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 1.101 2006/06/04 21:09:50 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/ipl.h>
75 #include <machine/smp.h>
79 #include <sys/stdint.h>
80 #include <libcaps/thread.h>
81 #include <sys/thread.h>
82 #include <sys/msgport.h>
83 #include <sys/errno.h>
84 #include <libcaps/globaldata.h>
85 #include <machine/cpufunc.h>
86 #include <sys/thread2.h>
87 #include <sys/msgport2.h>
91 #include <machine/lock.h>
92 #include <machine/atomic.h>
93 #include <machine/cpu.h>
97 static int untimely_switch = 0;
99 static int panic_on_cscount = 0;
101 static __int64_t switch_count = 0;
102 static __int64_t preempt_hit = 0;
103 static __int64_t preempt_miss = 0;
104 static __int64_t preempt_weird = 0;
105 static __int64_t token_contention_count = 0;
106 static __int64_t mplock_contention_count = 0;
110 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
112 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
119 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
120 &token_contention_count, 0, "spinning due to token contention");
121 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
122 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
131 #if !defined(KTR_GIANT_CONTENTION)
132 #define KTR_GIANT_CONTENTION KTR_ALL
135 KTR_INFO_MASTER(giant);
136 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
137 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
139 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
144 * These helper procedures handle the runq, they can only be called from
145 * within a critical section.
147 * WARNING! Prior to SMP being brought up it is possible to enqueue and
148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
149 * instead of 'mycpu' when referencing the globaldata structure. Once
150 * SMP live enqueuing and dequeueing only occurs on the current cpu.
154 _lwkt_dequeue(thread_t td)
156 if (td->td_flags & TDF_RUNQ) {
157 int nq = td->td_pri & TDPRI_MASK;
158 struct globaldata *gd = td->td_gd;
160 td->td_flags &= ~TDF_RUNQ;
161 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
162 /* runqmask is passively cleaned up by the switcher */
168 _lwkt_enqueue(thread_t td)
170 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
171 int nq = td->td_pri & TDPRI_MASK;
172 struct globaldata *gd = td->td_gd;
174 td->td_flags |= TDF_RUNQ;
175 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
176 gd->gd_runqmask |= 1 << nq;
181 * Schedule a thread to run. As the current thread we can always safely
182 * schedule ourselves, and a shortcut procedure is provided for that
185 * (non-blocking, self contained on a per cpu basis)
188 lwkt_schedule_self(thread_t td)
190 crit_enter_quick(td);
191 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
192 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
198 * Deschedule a thread.
200 * (non-blocking, self contained on a per cpu basis)
203 lwkt_deschedule_self(thread_t td)
205 crit_enter_quick(td);
213 * LWKTs operate on a per-cpu basis
215 * WARNING! Called from early boot, 'mycpu' may not work yet.
218 lwkt_gdinit(struct globaldata *gd)
222 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
223 TAILQ_INIT(&gd->gd_tdrunq[i]);
225 TAILQ_INIT(&gd->gd_tdallq);
231 * Create a new thread. The thread must be associated with a process context
232 * or LWKT start address before it can be scheduled. If the target cpu is
233 * -1 the thread will be created on the current cpu.
235 * If you intend to create a thread without a process context this function
236 * does everything except load the startup and switcher function.
239 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
242 globaldata_t gd = mycpu;
246 if (gd->gd_tdfreecount > 0) {
247 --gd->gd_tdfreecount;
248 td = TAILQ_FIRST(&gd->gd_tdfreeq);
249 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
250 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
251 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
253 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
257 td = zalloc(thread_zone);
259 td = malloc(sizeof(struct thread));
261 td->td_kstack = NULL;
262 td->td_kstack_size = 0;
263 flags |= TDF_ALLOCATED_THREAD;
266 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
267 if (flags & TDF_ALLOCATED_STACK) {
269 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
271 libcaps_free_stack(stack, td->td_kstack_size);
278 stack = (void *)kmem_alloc(kernel_map, stksize);
280 stack = libcaps_alloc_stack(stksize);
282 flags |= TDF_ALLOCATED_STACK;
285 lwkt_init_thread(td, stack, stksize, flags, mycpu);
287 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
294 * Initialize a preexisting thread structure. This function is used by
295 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
297 * All threads start out in a critical section at a priority of
298 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
299 * appropriate. This function may send an IPI message when the
300 * requested cpu is not the current cpu and consequently gd_tdallq may
301 * not be initialized synchronously from the point of view of the originating
304 * NOTE! we have to be careful in regards to creating threads for other cpus
305 * if SMP has not yet been activated.
