2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.47 2003/12/30 03:19:02 dillon Exp $
30 * Each cpu in a system has its own self-contained light weight kernel
31 * thread scheduler, which means that generally speaking we only need
32 * to use a critical section to avoid problems. Foreign thread
33 * scheduling is queued via (async) IPIs.
35 * NOTE: on UP machines smp_active is defined to be 0. On SMP machines
36 * smp_active is 0 prior to SMP activation, then it is 1. The LWKT module
37 * uses smp_active to optimize UP builds and to avoid sending IPIs during
38 * early boot (primarily interrupt and network thread initialization).
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
47 #include <sys/rtprio.h>
48 #include <sys/queue.h>
49 #include <sys/thread2.h>
50 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
56 #include <vm/vm_param.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_pager.h>
62 #include <vm/vm_extern.h>
63 #include <vm/vm_zone.h>
65 #include <machine/stdarg.h>
66 #include <machine/ipl.h>
67 #include <machine/smp.h>
69 #define THREAD_STACK (UPAGES * PAGE_SIZE)
73 #include <sys/stdint.h>
74 #include <libcaps/thread.h>
75 #include <sys/thread.h>
76 #include <sys/msgport.h>
77 #include <sys/errno.h>
78 #include <libcaps/globaldata.h>
79 #include <sys/thread2.h>
80 #include <sys/msgport2.h>
84 #include <machine/cpufunc.h>
85 #include <machine/lock.h>
89 static int untimely_switch = 0;
91 static int token_debug = 0;
93 static __int64_t switch_count = 0;
94 static __int64_t preempt_hit = 0;
95 static __int64_t preempt_miss = 0;
96 static __int64_t preempt_weird = 0;
98 static __int64_t ipiq_count = 0;
99 static __int64_t ipiq_fifofull = 0;
104 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
106 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
109 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
120 * These helper procedures handle the runq, they can only be called from
121 * within a critical section.
123 * WARNING! Prior to SMP being brought up it is possible to enqueue and
124 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
125 * instead of 'mycpu' when referencing the globaldata structure. Once
126 * SMP live enqueuing and dequeueing only occurs on the current cpu.
130 _lwkt_dequeue(thread_t td)
132 if (td->td_flags & TDF_RUNQ) {
133 int nq = td->td_pri & TDPRI_MASK;
134 struct globaldata *gd = td->td_gd;
136 td->td_flags &= ~TDF_RUNQ;
137 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
138 /* runqmask is passively cleaned up by the switcher */
144 _lwkt_enqueue(thread_t td)
146 if ((td->td_flags & TDF_RUNQ) == 0) {
147 int nq = td->td_pri & TDPRI_MASK;
148 struct globaldata *gd = td->td_gd;
150 td->td_flags |= TDF_RUNQ;
151 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
152 gd->gd_runqmask |= 1 << nq;
158 _lwkt_wantresched(thread_t ntd, thread_t cur)
160 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
166 * LWKTs operate on a per-cpu basis
168 * WARNING! Called from early boot, 'mycpu' may not work yet.
171 lwkt_gdinit(struct globaldata *gd)
175 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
176 TAILQ_INIT(&gd->gd_tdrunq[i]);
178 TAILQ_INIT(&gd->gd_tdallq);
184 * Initialize a thread wait structure prior to first use.
186 * NOTE! called from low level boot code, we cannot do anything fancy!
189 lwkt_init_wait(lwkt_wait_t w)
191 TAILQ_INIT(&w->wa_waitq);
195 * Create a new thread. The thread must be associated with a process context
196 * or LWKT start address before it can be scheduled. If the target cpu is
197 * -1 the thread will be created on the current cpu.
199 * If you intend to create a thread without a process context this function
200 * does everything except load the startup and switcher function.
