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 1.92 2006/03/01 00:17:55 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/thread2.h>
53 #include <sys/sysctl.h>
54 #include <sys/kthread.h>
55 #include <machine/cpu.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>
67 #include <vm/vm_zone.h>
69 #include <machine/stdarg.h>
70 #include <machine/ipl.h>
71 #include <machine/smp.h>
75 #include <sys/stdint.h>
76 #include <libcaps/thread.h>
77 #include <sys/thread.h>
78 #include <sys/msgport.h>
79 #include <sys/errno.h>
80 #include <libcaps/globaldata.h>
81 #include <machine/cpufunc.h>
82 #include <sys/thread2.h>
83 #include <sys/msgport2.h>
87 #include <machine/lock.h>
88 #include <machine/atomic.h>
89 #include <machine/cpu.h>
93 static int untimely_switch = 0;
95 static int panic_on_cscount = 0;
97 static __int64_t switch_count = 0;
98 static __int64_t preempt_hit = 0;
99 static __int64_t preempt_miss = 0;
100 static __int64_t preempt_weird = 0;
101 static __int64_t token_contention_count = 0;
102 static __int64_t mplock_contention_count = 0;
106 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
108 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
112 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
116 &token_contention_count, 0, "spinning due to token contention");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
118 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
133 _lwkt_dequeue(thread_t td)
135 if (td->td_flags & TDF_RUNQ) {
136 int nq = td->td_pri & TDPRI_MASK;
137 struct globaldata *gd = td->td_gd;
139 td->td_flags &= ~TDF_RUNQ;
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
147 _lwkt_enqueue(thread_t td)
149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
150 int nq = td->td_pri & TDPRI_MASK;
151 struct globaldata *gd = td->td_gd;
153 td->td_flags |= TDF_RUNQ;
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
160 * Schedule a thread to run. As the current thread we can always safely
161 * schedule ourselves, and a shortcut procedure is provided for that
164 * (non-blocking, self contained on a per cpu basis)
167 lwkt_schedule_self(thread_t td)
169 crit_enter_quick(td);
170 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
171 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
172 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
178 * Deschedule a thread.
180 * (non-blocking, self contained on a per cpu basis)
183 lwkt_deschedule_self(thread_t td)
185 crit_enter_quick(td);
186 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
194 * LWKTs operate on a per-cpu basis
196 * WARNING! Called from early boot, 'mycpu' may not work yet.
199 lwkt_gdinit(struct globaldata *gd)
203 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
204 TAILQ_INIT(&gd->gd_tdrunq[i]);
206 TAILQ_INIT(&gd->gd_tdallq);
212 * Initialize a thread wait structure prior to first use.
214 * NOTE! called from low level boot code, we cannot do anything fancy!
217 lwkt_wait_init(lwkt_wait_t w)
219 lwkt_token_init(&w->wa_token);
220 TAILQ_INIT(&w->wa_waitq);
226 * Create a new thread. The thread must be associated with a process context
227 * or LWKT start address before it can be scheduled. If the target cpu is
228 * -1 the thread will be created on the current cpu.
230 * If you intend to create a thread without a process context this function
231 * does everything except load the startup and switcher function.
234 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
237 globaldata_t gd = mycpu;
241 if (gd->gd_tdfreecount > 0) {
242 --gd->gd_tdfreecount;
243 td = TAILQ_FIRST(&gd->gd_tdfreeq);
244 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
245 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
246 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
248 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
252 td = zalloc(thread_zone);
254 td = malloc(sizeof(struct thread));
256 td->td_kstack = NULL;
257 td->td_kstack_size = 0;
258 flags |= TDF_ALLOCATED_THREAD;
261 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
262 if (flags & TDF_ALLOCATED_STACK) {
264 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
266 libcaps_free_stack(stack, td->td_kstack_size);
273 stack = (void *)kmem_alloc(kernel_map, stksize);
275 stack = libcaps_alloc_stack(stksize);
277 flags |= TDF_ALLOCATED_STACK;
280 lwkt_init_thread(td, stack, stksize, flags, mycpu);
282 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
289 * Initialize a preexisting thread structure. This function is used by
290 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
292 * All threads start out in a critical section at a priority of
293 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
294 * appropriate. This function may send an IPI message when the
295 * requested cpu is not the current cpu and consequently gd_tdallq may
296 * not be initialized synchronously from the point of view of the originating
299 * NOTE! we have to be careful in regards to creating threads for other cpus
300 * if SMP has not yet been activated.