310 lwkt_init_thread_remote(void *arg)
315 * Protected by critical section held by IPI dispatch
317 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
323 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
324 struct globaldata *gd)
326 globaldata_t mygd = mycpu;
328 bzero(td, sizeof(struct thread));
329 td->td_kstack = stack;
330 td->td_kstack_size = stksize;
331 td->td_flags = flags;
333 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
335 if ((flags & TDF_MPSAFE) == 0)
338 lwkt_initport(&td->td_msgport, td);
339 pmap_init_thread(td);
342 * Normally initializing a thread for a remote cpu requires sending an
343 * IPI. However, the idlethread is setup before the other cpus are
344 * activated so we have to treat it as a special case. XXX manipulation
345 * of gd_tdallq requires the BGL.
347 if (gd == mygd || td == &gd->gd_idlethread) {
349 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
352 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
356 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
364 lwkt_set_comm(thread_t td, const char *ctl, ...)
369 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
374 lwkt_hold(thread_t td)
380 lwkt_rele(thread_t td)
382 KKASSERT(td->td_refs > 0);
389 lwkt_wait_free(thread_t td)
392 tsleep(td, 0, "tdreap", hz);
398 lwkt_free_thread(thread_t td)
400 struct globaldata *gd = mycpu;
402 KASSERT((td->td_flags & TDF_RUNNING) == 0,
403 ("lwkt_free_thread: did not exit! %p", td));
406 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
407 (td->td_flags & TDF_ALLOCATED_THREAD)
409 ++gd->gd_tdfreecount;
410 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
414 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
416 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
418 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
421 td->td_kstack = NULL;
422 td->td_kstack_size = 0;
424 if (td->td_flags & TDF_ALLOCATED_THREAD) {
426 zfree(thread_zone, td);
436 * Switch to the next runnable lwkt. If no LWKTs are runnable then
437 * switch to the idlethread. Switching must occur within a critical
438 * section to avoid races with the scheduling queue.
440 * We always have full control over our cpu's run queue. Other cpus
441 * that wish to manipulate our queue must use the cpu_*msg() calls to
442 * talk to our cpu, so a critical section is all that is needed and
443 * the result is very, very fast thread switching.
445 * The LWKT scheduler uses a fixed priority model and round-robins at
446 * each priority level. User process scheduling is a totally
447 * different beast and LWKT priorities should not be confused with
448 * user process priorities.
450 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
451 * cleans it up. Note that the td_switch() function cannot do anything that
452 * requires the MP lock since the MP lock will have already been setup for
453 * the target thread (not the current thread). It's nice to have a scheduler
454 * that does not need the MP lock to work because it allows us to do some
455 * really cool high-performance MP lock optimizations.
457 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
458 * is not called by the current thread in the preemption case, only when
459 * the preempting thread blocks (in order to return to the original thread).
464 globaldata_t gd = mycpu;
465 thread_t td = gd->gd_curthread;
472 * Switching from within a 'fast' (non thread switched) interrupt or IPI
473 * is illegal. However, we may have to do it anyway if we hit a fatal
474 * kernel trap or we have paniced.
476 * If this case occurs save and restore the interrupt nesting level.