203 lwkt_alloc_thread(struct thread *td, int cpu)
210 if (mycpu->gd_tdfreecount > 0) {
211 --mycpu->gd_tdfreecount;
212 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
213 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
214 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
215 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
217 stack = td->td_kstack;
218 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
222 td = zalloc(thread_zone);
224 td = malloc(sizeof(struct thread));
226 td->td_kstack = NULL;
227 flags |= TDF_ALLOCATED_THREAD;
230 if ((stack = td->td_kstack) == NULL) {
232 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
234 stack = libcaps_alloc_stack(THREAD_STACK);
236 flags |= TDF_ALLOCATED_STACK;
239 lwkt_init_thread(td, stack, flags, mycpu);
241 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
248 * Initialize a preexisting thread structure. This function is used by
249 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
251 * All threads start out in a critical section at a priority of
252 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
253 * appropriate. This function may send an IPI message when the
254 * requested cpu is not the current cpu and consequently gd_tdallq may
255 * not be initialized synchronously from the point of view of the originating
258 * NOTE! we have to be careful in regards to creating threads for other cpus
259 * if SMP has not yet been activated.
262 lwkt_init_thread_remote(void *arg)
266 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
270 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
272 bzero(td, sizeof(struct thread));
273 td->td_kstack = stack;
274 td->td_flags |= flags;
276 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
277 lwkt_initport(&td->td_msgport, td);
278 pmap_init_thread(td);
279 if (smp_active == 0 || gd == mycpu) {
281 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
284 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
291 lwkt_set_comm(thread_t td, const char *ctl, ...)
296 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
301 lwkt_hold(thread_t td)
307 lwkt_rele(thread_t td)
309 KKASSERT(td->td_refs > 0);
316 lwkt_wait_free(thread_t td)
319 tsleep(td, 0, "tdreap", hz);
325 lwkt_free_thread(thread_t td)
327 struct globaldata *gd = mycpu;
329 KASSERT((td->td_flags & TDF_RUNNING) == 0,
330 ("lwkt_free_thread: did not exit! %p", td));
333 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
334 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
335 (td->td_flags & TDF_ALLOCATED_THREAD)
337 ++gd->gd_tdfreecount;
338 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
342 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
344 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
346 libcaps_free_stack(td->td_kstack, THREAD_STACK);
349 td->td_kstack = NULL;
351 if (td->td_flags & TDF_ALLOCATED_THREAD) {
353 zfree(thread_zone, td);
363 * Switch to the next runnable lwkt. If no LWKTs are runnable then
364 * switch to the idlethread. Switching must occur within a critical
365 * section to avoid races with the scheduling queue.
367 * We always have full control over our cpu's run queue. Other cpus
368 * that wish to manipulate our queue must use the cpu_*msg() calls to
369 * talk to our cpu, so a critical section is all that is needed and
370 * the result is very, very fast thread switching.
372 * The LWKT scheduler uses a fixed priority model and round-robins at
373 * each priority level. User process scheduling is a totally
374 * different beast and LWKT priorities should not be confused with
375 * user process priorities.
377 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
378 * cleans it up. Note that the td_switch() function cannot do anything that
379 * requires the MP lock since the MP lock will have already been setup for
380 * the target thread (not the current thread). It's nice to have a scheduler
381 * that does not need the MP lock to work because it allows us to do some
382 * really cool high-performance MP lock optimizations.
388 struct globaldata *gd;
389 thread_t td = curthread;
396 * Switching from within a 'fast' (non thread switched) interrupt is
399 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
400 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
404 * Passive release (used to transition from user to kernel mode
405 * when we block or switch rather then when we enter the kernel).
406 * This function is NOT called if we are switching into a preemption
407 * or returning from a preemption. Typically this causes us to lose
408 * our P_CURPROC designation (if we have one) and become a true LWKT
409 * thread, and may also hand P_CURPROC to another process and schedule
420 * td_mpcount cannot be used to determine if we currently hold the
421 * MP lock because get_mplock() will increment it prior to attempting
422 * to get the lock, and switch out if it can't. Our ownership of
423 * the actual lock will remain stable while we are in a critical section
424 * (but, of course, another cpu may own or release the lock so the
425 * actual value of mp_lock is not stable).
