu basis)
209 lwkt_deschedule_self(thread_t td)
211 crit_enter_quick(td);
219 * LWKTs operate on a per-cpu basis
221 * WARNING! Called from early boot, 'mycpu' may not work yet.
224 lwkt_gdinit(struct globaldata *gd)
228 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
229 TAILQ_INIT(&gd->gd_tdrunq[i]);
231 TAILQ_INIT(&gd->gd_tdallq);
237 * Create a new thread. The thread must be associated with a process context
238 * or LWKT start address before it can be scheduled. If the target cpu is
239 * -1 the thread will be created on the current cpu.
241 * If you intend to create a thread without a process context this function
242 * does everything except load the startup and switcher function.
245 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
248 globaldata_t gd = mycpu;
252 if (gd->gd_tdfreecount > 0) {
253 --gd->gd_tdfreecount;
254 td = TAILQ_FIRST(&gd->gd_tdfreeq);
255 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
256 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
257 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
259 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
263 td = zalloc(thread_zone);
265 td = malloc(sizeof(struct thread));
267 td->td_kstack = NULL;
268 td->td_kstack_size = 0;
269 flags |= TDF_ALLOCATED_THREAD;
272 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
273 if (flags & TDF_ALLOCATED_STACK) {
275 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
277 libcaps_free_stack(stack, td->td_kstack_size);
284 stack = (void *)kmem_alloc(&kernel_map, stksize);
286 stack = libcaps_alloc_stack(stksize);
288 flags |= TDF_ALLOCATED_STACK;
291 lwkt_init_thread(td, stack, stksize, flags, mycpu);
293 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
300 * Initialize a preexisting thread structure. This function is used by
301 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
303 * All threads start out in a critical section at a priority of
304 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
305 * appropriate. This function may send an IPI message when the
306 * requested cpu is not the current cpu and consequently gd_tdallq may
307 * not be initialized synchronously from the point of view of the originating
310 * NOTE! we have to be careful in regards to creating threads for other cpus
311 * if SMP has not yet been activated.
316 lwkt_init_thread_remote(void *arg)
321 * Protected by critical section held by IPI dispatch
323 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
329 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
330 struct globaldata *gd)
332 globaldata_t mygd = mycpu;
334 bzero(td, sizeof(struct thread));
335 td->td_kstack = stack;
336 td->td_kstack_size = stksize;
337 td->td_flags = flags;
339 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
341 if ((flags & TDF_MPSAFE) == 0)
344 if (lwkt_use_spin_port)
345 lwkt_initport_spin(&td->td_msgport);
347 lwkt_initport_thread(&td->td_msgport, td);
348 pmap_init_thread(td);
351 * Normally initializing a thread for a remote cpu requires sending an
352 * IPI. However, the idlethread is setup before the other cpus are
353 * activated so we have to treat it as a special case. XXX manipulation
354 * of gd_tdallq requires the BGL.
356 if (gd == mygd || td == &gd->gd_idlethread) {
358 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
361 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
365 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
373 lwkt_set_comm(thread_t td, const char *ctl, ...)
378 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
383 lwkt_hold(thread_t td)
389 lwkt_rele(thread_t td)
391 KKASSERT(td->td_refs > 0);
398 lwkt_wait_free(thread_t td)
401 tsleep(td, 0, "tdreap", hz);
407 lwkt_free_thread(thread_t td)
409 struct globaldata *gd = mycpu;
411 KASSERT((td->td_flags & TDF_RUNNING) == 0,
412 ("lwkt_free_thread: did not exit! %p", td));
415 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
416 (td->td_flags & TDF_ALLOCATED_THREAD)
418 ++gd->gd_tdfreecount;
419 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
423 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
425 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
427 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
430 td->td_kstack = NULL;
431 td->td_kstack_size = 0;
433 if (td->td_flags & TDF_ALLOCATED_THREAD) {
435 zfree(thread_zone, td);
445 * Switch to the next runnable lwkt. If no LWKTs are runnable then
446 * switch to the idlethread. Switching must occur within a critical
447 * section to avoid races with the scheduling queue.
449 * We always have full control over our cpu's run queue. Other cpus
450 * that wish to manipulate our queue must use the cpu_*msg() calls to
451 * talk to our cpu, so a critical section is all that is needed and
452 * the result is very, very fast thread switching.
