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 1.96 2006/05/21 20:23:25 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>
59 #include <sys/thread2.h>
60 #include <sys/spinlock2.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_zone.h>
72 #include <machine/stdarg.h>
73 #include <machine/ipl.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;
109 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
111 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
118 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
119 &token_contention_count, 0, "spinning due to token contention");
120 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
121 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
126 * These helper procedures handle the runq, they can only be called from
127 * within a critical section.
129 * WARNING! Prior to SMP being brought up it is possible to enqueue and
130 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
131 * instead of 'mycpu' when referencing the globaldata structure. Once
132 * SMP live enqueuing and dequeueing only occurs on the current cpu.
136 _lwkt_dequeue(thread_t td)
138 if (td->td_flags & TDF_RUNQ) {
139 int nq = td->td_pri & TDPRI_MASK;
140 struct globaldata *gd = td->td_gd;
142 td->td_flags &= ~TDF_RUNQ;
143 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
144 /* runqmask is passively cleaned up by the switcher */
150 _lwkt_enqueue(thread_t td)
152 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
153 int nq = td->td_pri & TDPRI_MASK;
154 struct globaldata *gd = td->td_gd;
156 td->td_flags |= TDF_RUNQ;
157 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
158 gd->gd_runqmask |= 1 << nq;
163 * Schedule a thread to run. As the current thread we can always safely
164 * schedule ourselves, and a shortcut procedure is provided for that
167 * (non-blocking, self contained on a per cpu basis)
170 lwkt_schedule_self(thread_t td)
172 crit_enter_quick(td);
173 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
174 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
175 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
181 * Deschedule a thread.
183 * (non-blocking, self contained on a per cpu basis)
186 lwkt_deschedule_self(thread_t td)
188 crit_enter_quick(td);
189 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
197 * LWKTs operate on a per-cpu basis
199 * WARNING! Called from early boot, 'mycpu' may not work yet.
202 lwkt_gdinit(struct globaldata *gd)
206 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
207 TAILQ_INIT(&gd->gd_tdrunq[i]);
209 TAILQ_INIT(&gd->gd_tdallq);
215 * Initialize a thread wait structure prior to first use.
217 * NOTE! called from low level boot code, we cannot do anything fancy!
220 lwkt_wait_init(lwkt_wait_t w)
222 spin_init(&w->wa_spinlock);
223 TAILQ_INIT(&w->wa_waitq);
229 * Create a new thread. The thread must be associated with a process context
230 * or LWKT start address before it can be scheduled. If the target cpu is
231 * -1 the thread will be created on the current cpu.
233 * If you intend to create a thread without a process context this function
234 * does everything except load the startup and switcher function.
237 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
240 globaldata_t gd = mycpu;
244 if (gd->gd_tdfreecount > 0) {
245 --gd->gd_tdfreecount;
246 td = TAILQ_FIRST(&gd->gd_tdfreeq);
247 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
248 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
249 TAILQ_REMOVE(&gd->gd_tdfreeq, td, td_threadq);
251 flags |= td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
255 td = zalloc(thread_zone);
257 td = malloc(sizeof(struct thread));
259 td->td_kstack = NULL;
260 td->td_kstack_size = 0;
261 flags |= TDF_ALLOCATED_THREAD;
264 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
265 if (flags & TDF_ALLOCATED_STACK) {
267 kmem_free(kernel_map, (vm_offset_t)stack, td->td_kstack_size);
269 libcaps_free_stack(stack, td->td_kstack_size);
276 stack = (void *)kmem_alloc(kernel_map, stksize);
278 stack = libcaps_alloc_stack(stksize);
280 flags |= TDF_ALLOCATED_STACK;
283 lwkt_init_thread(td, stack, stksize, flags, mycpu);
285 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
292 * Initialize a preexisting thread structure. This function is used by
293 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
295 * All threads start out in a critical section at a priority of
296 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
297 * appropriate. This function may send an IPI message when the
298 * requested cpu is not the current cpu and consequently gd_tdallq may
299 * not be initialized synchronously from the point of view of the originating
302 * NOTE! we have to be careful in regards to creating threads for other cpus
303 * if SMP has not yet been activated.