305 lwkt_init_thread_remote(void *arg)
309 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
315 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
316 struct globaldata *gd)
318 globaldata_t mygd = mycpu;
320 bzero(td, sizeof(struct thread));
321 td->td_kstack = stack;
322 td->td_kstack_size = stksize;
323 td->td_flags = flags;
325 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
327 if ((flags & TDF_MPSAFE) == 0)
330 lwkt_initport(&td->td_msgport, td);
331 pmap_init_thread(td);
334 * Normally initializing a thread for a remote cpu requires sending an
335 * IPI. However, the idlethread is setup before the other cpus are
336 * activated so we have to treat it as a special case. XXX manipulation
337 * of gd_tdallq requires the BGL.
339 if (gd == mygd || td == &gd->gd_idlethread) {
341 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
344 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
348 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
356 lwkt_set_comm(thread_t td, const char *ctl, ...)
361 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
366 lwkt_hold(thread_t td)
372 lwkt_rele(thread_t td)
374 KKASSERT(td->td_refs > 0);
381 lwkt_wait_free(thread_t td)
384 tsleep(td, 0, "tdreap", hz);
390 lwkt_free_thread(thread_t td)
392 struct globaldata *gd = mycpu;
394 KASSERT((td->td_flags & TDF_RUNNING) == 0,
395 ("lwkt_free_thread: did not exit! %p", td));
398 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* Protected by BGL */
399 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
400 (td->td_flags & TDF_ALLOCATED_THREAD)
402 ++gd->gd_tdfreecount;
403 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
407 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
409 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
411 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
414 td->td_kstack = NULL;
415 td->td_kstack_size = 0;
417 if (td->td_flags & TDF_ALLOCATED_THREAD) {
419 zfree(thread_zone, td);
429 * Switch to the next runnable lwkt. If no LWKTs are runnable then
430 * switch to the idlethread. Switching must occur within a critical
431 * section to avoid races with the scheduling queue.
433 * We always have full control over our cpu's run queue. Other cpus
434 * that wish to manipulate our queue must use the cpu_*msg() calls to
435 * talk to our cpu, so a critical section is all that is needed and
436 * the result is very, very fast thread switching.
438 * The LWKT scheduler uses a fixed priority model and round-robins at
439 * each priority level. User process scheduling is a totally
440 * different beast and LWKT priorities should not be confused with
441 * user process priorities.
443 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
444 * cleans it up. Note that the td_switch() function cannot do anything that
445 * requires the MP lock since the MP lock will have already been setup for
446 * the target thread (not the current thread). It's nice to have a scheduler
447 * that does not need the MP lock to work because it allows us to do some
448 * really cool high-performance MP lock optimizations.
450 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
451 * is not called by the current thread in the preemption case, only when
452 * the preempting thread blocks (in order to return to the original thread).
457 globaldata_t gd = mycpu;
458 thread_t td = gd->gd_curthread;
465 * We had better not be holding any spin locks.
467 KKASSERT(td->td_spinlocks == 0);
470 * Switching from within a 'fast' (non thread switched) interrupt or IPI
471 * is illegal. However, we may have to do it anyway if we hit a fatal
472 * kernel trap or we have paniced.
474 * If this case occurs save and restore the interrupt nesting level.
476 if (gd->gd_intr_nesting_level) {
480 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
481 panic("lwkt_switch: cannot switch from within "
482 "a fast interrupt, yet, td %p\n", td);
484 savegdnest = gd->gd_intr_nesting_level;
485 savegdtrap = gd->gd_trap_nesting_level;
486 gd->gd_intr_nesting_level = 0;
487 gd->gd_trap_nesting_level = 0;
488 if ((td->td_flags & TDF_PANICWARN) == 0) {
489 td->td_flags |= TDF_PANICWARN;
490 printf("Warning: thread switch from interrupt or IPI, "
491 "thread %p (%s)\n", td, td->td_comm);
493 db_print_backtrace();
497 gd->gd_intr_nesting_level = savegdnest;
498 gd->gd_trap_nesting_level = savegdtrap;
504 * Passive release (used to transition from user to kernel mode
505 * when we block or switch rather then when we enter the kernel).