478 if (gd->gd_intr_nesting_level) {
482 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
483 panic("lwkt_switch: cannot switch from within "
484 "a fast interrupt, yet, td %p\n", td);
486 savegdnest = gd->gd_intr_nesting_level;
487 savegdtrap = gd->gd_trap_nesting_level;
488 gd->gd_intr_nesting_level = 0;
489 gd->gd_trap_nesting_level = 0;
490 if ((td->td_flags & TDF_PANICWARN) == 0) {
491 td->td_flags |= TDF_PANICWARN;
492 printf("Warning: thread switch from interrupt or IPI, "
493 "thread %p (%s)\n", td, td->td_comm);
495 db_print_backtrace();
499 gd->gd_intr_nesting_level = savegdnest;
500 gd->gd_trap_nesting_level = savegdtrap;
506 * Passive release (used to transition from user to kernel mode
507 * when we block or switch rather then when we enter the kernel).
508 * This function is NOT called if we are switching into a preemption
509 * or returning from a preemption. Typically this causes us to lose
510 * our current process designation (if we have one) and become a true
511 * LWKT thread, and may also hand the current process designation to
512 * another process and schedule thread.
520 lwkt_relalltokens(td);
524 * We had better not be holding any spin locks, but don't get into an
525 * endless panic loop.
527 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
528 ("lwkt_switch: still holding a shared spinlock %p!",
529 gd->gd_spinlock_rd));
530 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
531 ("lwkt_switch: still holding %d exclusive spinlocks!",
532 gd->gd_spinlocks_wr));
537 * td_mpcount cannot be used to determine if we currently hold the
538 * MP lock because get_mplock() will increment it prior to attempting
539 * to get the lock, and switch out if it can't. Our ownership of
540 * the actual lock will remain stable while we are in a critical section
541 * (but, of course, another cpu may own or release the lock so the
542 * actual value of mp_lock is not stable).
544 mpheld = MP_LOCK_HELD();
546 if (td->td_cscount) {
547 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
549 if (panic_on_cscount)
550 panic("switching while mastering cpusync");
554 if ((ntd = td->td_preempted) != NULL) {
556 * We had preempted another thread on this cpu, resume the preempted
557 * thread. This occurs transparently, whether the preempted thread
558 * was scheduled or not (it may have been preempted after descheduling
561 * We have to setup the MP lock for the original thread after backing
562 * out the adjustment that was made to curthread when the original
565 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
567 if (ntd->td_mpcount && mpheld == 0) {
568 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
569 td, ntd, td->td_mpcount, ntd->td_mpcount);
571 if (ntd->td_mpcount) {
572 td->td_mpcount -= ntd->td_mpcount;
573 KKASSERT(td->td_mpcount >= 0);
576 ntd->td_flags |= TDF_PREEMPT_DONE;
579 * XXX. The interrupt may have woken a thread up, we need to properly
580 * set the reschedule flag if the originally interrupted thread is at
583 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
585 /* YYY release mp lock on switchback if original doesn't need it */
588 * Priority queue / round-robin at each priority. Note that user
589 * processes run at a fixed, low priority and the user process
590 * scheduler deals with interactions between user processes
591 * by scheduling and descheduling them from the LWKT queue as
594 * We have to adjust the MP lock for the target thread. If we
595 * need the MP lock and cannot obtain it we try to locate a
596 * thread that does not need the MP lock. If we cannot, we spin
599 * A similar issue exists for the tokens held by the target thread.
600 * If we cannot obtain ownership of the tokens we cannot immediately
601 * schedule the thread.
605 * If an LWKT reschedule was requested, well that is what we are
606 * doing now so clear it.
608 clear_lwkt_resched();
610 if (gd->gd_runqmask) {
611 int nq = bsrl(gd->gd_runqmask);
612 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
613 gd->gd_runqmask &= ~(1 << nq);
618 * THREAD SELECTION FOR AN SMP MACHINE BUILD
620 * If the target needs the MP lock and we couldn't get it,
621 * or if the target is holding tokens and we could not
622 * gain ownership of the tokens, continue looking for a
623 * thread to schedule and spin instead of HLT if we can't.