427 mpheld = MP_LOCK_HELD();
429 if ((ntd = td->td_preempted) != NULL) {
431 * We had preempted another thread on this cpu, resume the preempted
432 * thread. This occurs transparently, whether the preempted thread
433 * was scheduled or not (it may have been preempted after descheduling
436 * We have to setup the MP lock for the original thread after backing
437 * out the adjustment that was made to curthread when the original
440 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
442 if (ntd->td_mpcount && mpheld == 0) {
443 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
444 td, ntd, td->td_mpcount, ntd->td_mpcount);
446 if (ntd->td_mpcount) {
447 td->td_mpcount -= ntd->td_mpcount;
448 KKASSERT(td->td_mpcount >= 0);
451 ntd->td_flags |= TDF_PREEMPT_DONE;
452 /* YYY release mp lock on switchback if original doesn't need it */
455 * Priority queue / round-robin at each priority. Note that user
456 * processes run at a fixed, low priority and the user process
457 * scheduler deals with interactions between user processes
458 * by scheduling and descheduling them from the LWKT queue as
461 * We have to adjust the MP lock for the target thread. If we
462 * need the MP lock and cannot obtain it we try to locate a
463 * thread that does not need the MP lock.
467 if (gd->gd_runqmask) {
468 int nq = bsrl(gd->gd_runqmask);
469 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
470 gd->gd_runqmask &= ~(1 << nq);
474 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
476 * Target needs MP lock and we couldn't get it, try
477 * to locate a thread which does not need the MP lock
478 * to run. If we cannot locate a thread spin in idle.
480 u_int32_t rqmask = gd->gd_runqmask;
482 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
483 if (ntd->td_mpcount == 0)
488 rqmask &= ~(1 << nq);
492 ntd = &gd->gd_idlethread;
493 ntd->td_flags |= TDF_IDLE_NOHLT;
495 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
496 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
499 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
500 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
503 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
504 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
508 * We have nothing to run but only let the idle loop halt
509 * the cpu if there are no pending interrupts.
511 ntd = &gd->gd_idlethread;
512 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
513 ntd->td_flags |= TDF_IDLE_NOHLT;
516 KASSERT(ntd->td_pri >= TDPRI_CRIT,
517 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
520 * Do the actual switch. If the new target does not need the MP lock
521 * and we are holding it, release the MP lock. If the new target requires
522 * the MP lock we have already acquired it for the target.
525 if (ntd->td_mpcount == 0 ) {
529 ASSERT_MP_LOCK_HELD();
540 * Switch if another thread has a higher priority. Do not switch to other
541 * threads at the same priority.
546 struct globaldata *gd = mycpu;
547 struct thread *td = gd->gd_curthread;
549 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
555 * Request that the target thread preempt the current thread. Preemption
556 * only works under a specific set of conditions:
558 * - We are not preempting ourselves
559 * - The target thread is owned by the current cpu
560 * - We are not currently being preempted
561 * - The target is not currently being preempted
562 * - We are able to satisfy the target's MP lock requirements (if any).
564 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
565 * this is called via lwkt_schedule() through the td_preemptable callback.
566 * critpri is the managed critical priority that we should ignore in order
567 * to determine whether preemption is possible (aka usually just the crit
568 * priority of lwkt_schedule() itself).
570 * XXX at the moment we run the target thread in a critical section during
571 * the preemption in order to prevent the target from taking interrupts
572 * that *WE* can't. Preemption is strictly limited to interrupt threads
573 * and interrupt-like threads, outside of a critical section, and the
574 * preempted source thread will be resumed the instant the target blocks
575 * whether or not the source is scheduled (i.e. preemption is supposed to
576 * be as transparent as possible).
578 * The target thread inherits our MP count (added to its own) for the
579 * duration of the preemption in order to preserve the atomicy of the
580 * MP lock during the preemption. Therefore, any preempting targets must be
581 * careful in regards to MP assertions. Note that the MP count may be
582 * out of sync with the physical mp_lock, but we do not have to preserve
583 * the original ownership of the lock if it was out of synch (that is, we
584 * can leave it synchronized on return).