506 * This function is NOT called if we are switching into a preemption
507 * or returning from a preemption. Typically this causes us to lose
508 * our current process designation (if we have one) and become a true
509 * LWKT thread, and may also hand the current process designation to
510 * another process and schedule thread.
519 * td_mpcount cannot be used to determine if we currently hold the
520 * MP lock because get_mplock() will increment it prior to attempting
521 * to get the lock, and switch out if it can't. Our ownership of
522 * the actual lock will remain stable while we are in a critical section
523 * (but, of course, another cpu may own or release the lock so the
524 * actual value of mp_lock is not stable).
526 mpheld = MP_LOCK_HELD();
528 if (td->td_cscount) {
529 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
531 if (panic_on_cscount)
532 panic("switching while mastering cpusync");
536 if ((ntd = td->td_preempted) != NULL) {
538 * We had preempted another thread on this cpu, resume the preempted
539 * thread. This occurs transparently, whether the preempted thread
540 * was scheduled or not (it may have been preempted after descheduling
543 * We have to setup the MP lock for the original thread after backing
544 * out the adjustment that was made to curthread when the original
547 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
549 if (ntd->td_mpcount && mpheld == 0) {
550 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
551 td, ntd, td->td_mpcount, ntd->td_mpcount);
553 if (ntd->td_mpcount) {
554 td->td_mpcount -= ntd->td_mpcount;
555 KKASSERT(td->td_mpcount >= 0);
558 ntd->td_flags |= TDF_PREEMPT_DONE;
561 * XXX. The interrupt may have woken a thread up, we need to properly
562 * set the reschedule flag if the originally interrupted thread is at
565 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
567 /* YYY release mp lock on switchback if original doesn't need it */
570 * Priority queue / round-robin at each priority. Note that user
571 * processes run at a fixed, low priority and the user process
572 * scheduler deals with interactions between user processes
573 * by scheduling and descheduling them from the LWKT queue as
576 * We have to adjust the MP lock for the target thread. If we
577 * need the MP lock and cannot obtain it we try to locate a
578 * thread that does not need the MP lock. If we cannot, we spin
581 * A similar issue exists for the tokens held by the target thread.
582 * If we cannot obtain ownership of the tokens we cannot immediately
583 * schedule the thread.
587 * We are switching threads. If there are any pending requests for
588 * tokens we can satisfy all of them here.
591 if (gd->gd_tokreqbase)
592 lwkt_drain_token_requests();
596 * If an LWKT reschedule was requested, well that is what we are
597 * doing now so clear it.
599 clear_lwkt_resched();
601 if (gd->gd_runqmask) {
602 int nq = bsrl(gd->gd_runqmask);
603 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
604 gd->gd_runqmask &= ~(1 << nq);
609 * THREAD SELECTION FOR AN SMP MACHINE BUILD
611 * If the target needs the MP lock and we couldn't get it,
612 * or if the target is holding tokens and we could not
613 * gain ownership of the tokens, continue looking for a
614 * thread to schedule and spin instead of HLT if we can't.
616 * NOTE: the mpheld variable invalid after this conditional, it
617 * can change due to both cpu_try_mplock() returning success
618 * AND interactions in lwkt_chktokens() due to the fact that
619 * we are trying to check the mpcount of a thread other then
620 * the current thread. Because of this, if the current thread
621 * is not holding td_mpcount, an IPI indirectly run via
622 * lwkt_chktokens() can obtain and release the MP lock and
623 * cause the core MP lock to be released.
625 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
626 (ntd->td_toks && lwkt_chktokens(ntd) == 0)
628 u_int32_t rqmask = gd->gd_runqmask;
630 mpheld = MP_LOCK_HELD();
633 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
634 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
635 /* spinning due to MP lock being held */
637 ++mplock_contention_count;
639 /* mplock still not held, 'mpheld' still valid */
644 * mpheld state invalid after chktokens call returns
645 * failure, but the variable is only needed for
648 if (ntd->td_toks && !lwkt_chktokens(ntd)) {
649 /* spinning due to token contention */
651 ++token_contention_count;
653 mpheld = MP_LOCK_HELD();
660 rqmask &= ~(1 << nq);
664 ntd = &gd->gd_idlethread;
665 ntd->td_flags |= TDF_IDLE_NOHLT;
666 goto using_idle_thread;
668 ++gd->gd_cnt.v_swtch;
669 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
670 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
673 ++gd->gd_cnt.v_swtch;
674 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
675 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
679 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
680 * worry about tokens or the BGL.