454 * The LWKT scheduler uses a fixed priority model and round-robins at
455 * each priority level. User process scheduling is a totally
456 * different beast and LWKT priorities should not be confused with
457 * user process priorities.
459 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
460 * cleans it up. Note that the td_switch() function cannot do anything that
461 * requires the MP lock since the MP lock will have already been setup for
462 * the target thread (not the current thread). It's nice to have a scheduler
463 * that does not need the MP lock to work because it allows us to do some
464 * really cool high-performance MP lock optimizations.
466 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
467 * is not called by the current thread in the preemption case, only when
468 * the preempting thread blocks (in order to return to the original thread).
473 globaldata_t gd = mycpu;
474 thread_t td = gd->gd_curthread;
481 * Switching from within a 'fast' (non thread switched) interrupt or IPI
482 * is illegal. However, we may have to do it anyway if we hit a fatal
483 * kernel trap or we have paniced.
485 * If this case occurs save and restore the interrupt nesting level.
487 if (gd->gd_intr_nesting_level) {
491 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
492 panic("lwkt_switch: cannot switch from within "
493 "a fast interrupt, yet, td %p\n", td);
495 savegdnest = gd->gd_intr_nesting_level;
496 savegdtrap = gd->gd_trap_nesting_level;
497 gd->gd_intr_nesting_level = 0;
498 gd->gd_trap_nesting_level = 0;
499 if ((td->td_flags & TDF_PANICWARN) == 0) {
500 td->td_flags |= TDF_PANICWARN;
501 kprintf("Warning: thread switch from interrupt or IPI, "
502 "thread %p (%s)\n", td, td->td_comm);
504 db_print_backtrace();
508 gd->gd_intr_nesting_level = savegdnest;
509 gd->gd_trap_nesting_level = savegdtrap;
515 * Passive release (used to transition from user to kernel mode
516 * when we block or switch rather then when we enter the kernel).
517 * This function is NOT called if we are switching into a preemption
518 * or returning from a preemption. Typically this causes us to lose
519 * our current process designation (if we have one) and become a true
520 * LWKT thread, and may also hand the current process designation to
521 * another process and schedule thread.
529 lwkt_relalltokens(td);
533 * We had better not be holding any spin locks, but don't get into an
534 * endless panic loop.
536 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
537 ("lwkt_switch: still holding a shared spinlock %p!",
538 gd->gd_spinlock_rd));
539 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
540 ("lwkt_switch: still holding %d exclusive spinlocks!",
541 gd->gd_spinlocks_wr));
546 * td_mpcount cannot be used to determine if we currently hold the
547 * MP lock because get_mplock() will increment it prior to attempting
548 * to get the lock, and switch out if it can't. Our ownership of
549 * the actual lock will remain stable while we are in a critical section
550 * (but, of course, another cpu may own or release the lock so the
551 * actual value of mp_lock is not stable).
553 mpheld = MP_LOCK_HELD();
555 if (td->td_cscount) {
556 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
558 if (panic_on_cscount)
559 panic("switching while mastering cpusync");
563 if ((ntd = td->td_preempted) != NULL) {
565 * We had preempted another thread on this cpu, resume the preempted
566 * thread. This occurs transparently, whether the preempted thread
567 * was scheduled or not (it may have been preempted after descheduling
570 * We have to setup the MP lock for the original thread after backing
571 * out the adjustment that was made to curthread when the original
574 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
576 if (ntd->td_mpcount && mpheld == 0) {
577 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
578 td, ntd, td->td_mpcount, ntd->td_mpcount);
580 if (ntd->td_mpcount) {
581 td->td_mpcount -= ntd->td_mpcount;
582 KKASSERT(td->td_mpcount >= 0);
585 ntd->td_flags |= TDF_PREEMPT_DONE;
588 * XXX. The interrupt may have woken a thread up, we need to properly
589 * set the reschedule flag if the originally interrupted thread is at
592 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
594 /* YYY release mp lock on switchback if original doesn't need it */
597 * Priority queue / round-robin at each priority. Note that user
598 * processes run at a fixed, low priority and the user process
599 * scheduler deals with interactions between user processes
600 * by scheduling and descheduling them from the LWKT queue as
603 * We have to adjust the MP lock for the target thread. If we
604 * need the MP lock and cannot obtain it we try to locate a
605 * thread that does not need the MP lock. If we cannot, we spin
608 * A similar issue exists for the tokens held by the target thread.