587 lwkt_preempt(thread_t ntd, int critpri)
589 struct globaldata *gd = mycpu;
590 thread_t td = gd->gd_curthread;
597 * The caller has put us in a critical section. We can only preempt
598 * if the caller of the caller was not in a critical section (basically
599 * a local interrupt), as determined by the 'critpri' parameter. If
600 * we are unable to preempt
602 * YYY The target thread must be in a critical section (else it must
603 * inherit our critical section? I dunno yet).
605 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
608 if (!_lwkt_wantresched(ntd, td)) {
612 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
617 if (ntd->td_gd != gd) {
622 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
626 if (ntd->td_preempted) {
632 * note: an interrupt might have occured just as we were transitioning
633 * to or from the MP lock. In this case td_mpcount will be pre-disposed
634 * (non-zero) but not actually synchronized with the actual state of the
635 * lock. We can use it to imply an MP lock requirement for the
636 * preemption but we cannot use it to test whether we hold the MP lock
639 savecnt = td->td_mpcount;
640 mpheld = MP_LOCK_HELD();
641 ntd->td_mpcount += td->td_mpcount;
642 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
643 ntd->td_mpcount -= td->td_mpcount;
650 ntd->td_preempted = td;
651 td->td_flags |= TDF_PREEMPT_LOCK;
653 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
655 KKASSERT(savecnt == td->td_mpcount);
656 mpheld = MP_LOCK_HELD();
657 if (mpheld && td->td_mpcount == 0)
659 else if (mpheld == 0 && td->td_mpcount)
660 panic("lwkt_preempt(): MP lock was not held through");
662 ntd->td_preempted = NULL;
663 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
667 * Yield our thread while higher priority threads are pending. This is
668 * typically called when we leave a critical section but it can be safely
669 * called while we are in a critical section.
671 * This function will not generally yield to equal priority threads but it
672 * can occur as a side effect. Note that lwkt_switch() is called from
673 * inside the critical section to prevent its own crit_exit() from reentering
674 * lwkt_yield_quick().
676 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
677 * came along but was blocked and made pending.
679 * (self contained on a per cpu basis)
682 lwkt_yield_quick(void)
684 globaldata_t gd = mycpu;
685 thread_t td = gd->gd_curthread;
688 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
689 * it with a non-zero cpl then we might not wind up calling splz after
690 * a task switch when the critical section is exited even though the
691 * new task could accept the interrupt.
693 * XXX from crit_exit() only called after last crit section is released.
694 * If called directly will run splz() even if in a critical section.
696 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
697 * except for this special case, we MUST call splz() here to handle any
698 * pending ints, particularly after we switch, or we might accidently
699 * halt the cpu with interrupts pending.
701 if (gd->gd_reqflags && td->td_nest_count < 2)
705 * YYY enabling will cause wakeup() to task-switch, which really
706 * confused the old 4.x code. This is a good way to simulate
707 * preemption and MP without actually doing preemption or MP, because a
708 * lot of code assumes that wakeup() does not block.
710 if (untimely_switch && td->td_nest_count == 0 &&
711 gd->gd_intr_nesting_level == 0
715 * YYY temporary hacks until we disassociate the userland scheduler
716 * from the LWKT scheduler.
718 if (td->td_flags & TDF_RUNQ) {
719 lwkt_switch(); /* will not reenter yield function */
721 lwkt_schedule_self(); /* make sure we are scheduled */
722 lwkt_switch(); /* will not reenter yield function */
723 lwkt_deschedule_self(); /* make sure we are descheduled */
725 crit_exit_noyield(td);
730 * This implements a normal yield which, unlike _quick, will yield to equal
731 * priority threads as well. Note that gd_reqflags tests will be handled by
732 * the crit_exit() call in lwkt_switch().