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 * We have nothing to run but only let the idle loop halt
689 * the cpu if there are no pending interrupts.
691 ntd = &gd->gd_idlethread;
692 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
693 ntd->td_flags |= TDF_IDLE_NOHLT;
697 * The idle thread should not be holding the MP lock unless we
698 * are trapping in the kernel or in a panic. Since we select the
699 * idle thread unconditionally when no other thread is available,
700 * if the MP lock is desired during a panic or kernel trap, we
701 * have to loop in the scheduler until we get it.
703 if (ntd->td_mpcount) {
704 mpheld = MP_LOCK_HELD();
705 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
706 panic("Idle thread %p was holding the BGL!", ntd);
707 else if (mpheld == 0)
713 KASSERT(ntd->td_pri >= TDPRI_CRIT,
714 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
717 * Do the actual switch. If the new target does not need the MP lock
718 * and we are holding it, release the MP lock. If the new target requires
719 * the MP lock we have already acquired it for the target.
722 if (ntd->td_mpcount == 0 ) {
726 ASSERT_MP_LOCK_HELD(ntd);
733 /* NOTE: current cpu may have changed after switch */
738 * Request that the target thread preempt the current thread. Preemption
739 * only works under a specific set of conditions:
741 * - We are not preempting ourselves
742 * - The target thread is owned by the current cpu
743 * - We are not currently being preempted
744 * - The target is not currently being preempted
745 * - We are able to satisfy the target's MP lock requirements (if any).
747 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
748 * this is called via lwkt_schedule() through the td_preemptable callback.
749 * critpri is the managed critical priority that we should ignore in order
750 * to determine whether preemption is possible (aka usually just the crit
751 * priority of lwkt_schedule() itself).
753 * XXX at the moment we run the target thread in a critical section during
754 * the preemption in order to prevent the target from taking interrupts
755 * that *WE* can't. Preemption is strictly limited to interrupt threads
756 * and interrupt-like threads, outside of a critical section, and the
757 * preempted source thread will be resumed the instant the target blocks
758 * whether or not the source is scheduled (i.e. preemption is supposed to
759 * be as transparent as possible).
761 * The target thread inherits our MP count (added to its own) for the
762 * duration of the preemption in order to preserve the atomicy of the
763 * MP lock during the preemption. Therefore, any preempting targets must be
764 * careful in regards to MP assertions. Note that the MP count may be
765 * out of sync with the physical mp_lock, but we do not have to preserve
766 * the original ownership of the lock if it was out of synch (that is, we
767 * can leave it synchronized on return).
770 lwkt_preempt(thread_t ntd, int critpri)
772 struct globaldata *gd = mycpu;
780 * The caller has put us in a critical section. We can only preempt
781 * if the caller of the caller was not in a critical section (basically
782 * a local interrupt), as determined by the 'critpri' parameter.
784 * YYY The target thread must be in a critical section (else it must
785 * inherit our critical section? I dunno yet).
787 * Any tokens held by the target may not be held by thread(s) being
788 * preempted. We take the easy way out and do not preempt if
789 * the target is holding tokens.
791 * Set need_lwkt_resched() unconditionally for now YYY.
793 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
795 td = gd->gd_curthread;
796 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
800 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
806 if (ntd->td_gd != gd) {
813 * Take the easy way out and do not preempt if the target is holding
814 * one or more tokens. We could test whether the thread(s) being
815 * preempted interlock against the target thread's tokens and whether
816 * we can get all the target thread's tokens, but this situation
817 * should not occur very often so its easier to simply not preempt.