308 lwkt_init_thread_remote(void *arg)
312 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
318 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
319 struct globaldata *gd)
321 globaldata_t mygd = mycpu;
323 bzero(td, sizeof(struct thread));
324 td->td_kstack = stack;
325 td->td_kstack_size = stksize;
326 td->td_flags = flags;
328 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
330 if ((flags & TDF_MPSAFE) == 0)
333 lwkt_initport(&td->td_msgport, td);
334 pmap_init_thread(td);
337 * Normally initializing a thread for a remote cpu requires sending an
338 * IPI. However, the idlethread is setup before the other cpus are
339 * activated so we have to treat it as a special case. XXX manipulation
340 * of gd_tdallq requires the BGL.
342 if (gd == mygd || td == &gd->gd_idlethread) {
344 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
347 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
351 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
359 lwkt_set_comm(thread_t td, const char *ctl, ...)
364 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
369 lwkt_hold(thread_t td)
375 lwkt_rele(thread_t td)
377 KKASSERT(td->td_refs > 0);
384 lwkt_wait_free(thread_t td)
387 tsleep(td, 0, "tdreap", hz);
393 lwkt_free_thread(thread_t td)
395 struct globaldata *gd = mycpu;
397 KASSERT((td->td_flags & TDF_RUNNING) == 0,
398 ("lwkt_free_thread: did not exit! %p", td));
401 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* Protected by BGL */
402 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
403 (td->td_flags & TDF_ALLOCATED_THREAD)
405 ++gd->gd_tdfreecount;
406 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
410 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
412 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
414 libcaps_free_stack(td->td_kstack, td->td_kstack_size);
417 td->td_kstack = NULL;
418 td->td_kstack_size = 0;
420 if (td->td_flags & TDF_ALLOCATED_THREAD) {
422 zfree(thread_zone, td);
432 * Switch to the next runnable lwkt. If no LWKTs are runnable then
433 * switch to the idlethread. Switching must occur within a critical
434 * section to avoid races with the scheduling queue.
436 * We always have full control over our cpu's run queue. Other cpus
437 * that wish to manipulate our queue must use the cpu_*msg() calls to
438 * talk to our cpu, so a critical section is all that is needed and
439 * the result is very, very fast thread switching.
441 * The LWKT scheduler uses a fixed priority model and round-robins at
442 * each priority level. User process scheduling is a totally
443 * different beast and LWKT priorities should not be confused with
444 * user process priorities.
446 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
447 * cleans it up. Note that the td_switch() function cannot do anything that
448 * requires the MP lock since the MP lock will have already been setup for
449 * the target thread (not the current thread). It's nice to have a scheduler
450 * that does not need the MP lock to work because it allows us to do some
451 * really cool high-performance MP lock optimizations.
453 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
454 * is not called by the current thread in the preemption case, only when
455 * the preempting thread blocks (in order to return to the original thread).
460 globaldata_t gd = mycpu;
461 thread_t td = gd->gd_curthread;
468 * Switching from within a 'fast' (non thread switched) interrupt or IPI
469 * is illegal. However, we may have to do it anyway if we hit a fatal
470 * kernel trap or we have paniced.
472 * If this case occurs save and restore the interrupt nesting level.
474 if (gd->gd_intr_nesting_level) {
478 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
479 panic("lwkt_switch: cannot switch from within "
480 "a fast interrupt, yet, td %p\n", td);
482 savegdnest = gd->gd_intr_nesting_level;
483 savegdtrap = gd->gd_trap_nesting_level;
484 gd->gd_intr_nesting_level = 0;
485 gd->gd_trap_nesting_level = 0;
486 if ((td->td_flags & TDF_PANICWARN) == 0) {
487 td->td_flags |= TDF_PANICWARN;
488 printf("Warning: thread switch from interrupt or IPI, "
489 "thread %p (%s)\n", td, td->td_comm);
491 db_print_backtrace();
495 gd->gd_intr_nesting_level = savegdnest;
496 gd->gd_trap_nesting_level = savegdtrap;
502 * Passive release (used to transition from user to kernel mode
503 * when we block or switch rather then when we enter the kernel).
504 * This function is NOT called if we are switching into a preemption
505 * or returning from a preemption. Typically this causes us to lose
506 * our current process designation (if we have one) and become a true
507 * LWKT thread, and may also hand the current process designation to
508 * another process and schedule thread.
516 lwkt_relalltokens(td);
520 * We had better not be holding any spin locks, but don't get into an
521 * endless panic loop.