734 * (self contained on a per cpu basis)
739 lwkt_schedule_self();
744 * Schedule a thread to run. As the current thread we can always safely
745 * schedule ourselves, and a shortcut procedure is provided for that
748 * (non-blocking, self contained on a per cpu basis)
751 lwkt_schedule_self(void)
753 thread_t td = curthread;
756 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
759 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
760 panic("SCHED SELF PANIC");
766 * Generic schedule. Possibly schedule threads belonging to other cpus and
767 * deal with threads that might be blocked on a wait queue.
769 * YYY this is one of the best places to implement load balancing code.
770 * Load balancing can be accomplished by requesting other sorts of actions
771 * for the thread in question.
774 lwkt_schedule(thread_t td)
777 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
778 && td->td_proc->p_stat == SSLEEP
780 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
782 curthread->td_proc ? curthread->td_proc->p_pid : -1,
783 curthread->td_proc ? curthread->td_proc->p_stat : -1,
785 td->td_proc ? curthread->td_proc->p_pid : -1,
786 td->td_proc ? curthread->td_proc->p_stat : -1
788 panic("SCHED PANIC");
792 if (td == curthread) {
798 * If the thread is on a wait list we have to send our scheduling
799 * request to the owner of the wait structure. Otherwise we send
800 * the scheduling request to the cpu owning the thread. Races
801 * are ok, the target will forward the message as necessary (the
802 * message may chase the thread around before it finally gets
805 * (remember, wait structures use stable storage)
807 if ((w = td->td_wait) != NULL) {
808 if (lwkt_trytoken(&w->wa_token)) {
809 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
812 if (smp_active == 0 || td->td_gd == mycpu) {
814 if (td->td_preemptable) {
815 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
816 } else if (_lwkt_wantresched(td, curthread)) {
820 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
822 lwkt_reltoken(&w->wa_token);
824 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
828 * If the wait structure is NULL and we own the thread, there
829 * is no race (since we are in a critical section). If we
830 * do not own the thread there might be a race but the
831 * target cpu will deal with it.
833 if (smp_active == 0 || td->td_gd == mycpu) {
835 if (td->td_preemptable) {
836 td->td_preemptable(td, TDPRI_CRIT);
837 } else if (_lwkt_wantresched(td, curthread)) {
841 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
849 * Managed acquisition. This code assumes that the MP lock is held for
850 * the tdallq operation and that the thread has been descheduled from its
851 * original cpu. We also have to wait for the thread to be entirely switched
852 * out on its original cpu (this is usually fast enough that we never loop)
853 * since the LWKT system does not have to hold the MP lock while switching
854 * and the target may have released it before switching.
857 lwkt_acquire(thread_t td)
859 struct globaldata *gd;
862 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
863 while (td->td_flags & TDF_RUNNING) /* XXX spin */
867 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
870 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
876 * Deschedule a thread.
878 * (non-blocking, self contained on a per cpu basis)
881 lwkt_deschedule_self(void)
883 thread_t td = curthread;
886 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
892 * Generic deschedule. Descheduling threads other then your own should be
893 * done only in carefully controlled circumstances. Descheduling is
896 * This function may block if the cpu has run out of messages.
899 lwkt_deschedule(thread_t td)
902 if (td == curthread) {
905 if (td->td_gd == mycpu) {
908 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
915 * Set the target thread's priority. This routine does not automatically
916 * switch to a higher priority thread, LWKT threads are not designed for
917 * continuous priority changes. Yield if you want to switch.
919 * We have to retain the critical section count which uses the high bits
920 * of the td_pri field. The specified priority may also indicate zero or
921 * more critical sections by adding TDPRI_CRIT*N.
924 lwkt_setpri(thread_t td, int pri)
927 KKASSERT(td->td_gd == mycpu);
929 if (td->td_flags & TDF_RUNQ) {
931 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
934 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
940 lwkt_setpri_self(int pri)
942 thread_t td = curthread;
944 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
946 if (td->td_flags & TDF_RUNQ) {
948 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
951 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
957 lwkt_preempted_proc(void)
959 thread_t td = curthread;
960 while (td->td_preempted)
961 td = td->td_preempted;
965 typedef struct lwkt_gettoken_req {
973 * This function deschedules the current thread and blocks on the specified
974 * wait queue. We obtain ownership of the wait queue in order to block
975 * on it. A generation number is used to interlock the wait queue in case
976 * it gets signalled while we are blocked waiting on the token.