625 * NOTE: the mpheld variable invalid after this conditional, it
626 * can change due to both cpu_try_mplock() returning success
627 * AND interactions in lwkt_getalltokens() due to the fact that
628 * we are trying to check the mpcount of a thread other then
629 * the current thread. Because of this, if the current thread
630 * is not holding td_mpcount, an IPI indirectly run via
631 * lwkt_getalltokens() can obtain and release the MP lock and
632 * cause the core MP lock to be released.
634 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
635 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
637 u_int32_t rqmask = gd->gd_runqmask;
639 mpheld = MP_LOCK_HELD();
642 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
643 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
644 /* spinning due to MP lock being held */
646 ++mplock_contention_count;
648 /* mplock still not held, 'mpheld' still valid */
653 * mpheld state invalid after getalltokens call returns
654 * failure, but the variable is only needed for
657 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
658 /* spinning due to token contention */
660 ++token_contention_count;
662 mpheld = MP_LOCK_HELD();
669 rqmask &= ~(1 << nq);
673 ntd = &gd->gd_idlethread;
674 ntd->td_flags |= TDF_IDLE_NOHLT;
675 goto using_idle_thread;
677 ++gd->gd_cnt.v_swtch;
678 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
679 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
682 ++gd->gd_cnt.v_swtch;
683 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
684 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
688 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
689 * worry about tokens or the BGL.
691 ++gd->gd_cnt.v_swtch;
692 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
693 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
697 * We have nothing to run but only let the idle loop halt
698 * the cpu if there are no pending interrupts.
700 ntd = &gd->gd_idlethread;
701 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
702 ntd->td_flags |= TDF_IDLE_NOHLT;
706 * The idle thread should not be holding the MP lock unless we
707 * are trapping in the kernel or in a panic. Since we select the
708 * idle thread unconditionally when no other thread is available,
709 * if the MP lock is desired during a panic or kernel trap, we
710 * have to loop in the scheduler until we get it.
712 if (ntd->td_mpcount) {
713 mpheld = MP_LOCK_HELD();
714 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
715 panic("Idle thread %p was holding the BGL!", ntd);
716 else if (mpheld == 0)
722 KASSERT(ntd->td_pri >= TDPRI_CRIT,
723 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
726 * Do the actual switch. If the new target does not need the MP lock
727 * and we are holding it, release the MP lock. If the new target requires
728 * the MP lock we have already acquired it for the target.
731 if (ntd->td_mpcount == 0 ) {
735 ASSERT_MP_LOCK_HELD(ntd);
742 /* NOTE: current cpu may have changed after switch */
747 * Request that the target thread preempt the current thread. Preemption
748 * only works under a specific set of conditions:
750 * - We are not preempting ourselves
751 * - The target thread is owned by the current cpu
752 * - We are not currently being preempted
753 * - The target is not currently being preempted
754 * - We are able to satisfy the target's MP lock requirements (if any).
756 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
757 * this is called via lwkt_schedule() through the td_preemptable callback.
758 * critpri is the managed critical priority that we should ignore in order
759 * to determine whether preemption is possible (aka usually just the crit
760 * priority of lwkt_schedule() itself).
762 * XXX at the moment we run the target thread in a critical section during
763 * the preemption in order to prevent the target from taking interrupts
764 * that *WE* can't. Preemption is strictly limited to interrupt threads
765 * and interrupt-like threads, outside of a critical section, and the
766 * preempted source thread will be resumed the instant the target blocks
767 * whether or not the source is scheduled (i.e. preemption is supposed to
768 * be as transparent as possible).
770 * The target thread inherits our MP count (added to its own) for the
771 * duration of the preemption in order to preserve the atomicy of the
772 * MP lock during the preemption. Therefore, any preempting targets must be
773 * careful in regards to MP assertions. Note that the MP count may be
774 * out of sync with the physical mp_lock, but we do not have to preserve
775 * the original ownership of the lock if it was out of synch (that is, we
776 * can leave it synchronized on return).
779 lwkt_preempt(thread_t ntd, int critpri)
781 struct globaldata *gd = mycpu;
789 * The caller has put us in a critical section. We can only preempt
790 * if the caller of the caller was not in a critical section (basically
791 * a local interrupt), as determined by the 'critpri' parameter. We
792 * also acn't preempt if the caller is holding any spinlocks (even if
793 * he isn't in a critical section). This also handles the tokens test.