609 * If we cannot obtain ownership of the tokens we cannot immediately
610 * schedule the thread.
614 * If an LWKT reschedule was requested, well that is what we are
615 * doing now so clear it.
617 clear_lwkt_resched();
619 if (gd->gd_runqmask) {
620 int nq = bsrl(gd->gd_runqmask);
621 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
622 gd->gd_runqmask &= ~(1 << nq);
627 * THREAD SELECTION FOR AN SMP MACHINE BUILD
629 * If the target needs the MP lock and we couldn't get it,
630 * or if the target is holding tokens and we could not
631 * gain ownership of the tokens, continue looking for a
632 * thread to schedule and spin instead of HLT if we can't.
634 * NOTE: the mpheld variable invalid after this conditional, it
635 * can change due to both cpu_try_mplock() returning success
636 * AND interactions in lwkt_getalltokens() due to the fact that
637 * we are trying to check the mpcount of a thread other then
638 * the current thread. Because of this, if the current thread
639 * is not holding td_mpcount, an IPI indirectly run via
640 * lwkt_getalltokens() can obtain and release the MP lock and
641 * cause the core MP lock to be released.
643 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
644 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
646 u_int32_t rqmask = gd->gd_runqmask;
648 mpheld = MP_LOCK_HELD();
651 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
652 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
653 /* spinning due to MP lock being held */
655 ++mplock_contention_count;
657 /* mplock still not held, 'mpheld' still valid */
662 * mpheld state invalid after getalltokens call returns
663 * failure, but the variable is only needed for
666 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
667 /* spinning due to token contention */
669 ++token_contention_count;
671 mpheld = MP_LOCK_HELD();
678 rqmask &= ~(1 << nq);
682 cpu_mplock_contested();
683 ntd = &gd->gd_idlethread;
684 ntd->td_flags |= TDF_IDLE_NOHLT;
685 goto using_idle_thread;
687 ++gd->gd_cnt.v_swtch;
688 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
689 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
692 ++gd->gd_cnt.v_swtch;
693 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
694 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
698 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
699 * worry about tokens or the BGL.
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);
707 * We have nothing to run but only let the idle loop halt
708 * the cpu if there are no pending interrupts.
710 ntd = &gd->gd_idlethread;
711 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
712 ntd->td_flags |= TDF_IDLE_NOHLT;
716 * The idle thread should not be holding the MP lock unless we
717 * are trapping in the kernel or in a panic. Since we select the
718 * idle thread unconditionally when no other thread is available,
719 * if the MP lock is desired during a panic or kernel trap, we
720 * have to loop in the scheduler until we get it.
722 if (ntd->td_mpcount) {
723 mpheld = MP_LOCK_HELD();
724 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
725 panic("Idle thread %p was holding the BGL!", ntd);
726 } else if (mpheld == 0) {
727 cpu_mplock_contested();
734 KASSERT(ntd->td_pri >= TDPRI_CRIT,
735 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
738 * Do the actual switch. If the new target does not need the MP lock
739 * and we are holding it, release the MP lock. If the new target requires
740 * the MP lock we have already acquired it for the target.
743 if (ntd->td_mpcount == 0 ) {
747 ASSERT_MP_LOCK_HELD(ntd);
754 /* NOTE: current cpu may have changed after switch */
759 * Request that the target thread preempt the current thread. Preemption
760 * only works under a specific set of conditions:
762 * - We are not preempting ourselves
763 * - The target thread is owned by the current cpu
764 * - We are not currently being preempted
765 * - The target is not currently being preempted
766 * - We are able to satisfy the target's MP lock requirements (if any).
768 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
769 * this is called via lwkt_schedule() through the td_preemptable callback.
770 * critpri is the managed critical priority that we should ignore in order
771 * to determine whether preemption is possible (aka usually just the crit
772 * priority of lwkt_schedule() itself).
774 * XXX at the moment we run the target thread in a critical section during
775 * the preemption in order to prevent the target from taking interrupts
776 * that *WE* can't. Preemption is strictly limited to interrupt threads
777 * and interrupt-like threads, outside of a critical section, and the
778 * preempted source thread will be resumed the instant the target blocks
779 * whether or not the source is scheduled (i.e. preemption is supposed to
780 * be as transparent as possible).