978 * Note: alternatively we could dequeue our thread and then message the
979 * target cpu owning the wait queue. YYY implement as sysctl.
981 * Note: wait queue signals normally ping-pong the cpu as an optimization.
985 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
987 thread_t td = curthread;
989 lwkt_gettoken(&w->wa_token);
990 if (w->wa_gen == *gen) {
992 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
995 td->td_wmesg = wmesg;
998 lwkt_regettoken(&w->wa_token);
999 if (td->td_wmesg != NULL) {
1004 /* token might be lost, doesn't matter for gen update */
1006 lwkt_reltoken(&w->wa_token);
1010 * Signal a wait queue. We gain ownership of the wait queue in order to
1011 * signal it. Once a thread is removed from the wait queue we have to
1012 * deal with the cpu owning the thread.
1014 * Note: alternatively we could message the target cpu owning the wait
1015 * queue. YYY implement as sysctl.
1018 lwkt_signal(lwkt_wait_t w, int count)
1023 lwkt_gettoken(&w->wa_token);
1026 count = w->wa_count;
1027 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1030 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1032 td->td_wmesg = NULL;
1033 if (td->td_gd == mycpu) {
1036 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
1038 lwkt_regettoken(&w->wa_token);
1040 lwkt_reltoken(&w->wa_token);
1046 * Acquire ownership of a token
1048 * Acquire ownership of a token. The token may have spl and/or critical
1049 * section side effects, depending on its purpose. These side effects
1050 * guarentee that you will maintain ownership of the token as long as you
1051 * do not block. If you block you may lose access to the token (but you
1052 * must still release it even if you lose your access to it).
1054 * YYY for now we use a critical section to prevent IPIs from taking away
1055 * a token, but do we really only need to disable IPIs ?
1057 * YYY certain tokens could be made to act like mutexes when performance
1058 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1060 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1061 * means that 4.x does not expect a switch so for now we cannot switch
1062 * when waiting for an IPI to be returned.
1064 * YYY If the token is owned by another cpu we may have to send an IPI to
1065 * it and then block. The IPI causes the token to be given away to the
1066 * requesting cpu, unless it has already changed hands. Since only the
1067 * current cpu can give away a token it owns we do not need a memory barrier.
1068 * This needs serious optimization.
1075 lwkt_gettoken_remote(void *arg)
1077 lwkt_gettoken_req *req = arg;
1078 if (req->tok->t_cpu == mycpu->gd_cpuid) {
1081 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
1083 req->tok->t_cpu = req->cpu;
1084 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1085 /* else set reqcpu to point to current cpu for release */
1092 lwkt_gettoken(lwkt_token_t tok)
1095 * Prevent preemption so the token can't be taken away from us once
1096 * we gain ownership of it. Use a synchronous request which might
1097 * block. The request will be forwarded as necessary playing catchup
1103 if (curthread->td_pri > 1800) {
1104 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1105 tok, ((int **)&tok)[-1]);
1107 if (curthread->td_pri > 2000) {
1108 curthread->td_pri = 1000;
1113 while (tok->t_cpu != mycpu->gd_cpuid) {
1114 struct lwkt_gettoken_req req;
1118 req.cpu = mycpu->gd_cpuid;
1120 dcpu = (volatile int)tok->t_cpu;
1121 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1124 printf("REQT%d ", dcpu);
1126 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1127 lwkt_wait_ipiq(dcpu, seq);
1130 printf("REQR%d ", tok->t_cpu);
1135 * leave us in a critical section on return. This will be undone
1136 * by lwkt_reltoken(). Bump the generation number.
1138 return(++tok->t_gen);
1142 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1146 lwkt_trytoken(lwkt_token_t tok)
1150 if (tok->t_cpu != mycpu->gd_cpuid) {
1155 /* leave us in the critical section */
1161 * Release your ownership of a token. Releases must occur in reverse
1162 * order to aquisitions, eventually so priorities can be unwound properly
1163 * like SPLs. At the moment the actual implemention doesn't care.