523 KASSERT(gd->gd_spinlocks_rd == 0 || panicstr != NULL,
524 ("lwkt_switch: still holding %d shared spinlocks!",
525 gd->gd_spinlocks_rd));
526 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
527 ("lwkt_switch: still holding %d exclusive spinlocks!",
528 gd->gd_spinlocks_wr));
533 * td_mpcount cannot be used to determine if we currently hold the
534 * MP lock because get_mplock() will increment it prior to attempting
535 * to get the lock, and switch out if it can't. Our ownership of
536 * the actual lock will remain stable while we are in a critical section
537 * (but, of course, another cpu may own or release the lock so the
538 * actual value of mp_lock is not stable).
540 mpheld = MP_LOCK_HELD();
542 if (td->td_cscount) {
543 printf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
545 if (panic_on_cscount)
546 panic("switching while mastering cpusync");
550 if ((ntd = td->td_preempted) != NULL) {
552 * We had preempted another thread on this cpu, resume the preempted
553 * thread. This occurs transparently, whether the preempted thread
554 * was scheduled or not (it may have been preempted after descheduling
557 * We have to setup the MP lock for the original thread after backing
558 * out the adjustment that was made to curthread when the original
561 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
563 if (ntd->td_mpcount && mpheld == 0) {
564 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
565 td, ntd, td->td_mpcount, ntd->td_mpcount);
567 if (ntd->td_mpcount) {
568 td->td_mpcount -= ntd->td_mpcount;
569 KKASSERT(td->td_mpcount >= 0);
572 ntd->td_flags |= TDF_PREEMPT_DONE;
575 * XXX. The interrupt may have woken a thread up, we need to properly
576 * set the reschedule flag if the originally interrupted thread is at
579 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
581 /* YYY release mp lock on switchback if original doesn't need it */
584 * Priority queue / round-robin at each priority. Note that user
585 * processes run at a fixed, low priority and the user process
586 * scheduler deals with interactions between user processes
587 * by scheduling and descheduling them from the LWKT queue as
590 * We have to adjust the MP lock for the target thread. If we
591 * need the MP lock and cannot obtain it we try to locate a
592 * thread that does not need the MP lock. If we cannot, we spin
595 * A similar issue exists for the tokens held by the target thread.
596 * If we cannot obtain ownership of the tokens we cannot immediately
597 * schedule the thread.
601 * If an LWKT reschedule was requested, well that is what we are
602 * doing now so clear it.
604 clear_lwkt_resched();
606 if (gd->gd_runqmask) {
607 int nq = bsrl(gd->gd_runqmask);
608 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
609 gd->gd_runqmask &= ~(1 << nq);
614 * THREAD SELECTION FOR AN SMP MACHINE BUILD
616 * If the target needs the MP lock and we couldn't get it,
617 * or if the target is holding tokens and we could not
618 * gain ownership of the tokens, continue looking for a
619 * thread to schedule and spin instead of HLT if we can't.
621 * NOTE: the mpheld variable invalid after this conditional, it
622 * can change due to both cpu_try_mplock() returning success
623 * AND interactions in lwkt_getalltokens() due to the fact that
624 * we are trying to check the mpcount of a thread other then
625 * the current thread. Because of this, if the current thread
626 * is not holding td_mpcount, an IPI indirectly run via
627 * lwkt_getalltokens() can obtain and release the MP lock and
628 * cause the core MP lock to be released.
630 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
631 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
633 u_int32_t rqmask = gd->gd_runqmask;
635 mpheld = MP_LOCK_HELD();
638 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
639 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
640 /* spinning due to MP lock being held */
642 ++mplock_contention_count;
644 /* mplock still not held, 'mpheld' still valid */
649 * mpheld state invalid after getalltokens call returns
650 * failure, but the variable is only needed for
653 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
654 /* spinning due to token contention */
656 ++token_contention_count;
658 mpheld = MP_LOCK_HELD();
665 rqmask &= ~(1 << nq);
669 ntd = &gd->gd_idlethread;
670 ntd->td_flags |= TDF_IDLE_NOHLT;
671 goto using_idle_thread;
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);
678 ++gd->gd_cnt.v_swtch;
679 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
680 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
684 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
685 * worry about tokens or the BGL.
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);
693 * We have nothing to run but only let the idle loop halt
694 * the cpu if there are no pending interrupts.