782 * The target thread inherits our MP count (added to its own) for the
783 * duration of the preemption in order to preserve the atomicy of the
784 * MP lock during the preemption. Therefore, any preempting targets must be
785 * careful in regards to MP assertions. Note that the MP count may be
786 * out of sync with the physical mp_lock, but we do not have to preserve
787 * the original ownership of the lock if it was out of synch (that is, we
788 * can leave it synchronized on return).
791 lwkt_preempt(thread_t ntd, int critpri)
793 struct globaldata *gd = mycpu;
801 * The caller has put us in a critical section. We can only preempt
802 * if the caller of the caller was not in a critical section (basically
803 * a local interrupt), as determined by the 'critpri' parameter. We
804 * also acn't preempt if the caller is holding any spinlocks (even if
805 * he isn't in a critical section). This also handles the tokens test.
807 * YYY The target thread must be in a critical section (else it must
808 * inherit our critical section? I dunno yet).
810 * Set need_lwkt_resched() unconditionally for now YYY.
812 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
814 td = gd->gd_curthread;
815 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
819 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
825 if (ntd->td_gd != gd) {
832 * Take the easy way out and do not preempt if the target is holding
833 * any spinlocks. We could test whether the thread(s) being
834 * preempted interlock against the target thread's tokens and whether
835 * we can get all the target thread's tokens, but this situation
836 * should not occur very often so its easier to simply not preempt.
837 * Also, plain spinlocks are impossible to figure out at this point so
838 * just don't preempt.
840 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
845 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
850 if (ntd->td_preempted) {
857 * note: an interrupt might have occured just as we were transitioning
858 * to or from the MP lock. In this case td_mpcount will be pre-disposed
859 * (non-zero) but not actually synchronized with the actual state of the
860 * lock. We can use it to imply an MP lock requirement for the
861 * preemption but we cannot use it to test whether we hold the MP lock
864 savecnt = td->td_mpcount;
865 mpheld = MP_LOCK_HELD();
866 ntd->td_mpcount += td->td_mpcount;
867 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
868 ntd->td_mpcount -= td->td_mpcount;
876 * Since we are able to preempt the current thread, there is no need to
877 * call need_lwkt_resched().
880 ntd->td_preempted = td;
881 td->td_flags |= TDF_PREEMPT_LOCK;
883 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
885 KKASSERT(savecnt == td->td_mpcount);
886 mpheld = MP_LOCK_HELD();
887 if (mpheld && td->td_mpcount == 0)
889 else if (mpheld == 0 && td->td_mpcount)
890 panic("lwkt_preempt(): MP lock was not held through");
892 ntd->td_preempted = NULL;
893 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
897 * Yield our thread while higher priority threads are pending. This is
898 * typically called when we leave a critical section but it can be safely
899 * called while we are in a critical section.
901 * This function will not generally yield to equal priority threads but it
902 * can occur as a side effect. Note that lwkt_switch() is called from
903 * inside the critical section to prevent its own crit_exit() from reentering
904 * lwkt_yield_quick().
906 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
907 * came along but was blocked and made pending.
909 * (self contained on a per cpu basis)
912 lwkt_yield_quick(void)
914 globaldata_t gd = mycpu;
915 thread_t td = gd->gd_curthread;
918 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
919 * it with a non-zero cpl then we might not wind up calling splz after
920 * a task switch when the critical section is exited even though the
921 * new task could accept the interrupt.
923 * XXX from crit_exit() only called after last crit section is released.
924 * If called directly will run splz() even if in a critical section.
926 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
927 * except for this special case, we MUST call splz() here to handle any
928 * pending ints, particularly after we switch, or we might accidently
929 * halt the cpu with interrupts pending.
931 if (gd->gd_reqflags && td->td_nest_count < 2)
935 * YYY enabling will cause wakeup() to task-switch, which really
936 * confused the old 4.x code. This is a good way to simulate
937 * preemption and MP without actually doing preemption or MP, because a
938 * lot of code assumes that wakeup() does not block.
940 if (untimely_switch && td->td_nest_count == 0 &&
941 gd->gd_intr_nesting_level == 0
943 crit_enter_quick(td);
945 * YYY temporary hacks until we disassociate the userland scheduler
946 * from the LWKT scheduler.