819 if (ntd->td_toks != NULL) {
824 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
829 if (ntd->td_preempted) {
836 * note: an interrupt might have occured just as we were transitioning
837 * to or from the MP lock. In this case td_mpcount will be pre-disposed
838 * (non-zero) but not actually synchronized with the actual state of the
839 * lock. We can use it to imply an MP lock requirement for the
840 * preemption but we cannot use it to test whether we hold the MP lock
843 savecnt = td->td_mpcount;
844 mpheld = MP_LOCK_HELD();
845 ntd->td_mpcount += td->td_mpcount;
846 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
847 ntd->td_mpcount -= td->td_mpcount;
855 * Since we are able to preempt the current thread, there is no need to
856 * call need_lwkt_resched().
859 ntd->td_preempted = td;
860 td->td_flags |= TDF_PREEMPT_LOCK;
862 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
864 KKASSERT(savecnt == td->td_mpcount);
865 mpheld = MP_LOCK_HELD();
866 if (mpheld && td->td_mpcount == 0)
868 else if (mpheld == 0 && td->td_mpcount)
869 panic("lwkt_preempt(): MP lock was not held through");
871 ntd->td_preempted = NULL;
872 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
876 * Yield our thread while higher priority threads are pending. This is
877 * typically called when we leave a critical section but it can be safely
878 * called while we are in a critical section.
880 * This function will not generally yield to equal priority threads but it
881 * can occur as a side effect. Note that lwkt_switch() is called from
882 * inside the critical section to prevent its own crit_exit() from reentering
883 * lwkt_yield_quick().
885 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
886 * came along but was blocked and made pending.
888 * (self contained on a per cpu basis)
891 lwkt_yield_quick(void)
893 globaldata_t gd = mycpu;
894 thread_t td = gd->gd_curthread;
897 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
898 * it with a non-zero cpl then we might not wind up calling splz after
899 * a task switch when the critical section is exited even though the
900 * new task could accept the interrupt.
902 * XXX from crit_exit() only called after last crit section is released.
903 * If called directly will run splz() even if in a critical section.
905 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
906 * except for this special case, we MUST call splz() here to handle any
907 * pending ints, particularly after we switch, or we might accidently
908 * halt the cpu with interrupts pending.
910 if (gd->gd_reqflags && td->td_nest_count < 2)
914 * YYY enabling will cause wakeup() to task-switch, which really
915 * confused the old 4.x code. This is a good way to simulate
916 * preemption and MP without actually doing preemption or MP, because a
917 * lot of code assumes that wakeup() does not block.
919 if (untimely_switch && td->td_nest_count == 0 &&
920 gd->gd_intr_nesting_level == 0
922 crit_enter_quick(td);
924 * YYY temporary hacks until we disassociate the userland scheduler
925 * from the LWKT scheduler.
927 if (td->td_flags & TDF_RUNQ) {
928 lwkt_switch(); /* will not reenter yield function */
930 lwkt_schedule_self(td); /* make sure we are scheduled */
931 lwkt_switch(); /* will not reenter yield function */
932 lwkt_deschedule_self(td); /* make sure we are descheduled */
934 crit_exit_noyield(td);
939 * This implements a normal yield which, unlike _quick, will yield to equal
940 * priority threads as well. Note that gd_reqflags tests will be handled by
941 * the crit_exit() call in lwkt_switch().
943 * (self contained on a per cpu basis)
948 lwkt_schedule_self(curthread);
953 * Generic schedule. Possibly schedule threads belonging to other cpus and
954 * deal with threads that might be blocked on a wait queue.
956 * We have a little helper inline function which does additional work after
957 * the thread has been enqueued, including dealing with preemption and
958 * setting need_lwkt_resched() (which prevents the kernel from returning
959 * to userland until it has processed higher priority threads).
961 * It is possible for this routine to be called after a failed _enqueue
962 * (due to the target thread migrating, sleeping, or otherwise blocked).
963 * We have to check that the thread is actually on the run queue!
967 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
969 if (ntd->td_flags & TDF_RUNQ) {
970 if (ntd->td_preemptable) {
971 ntd->td_preemptable(ntd, cpri); /* YYY +token */
972 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
973 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
981 lwkt_schedule(thread_t td)
983 globaldata_t mygd = mycpu;
985 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
987 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
988 if (td == mygd->gd_curthread) {
994 * If the thread is on a wait list we have to send our scheduling
995 * request to the owner of the wait structure. Otherwise we send
996 * the scheduling request to the cpu owning the thread. Races
997 * are ok, the target will forward the message as necessary (the
998 * message may chase the thread around before it finally gets
1001 * (remember, wait structures use stable storage)
1003 * NOTE: we have to account for the number of critical sections
1004 * under our control when calling _lwkt_schedule_post() so it
1005 * can figure out whether preemption is allowed.