696 ntd = &gd->gd_idlethread;
697 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
698 ntd->td_flags |= TDF_IDLE_NOHLT;
702 * The idle thread should not be holding the MP lock unless we
703 * are trapping in the kernel or in a panic. Since we select the
704 * idle thread unconditionally when no other thread is available,
705 * if the MP lock is desired during a panic or kernel trap, we
706 * have to loop in the scheduler until we get it.
708 if (ntd->td_mpcount) {
709 mpheld = MP_LOCK_HELD();
710 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
711 panic("Idle thread %p was holding the BGL!", ntd);
712 else if (mpheld == 0)
718 KASSERT(ntd->td_pri >= TDPRI_CRIT,
719 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
722 * Do the actual switch. If the new target does not need the MP lock
723 * and we are holding it, release the MP lock. If the new target requires
724 * the MP lock we have already acquired it for the target.
727 if (ntd->td_mpcount == 0 ) {
731 ASSERT_MP_LOCK_HELD(ntd);
738 /* NOTE: current cpu may have changed after switch */
743 * Request that the target thread preempt the current thread. Preemption
744 * only works under a specific set of conditions:
746 * - We are not preempting ourselves
747 * - The target thread is owned by the current cpu
748 * - We are not currently being preempted
749 * - The target is not currently being preempted
750 * - We are able to satisfy the target's MP lock requirements (if any).
752 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
753 * this is called via lwkt_schedule() through the td_preemptable callback.
754 * critpri is the managed critical priority that we should ignore in order
755 * to determine whether preemption is possible (aka usually just the crit
756 * priority of lwkt_schedule() itself).
758 * XXX at the moment we run the target thread in a critical section during
759 * the preemption in order to prevent the target from taking interrupts
760 * that *WE* can't. Preemption is strictly limited to interrupt threads
761 * and interrupt-like threads, outside of a critical section, and the
762 * preempted source thread will be resumed the instant the target blocks
763 * whether or not the source is scheduled (i.e. preemption is supposed to
764 * be as transparent as possible).
766 * The target thread inherits our MP count (added to its own) for the
767 * duration of the preemption in order to preserve the atomicy of the
768 * MP lock during the preemption. Therefore, any preempting targets must be
769 * careful in regards to MP assertions. Note that the MP count may be
770 * out of sync with the physical mp_lock, but we do not have to preserve
771 * the original ownership of the lock if it was out of synch (that is, we
772 * can leave it synchronized on return).
775 lwkt_preempt(thread_t ntd, int critpri)
777 struct globaldata *gd = mycpu;
785 * The caller has put us in a critical section. We can only preempt
786 * if the caller of the caller was not in a critical section (basically
787 * a local interrupt), as determined by the 'critpri' parameter. We
788 * also acn't preempt if the caller is holding any spinlocks (even if
789 * he isn't in a critical section). This also handles the tokens test.
791 * YYY The target thread must be in a critical section (else it must
792 * inherit our critical section? I dunno yet).
794 * Set need_lwkt_resched() unconditionally for now YYY.
796 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
798 td = gd->gd_curthread;
799 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
803 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
809 if (ntd->td_gd != gd) {
816 * Take the easy way out and do not preempt if the target is holding
817 * any spinlocks. We could test whether the thread(s) being
818 * preempted interlock against the target thread's tokens and whether
819 * we can get all the target thread's tokens, but this situation
820 * should not occur very often so its easier to simply not preempt.
821 * Also, plain spinlocks are impossible to figure out at this point so
822 * just don't preempt.
824 if (gd->gd_spinlocks_rd + gd->gd_spinlocks_wr != 0) {
829 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
834 if (ntd->td_preempted) {
841 * note: an interrupt might have occured just as we were transitioning
842 * to or from the MP lock. In this case td_mpcount will be pre-disposed
843 * (non-zero) but not actually synchronized with the actual state of the
844 * lock. We can use it to imply an MP lock requirement for the
845 * preemption but we cannot use it to test whether we hold the MP lock
848 savecnt = td->td_mpcount;
849 mpheld = MP_LOCK_HELD();
850 ntd->td_mpcount += td->td_mpcount;
851 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
852 ntd->td_mpcount -= td->td_mpcount;
860 * Since we are able to preempt the current thread, there is no need to
861 * call need_lwkt_resched().