948 if (td->td_flags & TDF_RUNQ) {
949 lwkt_switch(); /* will not reenter yield function */
951 lwkt_schedule_self(td); /* make sure we are scheduled */
952 lwkt_switch(); /* will not reenter yield function */
953 lwkt_deschedule_self(td); /* make sure we are descheduled */
955 crit_exit_noyield(td);
960 * This implements a normal yield which, unlike _quick, will yield to equal
961 * priority threads as well. Note that gd_reqflags tests will be handled by
962 * the crit_exit() call in lwkt_switch().
964 * (self contained on a per cpu basis)
969 lwkt_schedule_self(curthread);
974 * Generic schedule. Possibly schedule threads belonging to other cpus and
975 * deal with threads that might be blocked on a wait queue.
977 * We have a little helper inline function which does additional work after
978 * the thread has been enqueued, including dealing with preemption and
979 * setting need_lwkt_resched() (which prevents the kernel from returning
980 * to userland until it has processed higher priority threads).
982 * It is possible for this routine to be called after a failed _enqueue
983 * (due to the target thread migrating, sleeping, or otherwise blocked).
984 * We have to check that the thread is actually on the run queue!
988 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
990 if (ntd->td_flags & TDF_RUNQ) {
991 if (ntd->td_preemptable) {
992 ntd->td_preemptable(ntd, cpri); /* YYY +token */
993 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
994 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
1002 lwkt_schedule(thread_t td)
1004 globaldata_t mygd = mycpu;
1006 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1007 crit_enter_gd(mygd);
1008 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1009 if (td == mygd->gd_curthread) {
1013 * If we own the thread, there is no race (since we are in a
1014 * critical section). If we do not own the thread there might
1015 * be a race but the target cpu will deal with it.
1018 if (td->td_gd == mygd) {
1020 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1022 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1026 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1035 * Thread migration using a 'Pull' method. The thread may or may not be
1036 * the current thread. It MUST be descheduled and in a stable state.
1037 * lwkt_giveaway() must be called on the cpu owning the thread.
1039 * At any point after lwkt_giveaway() is called, the target cpu may
1040 * 'pull' the thread by calling lwkt_acquire().
1042 * MPSAFE - must be called under very specific conditions.
1045 lwkt_giveaway(thread_t td)
1047 globaldata_t gd = mycpu;
1050 KKASSERT(td->td_gd == gd);
1051 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1052 td->td_flags |= TDF_MIGRATING;
1057 lwkt_acquire(thread_t td)
1062 KKASSERT(td->td_flags & TDF_MIGRATING);
1067 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1068 crit_enter_gd(mygd);
1069 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1072 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1073 td->td_flags &= ~TDF_MIGRATING;
1076 crit_enter_gd(mygd);
1077 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1078 td->td_flags &= ~TDF_MIGRATING;
1086 * Generic deschedule. Descheduling threads other then your own should be
1087 * done only in carefully controlled circumstances. Descheduling is
1090 * This function may block if the cpu has run out of messages.
1093 lwkt_deschedule(thread_t td)
1097 if (td == curthread) {
1100 if (td->td_gd == mycpu) {
1103 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1113 * Set the target thread's priority. This routine does not automatically
1114 * switch to a higher priority thread, LWKT threads are not designed for
1115 * continuous priority changes. Yield if you want to switch.
1117 * We have to retain the critical section count which uses the high bits
1118 * of the td_pri field. The specified priority may also indicate zero or
1119 * more critical sections by adding TDPRI_CRIT*N.
1121 * Note that we requeue the thread whether it winds up on a different runq
1122 * or not. uio_yield() depends on this and the routine is not normally
1123 * called with the same priority otherwise.
1126 lwkt_setpri(thread_t td, int pri)
1129 KKASSERT(td->td_gd == mycpu);
1131 if (td->td_flags & TDF_RUNQ) {
1133 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1136 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1142 lwkt_setpri_self(int pri)
1144 thread_t td = curthread;
1146 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1148 if (td->td_flags & TDF_RUNQ) {
1150 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1153 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1159 * Determine if there is a runnable thread at a higher priority then
1160 * the current thread. lwkt_setpri() does not check this automatically.