795 * YYY The target thread must be in a critical section (else it must
796 * inherit our critical section? I dunno yet).
798 * Set need_lwkt_resched() unconditionally for now YYY.
800 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
802 td = gd->gd_curthread;
803 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
807 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
813 if (ntd->td_gd != gd) {
820 * Take the easy way out and do not preempt if the target is holding
821 * any spinlocks. We could test whether the thread(s) being
822 * preempted interlock against the target thread's tokens and whether
823 * we can get all the target thread's tokens, but this situation
824 * should not occur very often so its easier to simply not preempt.
825 * Also, plain spinlocks are impossible to figure out at this point so
826 * just don't preempt.
828 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
833 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
838 if (ntd->td_preempted) {
845 * note: an interrupt might have occured just as we were transitioning
846 * to or from the MP lock. In this case td_mpcount will be pre-disposed
847 * (non-zero) but not actually synchronized with the actual state of the
848 * lock. We can use it to imply an MP lock requirement for the
849 * preemption but we cannot use it to test whether we hold the MP lock
852 savecnt = td->td_mpcount;
853 mpheld = MP_LOCK_HELD();
854 ntd->td_mpcount += td->td_mpcount;
855 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
856 ntd->td_mpcount -= td->td_mpcount;
864 * Since we are able to preempt the current thread, there is no need to
865 * call need_lwkt_resched().
868 ntd->td_preempted = td;
869 td->td_flags |= TDF_PREEMPT_LOCK;
871 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
873 KKASSERT(savecnt == td->td_mpcount);
874 mpheld = MP_LOCK_HELD();
875 if (mpheld && td->td_mpcount == 0)
877 else if (mpheld == 0 && td->td_mpcount)
878 panic("lwkt_preempt(): MP lock was not held through");
880 ntd->td_preempted = NULL;
881 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
885 * Yield our thread while higher priority threads are pending. This is
886 * typically called when we leave a critical section but it can be safely
887 * called while we are in a critical section.
889 * This function will not generally yield to equal priority threads but it
890 * can occur as a side effect. Note that lwkt_switch() is called from
891 * inside the critical section to prevent its own crit_exit() from reentering
892 * lwkt_yield_quick().
894 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
895 * came along but was blocked and made pending.
897 * (self contained on a per cpu basis)
900 lwkt_yield_quick(void)
902 globaldata_t gd = mycpu;
903 thread_t td = gd->gd_curthread;
906 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
907 * it with a non-zero cpl then we might not wind up calling splz after
908 * a task switch when the critical section is exited even though the
909 * new task could accept the interrupt.
911 * XXX from crit_exit() only called after last crit section is released.
912 * If called directly will run splz() even if in a critical section.
914 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
915 * except for this special case, we MUST call splz() here to handle any
916 * pending ints, particularly after we switch, or we might accidently
917 * halt the cpu with interrupts pending.
919 if (gd->gd_reqflags && td->td_nest_count < 2)
923 * YYY enabling will cause wakeup() to task-switch, which really
924 * confused the old 4.x code. This is a good way to simulate
925 * preemption and MP without actually doing preemption or MP, because a
926 * lot of code assumes that wakeup() does not block.
928 if (untimely_switch && td->td_nest_count == 0 &&
929 gd->gd_intr_nesting_level == 0
931 crit_enter_quick(td);
933 * YYY temporary hacks until we disassociate the userland scheduler
934 * from the LWKT scheduler.
936 if (td->td_flags & TDF_RUNQ) {
937 lwkt_switch(); /* will not reenter yield function */
939 lwkt_schedule_self(td); /* make sure we are scheduled */
940 lwkt_switch(); /* will not reenter yield function */
941 lwkt_deschedule_self(td); /* make sure we are descheduled */
943 crit_exit_noyield(td);
948 * This implements a normal yield which, unlike _quick, will yield to equal
949 * priority threads as well. Note that gd_reqflags tests will be handled by
950 * the crit_exit() call in lwkt_switch().