864 ntd->td_preempted = td;
865 td->td_flags |= TDF_PREEMPT_LOCK;
867 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
869 KKASSERT(savecnt == td->td_mpcount);
870 mpheld = MP_LOCK_HELD();
871 if (mpheld && td->td_mpcount == 0)
873 else if (mpheld == 0 && td->td_mpcount)
874 panic("lwkt_preempt(): MP lock was not held through");
876 ntd->td_preempted = NULL;
877 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
881 * Yield our thread while higher priority threads are pending. This is
882 * typically called when we leave a critical section but it can be safely
883 * called while we are in a critical section.
885 * This function will not generally yield to equal priority threads but it
886 * can occur as a side effect. Note that lwkt_switch() is called from
887 * inside the critical section to prevent its own crit_exit() from reentering
888 * lwkt_yield_quick().
890 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
891 * came along but was blocked and made pending.
893 * (self contained on a per cpu basis)
896 lwkt_yield_quick(void)
898 globaldata_t gd = mycpu;
899 thread_t td = gd->gd_curthread;
902 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
903 * it with a non-zero cpl then we might not wind up calling splz after
904 * a task switch when the critical section is exited even though the
905 * new task could accept the interrupt.
907 * XXX from crit_exit() only called after last crit section is released.
908 * If called directly will run splz() even if in a critical section.
910 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
911 * except for this special case, we MUST call splz() here to handle any
912 * pending ints, particularly after we switch, or we might accidently
913 * halt the cpu with interrupts pending.
915 if (gd->gd_reqflags && td->td_nest_count < 2)
919 * YYY enabling will cause wakeup() to task-switch, which really
920 * confused the old 4.x code. This is a good way to simulate
921 * preemption and MP without actually doing preemption or MP, because a
922 * lot of code assumes that wakeup() does not block.
924 if (untimely_switch && td->td_nest_count == 0 &&
925 gd->gd_intr_nesting_level == 0
927 crit_enter_quick(td);
929 * YYY temporary hacks until we disassociate the userland scheduler
930 * from the LWKT scheduler.
932 if (td->td_flags & TDF_RUNQ) {
933 lwkt_switch(); /* will not reenter yield function */
935 lwkt_schedule_self(td); /* make sure we are scheduled */
936 lwkt_switch(); /* will not reenter yield function */
937 lwkt_deschedule_self(td); /* make sure we are descheduled */
939 crit_exit_noyield(td);
944 * This implements a normal yield which, unlike _quick, will yield to equal
945 * priority threads as well. Note that gd_reqflags tests will be handled by
946 * the crit_exit() call in lwkt_switch().
948 * (self contained on a per cpu basis)
953 lwkt_schedule_self(curthread);
958 * Generic schedule. Possibly schedule threads belonging to other cpus and
959 * deal with threads that might be blocked on a wait queue.
961 * We have a little helper inline function which does additional work after
962 * the thread has been enqueued, including dealing with preemption and
963 * setting need_lwkt_resched() (which prevents the kernel from returning
964 * to userland until it has processed higher priority threads).
966 * It is possible for this routine to be called after a failed _enqueue
967 * (due to the target thread migrating, sleeping, or otherwise blocked).
968 * We have to check that the thread is actually on the run queue!
972 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
974 if (ntd->td_flags & TDF_RUNQ) {
975 if (ntd->td_preemptable) {
976 ntd->td_preemptable(ntd, cpri); /* YYY +token */
977 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
978 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
986 lwkt_schedule(thread_t td)
988 globaldata_t mygd = mycpu;
990 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
992 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
993 if (td == mygd->gd_curthread) {
999 * If the thread is on a wait list we have to send our scheduling
1000 * request to the owner of the wait structure. Otherwise we send
1001 * the scheduling request to the cpu owning the thread. Races
1002 * are ok, the target will forward the message as necessary (the
1003 * message may chase the thread around before it finally gets
1006 * (remember, wait structures use stable storage)
1008 * NOTE: we have to account for the number of critical sections
1009 * under our control when calling _lwkt_schedule_post() so it
1010 * can figure out whether preemption is allowed.
1012 * NOTE: The wait structure algorithms are a mess and need to be
1015 * NOTE: We cannot safely acquire or release a token, even
1016 * non-blocking, because this routine may be called in the context
1017 * of a thread already holding the token and thus not provide any
1018 * interlock protection. We cannot safely manipulate the td_toks
1019 * list for the same reason. Instead we depend on our critical
1020 * section if the token is owned by our cpu.