1007 * NOTE: The wait structure algorithms are a mess and need to be
1010 * NOTE: We cannot safely acquire or release a token, even
1011 * non-blocking, because this routine may be called in the context
1012 * of a thread already holding the token and thus not provide any
1013 * interlock protection. We cannot safely manipulate the td_toks
1014 * list for the same reason. Instead we depend on our critical
1015 * section if the token is owned by our cpu.
1017 if ((w = td->td_wait) != NULL) {
1018 if (w->wa_token.t_cpu == mygd) {
1019 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1023 if (td->td_gd == mygd) {
1025 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1027 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1031 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1035 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc1_t)lwkt_schedule, td);
1037 panic("bad token %p", &w->wa_token);
1042 * If the wait structure is NULL and we own the thread, there
1043 * is no race (since we are in a critical section). If we
1044 * do not own the thread there might be a race but the
1045 * target cpu will deal with it.
1048 if (td->td_gd == mygd) {
1050 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1052 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1056 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1064 * Managed acquisition. This code assumes that the MP lock is held for
1065 * the tdallq operation and that the thread has been descheduled from its
1066 * original cpu. We also have to wait for the thread to be entirely switched
1067 * out on its original cpu (this is usually fast enough that we never loop)
1068 * since the LWKT system does not have to hold the MP lock while switching
1069 * and the target may have released it before switching.
1072 lwkt_acquire(thread_t td)
1080 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1081 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */
1084 crit_enter_gd(mygd);
1085 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
1087 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1093 * Generic deschedule. Descheduling threads other then your own should be
1094 * done only in carefully controlled circumstances. Descheduling is
1097 * This function may block if the cpu has run out of messages.
1100 lwkt_deschedule(thread_t td)
1104 if (td == curthread) {
1107 if (td->td_gd == mycpu) {
1110 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1120 * Set the target thread's priority. This routine does not automatically
1121 * switch to a higher priority thread, LWKT threads are not designed for
1122 * continuous priority changes. Yield if you want to switch.
1124 * We have to retain the critical section count which uses the high bits
1125 * of the td_pri field. The specified priority may also indicate zero or
1126 * more critical sections by adding TDPRI_CRIT*N.
1128 * Note that we requeue the thread whether it winds up on a different runq
1129 * or not. uio_yield() depends on this and the routine is not normally
1130 * called with the same priority otherwise.
1133 lwkt_setpri(thread_t td, int pri)
1136 KKASSERT(td->td_gd == mycpu);
1138 if (td->td_flags & TDF_RUNQ) {
1140 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1143 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1149 lwkt_setpri_self(int pri)
1151 thread_t td = curthread;
1153 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1155 if (td->td_flags & TDF_RUNQ) {
1157 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1160 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1166 * Determine if there is a runnable thread at a higher priority then
1167 * the current thread. lwkt_setpri() does not check this automatically.
1168 * Return 1 if there is, 0 if there isn't.
1170 * Example: if bit 31 of runqmask is set and the current thread is priority
1171 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1173 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1174 * up comparing against 0xffffffff, a comparison that will always be false.
1177 lwkt_checkpri_self(void)
1179 globaldata_t gd = mycpu;
1180 thread_t td = gd->gd_curthread;
1181 int nq = td->td_pri & TDPRI_MASK;
1183 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1184 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1192 * Migrate the current thread to the specified cpu. The BGL must be held
1193 * (for the gd_tdallq manipulation XXX). This is accomplished by
1194 * descheduling ourselves from the current cpu, moving our thread to the
1195 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1196 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1199 static void lwkt_setcpu_remote(void *arg);
1203 lwkt_setcpu_self(globaldata_t rgd)
1206 thread_t td = curthread;
1208 if (td->td_gd != rgd) {
1209 crit_enter_quick(td);
1210 td->td_flags |= TDF_MIGRATING;
1211 lwkt_deschedule_self(td);
1212 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1213 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1214 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1216 /* we are now on the target cpu */
1217 crit_exit_quick(td);
1223 lwkt_migratecpu(int cpuid)
1228 rgd = globaldata_find(cpuid);
1229 lwkt_setcpu_self(rgd);
1234 * Remote IPI for cpu migration (called while in a critical section so we
1235 * do not have to enter another one). The thread has already been moved to
1236 * our cpu's allq, but we must wait for the thread to be completely switched
1237 * out on the originating cpu before we schedule it on ours or the stack
1238 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1239 * change to main memory.