952 * (self contained on a per cpu basis)
957 lwkt_schedule_self(curthread);
962 * Generic schedule. Possibly schedule threads belonging to other cpus and
963 * deal with threads that might be blocked on a wait queue.
965 * We have a little helper inline function which does additional work after
966 * the thread has been enqueued, including dealing with preemption and
967 * setting need_lwkt_resched() (which prevents the kernel from returning
968 * to userland until it has processed higher priority threads).
970 * It is possible for this routine to be called after a failed _enqueue
971 * (due to the target thread migrating, sleeping, or otherwise blocked).
972 * We have to check that the thread is actually on the run queue!
976 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
978 if (ntd->td_flags & TDF_RUNQ) {
979 if (ntd->td_preemptable) {
980 ntd->td_preemptable(ntd, cpri); /* YYY +token */
981 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
982 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
990 lwkt_schedule(thread_t td)
992 globaldata_t mygd = mycpu;
994 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
996 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
997 if (td == mygd->gd_curthread) {
1001 * If we own the thread, there is no race (since we are in a
1002 * critical section). If we do not own the thread there might
1003 * be a race but the target cpu will deal with it.
1006 if (td->td_gd == mygd) {
1008 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1010 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1014 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1023 * Thread migration using a 'Pull' method. The thread may or may not be
1024 * the current thread. It MUST be descheduled and in a stable state.
1025 * lwkt_giveaway() must be called on the cpu owning the thread.
1027 * At any point after lwkt_giveaway() is called, the target cpu may
1028 * 'pull' the thread by calling lwkt_acquire().
1030 * MPSAFE - must be called under very specific conditions.
1033 lwkt_giveaway(thread_t td)
1035 globaldata_t gd = mycpu;
1038 KKASSERT(td->td_gd == gd);
1039 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1040 td->td_flags |= TDF_MIGRATING;
1045 lwkt_acquire(thread_t td)
1050 KKASSERT(td->td_flags & TDF_MIGRATING);
1055 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1056 crit_enter_gd(mygd);
1057 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1060 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1061 td->td_flags &= ~TDF_MIGRATING;
1064 crit_enter_gd(mygd);
1065 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1066 td->td_flags &= ~TDF_MIGRATING;
1074 * Generic deschedule. Descheduling threads other then your own should be
1075 * done only in carefully controlled circumstances. Descheduling is
1078 * This function may block if the cpu has run out of messages.
1081 lwkt_deschedule(thread_t td)
1085 if (td == curthread) {
1088 if (td->td_gd == mycpu) {
1091 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1101 * Set the target thread's priority. This routine does not automatically
1102 * switch to a higher priority thread, LWKT threads are not designed for
1103 * continuous priority changes. Yield if you want to switch.
1105 * We have to retain the critical section count which uses the high bits
1106 * of the td_pri field. The specified priority may also indicate zero or
1107 * more critical sections by adding TDPRI_CRIT*N.
1109 * Note that we requeue the thread whether it winds up on a different runq
1110 * or not. uio_yield() depends on this and the routine is not normally
1111 * called with the same priority otherwise.
1114 lwkt_setpri(thread_t td, int pri)
1117 KKASSERT(td->td_gd == mycpu);
1119 if (td->td_flags & TDF_RUNQ) {
1121 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1124 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1130 lwkt_setpri_self(int pri)
1132 thread_t td = curthread;
1134 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1136 if (td->td_flags & TDF_RUNQ) {
1138 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1141 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1147 * Determine if there is a runnable thread at a higher priority then
1148 * the current thread. lwkt_setpri() does not check this automatically.
1149 * Return 1 if there is, 0 if there isn't.
1151 * Example: if bit 31 of runqmask is set and the current thread is priority
1152 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1154 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1155 * up comparing against 0xffffffff, a comparison that will always be false.
1158 lwkt_checkpri_self(void)
1160 globaldata_t gd = mycpu;
1161 thread_t td = gd->gd_curthread;
1162 int nq = td->td_pri & TDPRI_MASK;
1164 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1165 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1173 * Migrate the current thread to the specified cpu.