1022 if ((w = td->td_wait) != NULL) {
1023 spin_lock_wr(&w->wa_spinlock);
1024 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1027 spin_unlock_wr(&w->wa_spinlock);
1029 if (td->td_gd == mygd) {
1031 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1033 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1037 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1041 * If the wait structure is NULL and we own the thread, there
1042 * is no race (since we are in a critical section). If we
1043 * do not own the thread there might be a race but the
1044 * target cpu will deal with it.
1047 if (td->td_gd == mygd) {
1049 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1051 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1055 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1063 * Managed acquisition. This code assumes that the MP lock is held for
1064 * the tdallq operation and that the thread has been descheduled from its
1065 * original cpu. We also have to wait for the thread to be entirely switched
1066 * out on its original cpu (this is usually fast enough that we never loop)
1067 * since the LWKT system does not have to hold the MP lock while switching
1068 * and the target may have released it before switching.
1071 lwkt_acquire(thread_t td)
1079 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1080 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) /* XXX spin */
1083 crit_enter_gd(mygd);
1084 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
1086 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq); /* protected by BGL */
1092 * Generic deschedule. Descheduling threads other then your own should be
1093 * done only in carefully controlled circumstances. Descheduling is
1096 * This function may block if the cpu has run out of messages.
1099 lwkt_deschedule(thread_t td)
1103 if (td == curthread) {
1106 if (td->td_gd == mycpu) {
1109 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1119 * Set the target thread's priority. This routine does not automatically
1120 * switch to a higher priority thread, LWKT threads are not designed for
1121 * continuous priority changes. Yield if you want to switch.
1123 * We have to retain the critical section count which uses the high bits
1124 * of the td_pri field. The specified priority may also indicate zero or
1125 * more critical sections by adding TDPRI_CRIT*N.
1127 * Note that we requeue the thread whether it winds up on a different runq
1128 * or not. uio_yield() depends on this and the routine is not normally
1129 * called with the same priority otherwise.
1132 lwkt_setpri(thread_t td, int pri)
1135 KKASSERT(td->td_gd == mycpu);
1137 if (td->td_flags & TDF_RUNQ) {
1139 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1142 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1148 lwkt_setpri_self(int pri)
1150 thread_t td = curthread;
1152 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1154 if (td->td_flags & TDF_RUNQ) {
1156 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1159 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1165 * Determine if there is a runnable thread at a higher priority then
1166 * the current thread. lwkt_setpri() does not check this automatically.
1167 * Return 1 if there is, 0 if there isn't.
1169 * Example: if bit 31 of runqmask is set and the current thread is priority
1170 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1172 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1173 * up comparing against 0xffffffff, a comparison that will always be false.
1176 lwkt_checkpri_self(void)
1178 globaldata_t gd = mycpu;
1179 thread_t td = gd->gd_curthread;
1180 int nq = td->td_pri & TDPRI_MASK;
1182 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1183 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1191 * Migrate the current thread to the specified cpu. The BGL must be held
1192 * (for the gd_tdallq manipulation XXX). This is accomplished by
1193 * descheduling ourselves from the current cpu, moving our thread to the
1194 * tdallq of the target cpu, IPI messaging the target cpu, and switching out.
1195 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1198 static void lwkt_setcpu_remote(void *arg);
1202 lwkt_setcpu_self(globaldata_t rgd)
1205 thread_t td = curthread;
1207 if (td->td_gd != rgd) {
1208 crit_enter_quick(td);
1209 td->td_flags |= TDF_MIGRATING;
1210 lwkt_deschedule_self(td);
1211 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq); /* protected by BGL */
1212 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq); /* protected by BGL */
1213 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1215 /* we are now on the target cpu */
1216 crit_exit_quick(td);
1222 lwkt_migratecpu(int cpuid)
1227 rgd = globaldata_find(cpuid);
1228 lwkt_setcpu_self(rgd);
1233 * Remote IPI for cpu migration (called while in a critical section so we
1234 * do not have to enter another one). The thread has already been moved to
1235 * our cpu's allq, but we must wait for the thread to be completely switched
1236 * out on the originating cpu before we schedule it on ours or the stack
1237 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1238 * change to main memory.