1165 * We can safely hand a token that we own to another cpu without notifying
1166 * it, but once we do we can't get it back without requesting it (unless
1167 * the other cpu hands it back to us before we check).
1169 * We might have lost the token, so check that.
1171 * Return the token's generation number. The number is useful to callers
1172 * who may want to know if the token was stolen during potential blockages.
1175 lwkt_reltoken(lwkt_token_t tok)
1179 if (tok->t_cpu == mycpu->gd_cpuid) {
1180 tok->t_cpu = tok->t_reqcpu;
1188 * Reacquire a token that might have been lost. 0 is returned if the
1189 * generation has not changed (nobody stole the token from us), -1 is
1190 * returned otherwise. The token is reacquired regardless but the
1191 * generation number is not bumped further if we already own the token.
1193 * For efficiency we inline the best-case situation for lwkt_regettoken()
1194 * (i.e .we still own the token).
1197 lwkt_gentoken(lwkt_token_t tok, int *gen)
1199 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
1201 *gen = lwkt_regettoken(tok);
1206 * Re-acquire a token that might have been lost. The generation number
1207 * is bumped and returned regardless of whether the token had been lost
1208 * or not (because we only have cpu granularity we have to bump the token
1212 lwkt_regettoken(lwkt_token_t tok)
1214 /* assert we are in a critical section */
1215 if (tok->t_cpu != mycpu->gd_cpuid) {
1217 while (tok->t_cpu != mycpu->gd_cpuid) {
1218 struct lwkt_gettoken_req req;
1222 req.cpu = mycpu->gd_cpuid;
1224 dcpu = (volatile int)tok->t_cpu;
1225 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1228 printf("REQT%d ", dcpu);
1230 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1231 lwkt_wait_ipiq(dcpu, seq);
1234 printf("REQR%d ", tok->t_cpu);
1244 lwkt_inittoken(lwkt_token_t tok)
1247 * Zero structure and set cpu owner and reqcpu to cpu 0.
1249 bzero(tok, sizeof(*tok));
1253 * Create a kernel process/thread/whatever. It shares it's address space
1254 * with proc0 - ie: kernel only.
1256 * NOTE! By default new threads are created with the MP lock held. A
1257 * thread which does not require the MP lock should release it by calling
1258 * rel_mplock() at the start of the new thread.
1261 lwkt_create(void (*func)(void *), void *arg,
1262 struct thread **tdp, thread_t template, int tdflags, int cpu,
1263 const char *fmt, ...)
1268 td = lwkt_alloc_thread(template, cpu);
1271 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1272 td->td_flags |= TDF_VERBOSE | tdflags;
1278 * Set up arg0 for 'ps' etc
1280 __va_start(ap, fmt);
1281 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1285 * Schedule the thread to run
1287 if ((td->td_flags & TDF_STOPREQ) == 0)
1290 td->td_flags &= ~TDF_STOPREQ;
1295 * kthread_* is specific to the kernel and is not needed by userland.
1300 * Destroy an LWKT thread. Warning! This function is not called when
1301 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1302 * uses a different reaping mechanism.
1307 thread_t td = curthread;
1309 if (td->td_flags & TDF_VERBOSE)
1310 printf("kthread %p %s has exited\n", td, td->td_comm);
1312 lwkt_deschedule_self();
1313 ++mycpu->gd_tdfreecount;
1314 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1319 * Create a kernel process/thread/whatever. It shares it's address space
1320 * with proc0 - ie: kernel only. 5.x compatible.
1322 * NOTE! By default kthreads are created with the MP lock held. A
1323 * thread which does not require the MP lock should release it by calling
1324 * rel_mplock() at the start of the new thread.
1327 kthread_create(void (*func)(void *), void *arg,
1328 struct thread **tdp, const char *fmt, ...)