1175 * This is accomplished by descheduling ourselves from the current cpu,
1176 * moving our thread to the tdallq of the target cpu, IPI messaging the
1177 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1178 * races while the thread is being migrated.
1181 static void lwkt_setcpu_remote(void *arg);
1185 lwkt_setcpu_self(globaldata_t rgd)
1188 thread_t td = curthread;
1190 if (td->td_gd != rgd) {
1191 crit_enter_quick(td);
1192 td->td_flags |= TDF_MIGRATING;
1193 lwkt_deschedule_self(td);
1194 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1195 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1197 /* we are now on the target cpu */
1198 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1199 crit_exit_quick(td);
1205 lwkt_migratecpu(int cpuid)
1210 rgd = globaldata_find(cpuid);
1211 lwkt_setcpu_self(rgd);
1216 * Remote IPI for cpu migration (called while in a critical section so we
1217 * do not have to enter another one). The thread has already been moved to
1218 * our cpu's allq, but we must wait for the thread to be completely switched
1219 * out on the originating cpu before we schedule it on ours or the stack
1220 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1221 * change to main memory.
1223 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1224 * against wakeups. It is best if this interface is used only when there
1225 * are no pending events that might try to schedule the thread.
1229 lwkt_setcpu_remote(void *arg)
1232 globaldata_t gd = mycpu;
1234 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1238 td->td_flags &= ~TDF_MIGRATING;
1239 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
1245 lwkt_preempted_proc(void)
1247 thread_t td = curthread;
1248 while (td->td_preempted)
1249 td = td->td_preempted;
1254 * Create a kernel process/thread/whatever. It shares it's address space
1255 * with proc0 - ie: kernel only.
1257 * NOTE! By default new threads are created with the MP lock held. A
1258 * thread which does not require the MP lock should release it by calling
1259 * rel_mplock() at the start of the new thread.
1262 lwkt_create(void (*func)(void *), void *arg,
1263 struct thread **tdp, thread_t template, int tdflags, int cpu,
1264 const char *fmt, ...)
1269 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1270 tdflags | TDF_VERBOSE);
1273 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1276 * Set up arg0 for 'ps' etc
1278 __va_start(ap, fmt);
1279 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1283 * Schedule the thread to run
1285 if ((td->td_flags & TDF_STOPREQ) == 0)
1288 td->td_flags &= ~TDF_STOPREQ;
1293 * kthread_* is specific to the kernel and is not needed by userland.
1298 * Destroy an LWKT thread. Warning! This function is not called when
1299 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1300 * uses a different reaping mechanism.
1305 thread_t td = curthread;
1308 if (td->td_flags & TDF_VERBOSE)
1309 printf("kthread %p %s has exited\n", td, td->td_comm);
1311 crit_enter_quick(td);
1312 lwkt_deschedule_self(td);
1314 lwkt_remove_tdallq(td);
1315 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1316 ++gd->gd_tdfreecount;
1317 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1323 lwkt_remove_tdallq(thread_t td)
1325 KKASSERT(td->td_gd == mycpu);
1326 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1329 #endif /* _KERNEL */
1334 thread_t td = curthread;
1335 int lpri = td->td_pri;
1338 panic("td_pri is/would-go negative! %p %d", td, lpri);
1344 * Called from debugger/panic on cpus which have been stopped. We must still
1345 * process the IPIQ while stopped, even if we were stopped while in a critical
1348 * If we are dumping also try to process any pending interrupts. This may
1349 * or may not work depending on the state of the cpu at the point it was
1353 lwkt_smp_stopped(void)
1355 globaldata_t gd = mycpu;
1359 lwkt_process_ipiq();
1362 lwkt_process_ipiq();
1368 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1369 * get_mplock() has already incremented td_mpcount. We must block and
1370 * not return until giant is held.
1372 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1373 * reschedule the thread until it can obtain the giant lock for it.
1376 lwkt_mp_lock_contested(void)