1240 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1241 * against wakeups. It is best if this interface is used only when there
1242 * are no pending events that might try to schedule the thread.
1246 lwkt_setcpu_remote(void *arg)
1249 globaldata_t gd = mycpu;
1251 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1255 td->td_flags &= ~TDF_MIGRATING;
1256 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
1262 lwkt_preempted_proc(void)
1264 thread_t td = curthread;
1265 while (td->td_preempted)
1266 td = td->td_preempted;
1271 * Block on the specified wait queue until signaled. A generation number
1272 * must be supplied to interlock the wait queue. The function will
1273 * return immediately if the generation number does not match the wait
1274 * structure's generation number.
1277 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
1279 thread_t td = curthread;
1281 spin_lock_wr(&w->wa_spinlock);
1282 if (w->wa_gen == *gen) {
1284 td->td_flags |= TDF_BLOCKQ;
1285 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
1288 td->td_wmesg = wmesg;
1289 spin_unlock_wr(&w->wa_spinlock);
1291 KKASSERT((td->td_flags & TDF_BLOCKQ) == 0);
1292 td->td_wmesg = NULL;
1296 spin_unlock_wr(&w->wa_spinlock);
1301 * Signal a wait queue. We gain ownership of the wait queue in order to
1302 * signal it. Once a thread is removed from the wait queue we have to
1303 * deal with the cpu owning the thread.
1305 * Note: alternatively we could message the target cpu owning the wait
1306 * queue. YYY implement as sysctl.
1309 lwkt_signal(lwkt_wait_t w, int count)
1313 spin_lock_wr(&w->wa_spinlock);
1316 count = w->wa_count;
1317 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1320 KKASSERT(td->td_flags & TDF_BLOCKQ);
1321 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1322 td->td_flags &= ~TDF_BLOCKQ;
1324 spin_unlock_wr(&w->wa_spinlock);
1325 KKASSERT(td->td_proc == NULL || (td->td_proc->p_flag & P_ONRUNQ) == 0);
1327 if (td->td_gd == mycpu) {
1330 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
1335 spin_lock_wr(&w->wa_spinlock);
1337 spin_unlock_wr(&w->wa_spinlock);
1341 * Create a kernel process/thread/whatever. It shares it's address space
1342 * with proc0 - ie: kernel only.
1344 * NOTE! By default new threads are created with the MP lock held. A
1345 * thread which does not require the MP lock should release it by calling
1346 * rel_mplock() at the start of the new thread.
1349 lwkt_create(void (*func)(void *), void *arg,
1350 struct thread **tdp, thread_t template, int tdflags, int cpu,
1351 const char *fmt, ...)
1356 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1357 tdflags | TDF_VERBOSE);
1360 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1363 * Set up arg0 for 'ps' etc
1365 __va_start(ap, fmt);
1366 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1370 * Schedule the thread to run
1372 if ((td->td_flags & TDF_STOPREQ) == 0)
1375 td->td_flags &= ~TDF_STOPREQ;
1380 * kthread_* is specific to the kernel and is not needed by userland.
1385 * Destroy an LWKT thread. Warning! This function is not called when
1386 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1387 * uses a different reaping mechanism.
1392 thread_t td = curthread;
1395 if (td->td_flags & TDF_VERBOSE)
1396 printf("kthread %p %s has exited\n", td, td->td_comm);
1398 crit_enter_quick(td);
1399 lwkt_deschedule_self(td);
1401 KKASSERT(gd == td->td_gd);
1402 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1403 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1404 ++gd->gd_tdfreecount;
1405 TAILQ_INSERT_TAIL(&gd->gd_tdfreeq, td, td_threadq);
1410 #endif /* _KERNEL */
1415 thread_t td = curthread;
1416 int lpri = td->td_pri;
1419 panic("td_pri is/would-go negative! %p %d", td, lpri);
1425 * Called from debugger/panic on cpus which have been stopped. We must still
1426 * process the IPIQ while stopped, even if we were stopped while in a critical
1429 * If we are dumping also try to process any pending interrupts. This may
1430 * or may not work depending on the state of the cpu at the point it was
1434 lwkt_smp_stopped(void)
1436 globaldata_t gd = mycpu;
1440 lwkt_process_ipiq();
1443 lwkt_process_ipiq();