1333 td = lwkt_alloc_thread(NULL, -1);
1336 cpu_set_thread_handler(td, kthread_exit, func, arg);
1337 td->td_flags |= TDF_VERBOSE;
1343 * Set up arg0 for 'ps' etc
1345 __va_start(ap, fmt);
1346 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1350 * Schedule the thread to run
1357 * Destroy an LWKT thread. Warning! This function is not called when
1358 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1359 * uses a different reaping mechanism.
1361 * XXX duplicates lwkt_exit()
1369 #endif /* _KERNEL */
1374 thread_t td = curthread;
1375 int lpri = td->td_pri;
1378 panic("td_pri is/would-go negative! %p %d", td, lpri);
1384 * Send a function execution request to another cpu. The request is queued
1385 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1386 * possible target cpu. The FIFO can be written.
1388 * YYY If the FIFO fills up we have to enable interrupts and process the
1389 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1390 * Create a CPU_*() function to do this!
1392 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1393 * end will take care of any pending interrupts.
1395 * Must be called from a critical section.
1398 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1402 struct globaldata *gd = mycpu;
1404 if (dcpu == gd->gd_cpuid) {
1409 ++gd->gd_intr_nesting_level;
1411 if (gd->gd_intr_nesting_level > 20)
1412 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1414 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1415 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1417 ip = &gd->gd_ipiq[dcpu];
1420 * We always drain before the FIFO becomes full so it should never
1421 * become full. We need to leave enough entries to deal with
1424 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1425 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1426 ip->ip_func[windex] = func;
1427 ip->ip_arg[windex] = arg;
1428 /* YYY memory barrier */
1430 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1431 unsigned int eflags = read_eflags();
1434 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1435 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1436 lwkt_process_ipiq();
1438 write_eflags(eflags);
1440 --gd->gd_intr_nesting_level;
1441 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1443 return(ip->ip_windex);
1447 * Send a message to several target cpus. Typically used for scheduling.
1448 * The message will not be sent to stopped cpus.
1451 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1455 mask &= ~stopped_cpus;
1458 lwkt_send_ipiq(cpuid, func, arg);
1459 mask &= ~(1 << cpuid);
1464 * Wait for the remote cpu to finish processing a function.
1466 * YYY we have to enable interrupts and process the IPIQ while waiting
1467 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1468 * function to do this! YYY we really should 'block' here.
1470 * Must be called from a critical section. Thsi routine may be called
1471 * from an interrupt (for example, if an interrupt wakes a foreign thread
1475 lwkt_wait_ipiq(int dcpu, int seq)
1478 int maxc = 100000000;
1480 if (dcpu != mycpu->gd_cpuid) {
1481 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1482 ip = &mycpu->gd_ipiq[dcpu];
1483 if ((int)(ip->ip_xindex - seq) < 0) {
1484 unsigned int eflags = read_eflags();
1486 while ((int)(ip->ip_xindex - seq) < 0) {
1487 lwkt_process_ipiq();
1489 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1490 if (maxc < -1000000)
1491 panic("LWKT_WAIT_IPIQ");
1493 write_eflags(eflags);
1499 * Called from IPI interrupt (like a fast interrupt), which has placed
1500 * us in a critical section. The MP lock may or may not be held.
1501 * May also be called from doreti or splz, or be reentrantly called
1502 * indirectly through the ip_func[] we run.
1505 lwkt_process_ipiq(void)
1508 int cpuid = mycpu->gd_cpuid;
1510 for (n = 0; n < ncpus; ++n) {
1516 ip = globaldata_find(n)->gd_ipiq;
1522 * Note: xindex is only updated after we are sure the function has
1523 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1524 * function may send an IPI which may block/drain.
1526 while (ip->ip_rindex != ip->ip_windex) {
1527 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1529 ip->ip_func[ri](ip->ip_arg[ri]);
1530 /* YYY memory barrier */
1531 ip->ip_xindex = ip->ip_rindex;
1539 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1541 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1542 return(0); /* NOT REACHED */
1546 lwkt_wait_ipiq(int dcpu, int seq)
1548 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);