1161 * Return 1 if there is, 0 if there isn't.
1163 * Example: if bit 31 of runqmask is set and the current thread is priority
1164 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1166 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1167 * up comparing against 0xffffffff, a comparison that will always be false.
1170 lwkt_checkpri_self(void)
1172 globaldata_t gd = mycpu;
1173 thread_t td = gd->gd_curthread;
1174 int nq = td->td_pri & TDPRI_MASK;
1176 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1177 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1185 * Migrate the current thread to the specified cpu.
1187 * This is accomplished by descheduling ourselves from the current cpu,
1188 * moving our thread to the tdallq of the target cpu, IPI messaging the
1189 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1190 * races while the thread is being migrated.
1193 static void lwkt_setcpu_remote(void *arg);
1197 lwkt_setcpu_self(globaldata_t rgd)
1200 thread_t td = curthread;
1202 if (td->td_gd != rgd) {
1203 crit_enter_quick(td);
1204 td->td_flags |= TDF_MIGRATING;
1205 lwkt_deschedule_self(td);
1206 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1207 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1209 /* we are now on the target cpu */
1210 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1211 crit_exit_quick(td);
1217 lwkt_migratecpu(int cpuid)
1222 rgd = globaldata_find(cpuid);
1223 lwkt_setcpu_self(rgd);
1228 * Remote IPI for cpu migration (called while in a critical section so we
1229 * do not have to enter another one). The thread has already been moved to
1230 * our cpu's allq, but we must wait for the thread to be completely switched
1231 * out on the originating cpu before we schedule it on ours or the stack
1232 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1233 * change to main memory.
1235 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1236 * against wakeups. It is best if this interface is used only when there
1237 * are no pending events that might try to schedule the thread.
1241 lwkt_setcpu_remote(void *arg)
1244 globaldata_t gd = mycpu;
1246 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1250 td->td_flags &= ~TDF_MIGRATING;
1251 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1257 lwkt_preempted_proc(void)
1259 thread_t td = curthread;
1260 while (td->td_preempted)
1261 td = td->td_preempted;
1266 * Create a kernel process/thread/whatever. It shares it's address space
1267 * with proc0 - ie: kernel only.
1269 * NOTE! By default new threads are created with the MP lock held. A
1270 * thread which does not require the MP lock should release it by calling
1271 * rel_mplock() at the start of the new thread.
1274 lwkt_create(void (*func)(void *), void *arg,
1275 struct thread **tdp, thread_t template, int tdflags, int cpu,
1276 const char *fmt, ...)
1281 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1285 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1288 * Set up arg0 for 'ps' etc
1290 __va_start(ap, fmt);
1291 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1295 * Schedule the thread to run
1297 if ((td->td_flags & TDF_STOPREQ) == 0)
1300 td->td_flags &= ~TDF_STOPREQ;
1305 * kthread_* is specific to the kernel and is not needed by userland.
1310 * Destroy an LWKT thread. Warning! This function is not called when
1311 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1312 * uses a different reaping mechanism.
1317 thread_t td = curthread;
1320 if (td->td_flags & TDF_VERBOSE)
1321 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1323 crit_enter_quick(td);
1324 lwkt_deschedule_self(td);
1326 lwkt_remove_tdallq(td);
1327 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1328 ++gd->gd_tdfreecount;
1329 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1335 lwkt_remove_tdallq(thread_t td)
1337 KKASSERT(td->td_gd == mycpu);
1338 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1341 #endif /* _KERNEL */
1346 thread_t td = curthread;
1347 int lpri = td->td_pri;
1350 panic("td_pri is/would-go negative! %p %d", td, lpri);
1356 * Called from debugger/panic on cpus which have been stopped. We must still
1357 * process the IPIQ while stopped, even if we were stopped while in a critical
1360 * If we are dumping also try to process any pending interrupts. This may
1361 * or may not work depending on the state of the cpu at the point it was
1365 lwkt_smp_stopped(void)
1367 globaldata_t gd = mycpu;
1371 lwkt_process_ipiq();
1374 lwkt_process_ipiq();
1380 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1381 * get_mplock() has already incremented td_mpcount. We must block and
1382 * not return until giant is held.
1384 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1385 * reschedule the thread until it can obtain the giant lock for it.
1388 lwkt_mp_lock_contested(void)