1241 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1242 * against wakeups. It is best if this interface is used only when there
1243 * are no pending events that might try to schedule the thread.
1247 lwkt_setcpu_remote(void *arg)
1250 globaldata_t gd = mycpu;
1252 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1256 td->td_flags &= ~TDF_MIGRATING;
1257 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
1263 lwkt_preempted_proc(void)
1265 thread_t td = curthread;
1266 while (td->td_preempted)
1267 td = td->td_preempted;
1272 * Block on the specified wait queue until signaled. A generation number
1273 * must be supplied to interlock the wait queue. The function will
1274 * return immediately if the generation number does not match the wait
1275 * structure's generation number.
1278 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1280 thread_t td = curthread;
1283 lwkt_gettoken(&ilock, &w->wa_token);
1285 if (w->wa_gen == *gen) {
1287 td->td_flags |= TDF_BLOCKQ;
1288 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1291 td->td_wmesg = wmesg;
1293 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1294 td->td_wmesg = NULL;
1298 lwkt_reltoken(&ilock);
1302 * Signal a wait queue. We gain ownership of the wait queue in order to
1303 * signal it. Once a thread is removed from the wait queue we have to
1304 * deal with the cpu owning the thread.
1306 * Note: alternatively we could message the target cpu owning the wait
1307 * queue. YYY implement as sysctl.
1310 lwkt_signal(lwkt_wait_t w, int count)
1315 lwkt_gettoken(&ilock, &w->wa_token);
1319 count = w->wa_count;
1320 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1323 KKASSERT(td->td_flags & TDF_BLOCKQ);
1324 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1325 td->td_flags &= ~TDF_BLOCKQ;
1327 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
1329 if (td->td_gd == mycpu) {
1332 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1339 lwkt_reltoken(&ilock);
1343 * Create a kernel process/thread/whatever. It shares it's address space
1344 * with proc0 - ie: kernel only.
1346 * NOTE! By default new threads are created with the MP lock held. A
1347 * thread which does not require the MP lock should release it by calling
1348 * rel_mplock() at the start of the new thread.
1351 lwkt_create(void (*func)(void *), void *arg,
1352 struct thread **tdp, thread_t template, int tdflags, int cpu,
1353 const char *fmt, ...)
1358 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1359 tdflags | TDF_VERBOSE);
1362 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1365 * Set up arg0 for 'ps' etc
1367 __va_start(ap, fmt);
1368 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1372 * Schedule the thread to run
1374 if ((td->td_flags & TDF_STOPREQ) == 0)
1377 td->td_flags &= ~TDF_STOPREQ;
1382 * kthread_* is specific to the kernel and is not needed by userland.
1387 * Destroy an LWKT thread. Warning! This function is not called when
1388 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1389 * uses a different reaping mechanism.
1394 thread_t td = curthread;
1397 if (td->td_flags & TDF_VERBOSE)
1398 printf("kthread %p %s has exited\n", td, td->td_comm);
1400 crit_enter_quick(td);
1401 lwkt_deschedule_self(td);
1403 KKASSERT(gd == td->td_gd);
1404 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1405 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1406 ++gd->gd_tdfreecount;
1407 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1412 #endif /* _KERNEL */
1417 thread_t td = curthread;
1418 int lpri = td->td_pri;
1421 panic("td_pri is/would-go negative! %p %d", td, lpri);
1427 * Called from debugger/panic on cpus which have been stopped. We must still
1428 * process the IPIQ while stopped, even if we were stopped while in a critical
1431 * If we are dumping also try to process any pending interrupts. This may
1432 * or may not work depending on the state of the cpu at the point it was
1436 lwkt_smp_stopped(void)
1438 globaldata_t gd = mycpu;
1442 lwkt_process_ipiq();
1445 lwkt_process_ipiq();