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 1.41 2003/11/09 02:22:36 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).
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
45 #include <sys/rtprio.h>
46 #include <sys/queue.h>
47 #include <sys/thread2.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
54 #include <vm/vm_param.h>
55 #include <vm/vm_kern.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_page.h>
58 #include <vm/vm_map.h>
59 #include <vm/vm_pager.h>
60 #include <vm/vm_extern.h>
61 #include <vm/vm_zone.h>
63 #include <machine/stdarg.h>
64 #include <machine/ipl.h>
65 #include <machine/smp.h>
67 static int untimely_switch = 0;
68 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
70 static int token_debug = 0;
71 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
73 static quad_t switch_count = 0;
74 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
75 static quad_t preempt_hit = 0;
76 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
77 static quad_t preempt_miss = 0;
78 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
79 static quad_t preempt_weird = 0;
80 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
81 static quad_t ipiq_count = 0;
82 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
83 static quad_t ipiq_fifofull = 0;
84 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
87 * These helper procedures handle the runq, they can only be called from
88 * within a critical section.
90 * WARNING! Prior to SMP being brought up it is possible to enqueue and
91 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
92 * instead of 'mycpu' when referencing the globaldata structure. Once
93 * SMP live enqueuing and dequeueing only occurs on the current cpu.
97 _lwkt_dequeue(thread_t td)
99 if (td->td_flags & TDF_RUNQ) {
100 int nq = td->td_pri & TDPRI_MASK;
101 struct globaldata *gd = td->td_gd;
103 td->td_flags &= ~TDF_RUNQ;
104 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
105 /* runqmask is passively cleaned up by the switcher */
111 _lwkt_enqueue(thread_t td)
113 if ((td->td_flags & TDF_RUNQ) == 0) {
114 int nq = td->td_pri & TDPRI_MASK;
115 struct globaldata *gd = td->td_gd;
117 td->td_flags |= TDF_RUNQ;
118 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
119 gd->gd_runqmask |= 1 << nq;
125 _lwkt_wantresched(thread_t ntd, thread_t cur)
127 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
131 * LWKTs operate on a per-cpu basis
133 * WARNING! Called from early boot, 'mycpu' may not work yet.
136 lwkt_gdinit(struct globaldata *gd)
140 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
141 TAILQ_INIT(&gd->gd_tdrunq[i]);
143 TAILQ_INIT(&gd->gd_tdallq);
147 * Initialize a thread wait structure prior to first use.
149 * NOTE! called from low level boot code, we cannot do anything fancy!
152 lwkt_init_wait(lwkt_wait_t w)
154 TAILQ_INIT(&w->wa_waitq);
158 * Create a new thread. The thread must be associated with a process context
159 * or LWKT start address before it can be scheduled. If the target cpu is
160 * -1 the thread will be created on the current cpu.
162 * If you intend to create a thread without a process context this function
163 * does everything except load the startup and switcher function.
166 lwkt_alloc_thread(struct thread *td, int cpu)
173 if (mycpu->gd_tdfreecount > 0) {
174 --mycpu->gd_tdfreecount;
175 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
176 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
177 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
178 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
180 stack = td->td_kstack;
181 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
184 td = zalloc(thread_zone);
185 td->td_kstack = NULL;
186 flags |= TDF_ALLOCATED_THREAD;
189 if ((stack = td->td_kstack) == NULL) {
190 stack = (void *)kmem_alloc(kernel_map, UPAGES * PAGE_SIZE);
191 flags |= TDF_ALLOCATED_STACK;
194 lwkt_init_thread(td, stack, flags, mycpu);
196 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
201 * Initialize a preexisting thread structure. This function is used by
202 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
204 * All threads start out in a critical section at a priority of
205 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
206 * appropriate. This function may send an IPI message when the
207 * requested cpu is not the current cpu and consequently gd_tdallq may
208 * not be initialized synchronously from the point of view of the originating
211 * NOTE! we have to be careful in regards to creating threads for other cpus
212 * if SMP has not yet been activated.
215 lwkt_init_thread_remote(void *arg)
219 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
223 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
225 bzero(td, sizeof(struct thread));
226 td->td_kstack = stack;
227 td->td_flags |= flags;
229 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
230 lwkt_init_port(&td->td_msgport, td);
231 pmap_init_thread(td);
232 if (smp_active == 0 || gd == mycpu) {
234 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
237 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
242 lwkt_set_comm(thread_t td, const char *ctl, ...)
247 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
252 lwkt_hold(thread_t td)
258 lwkt_rele(thread_t td)
260 KKASSERT(td->td_refs > 0);
265 lwkt_wait_free(thread_t td)
268 tsleep(td, 0, "tdreap", hz);
272 lwkt_free_thread(thread_t td)
274 struct globaldata *gd = mycpu;
276 KASSERT((td->td_flags & TDF_RUNNING) == 0,
277 ("lwkt_free_thread: did not exit! %p", td));
280 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
281 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
282 (td->td_flags & TDF_ALLOCATED_THREAD)
284 ++gd->gd_tdfreecount;
285 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
289 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
290 kmem_free(kernel_map,
291 (vm_offset_t)td->td_kstack, UPAGES * PAGE_SIZE);
293 td->td_kstack = NULL;
295 if (td->td_flags & TDF_ALLOCATED_THREAD)
296 zfree(thread_zone, td);
302 * Switch to the next runnable lwkt. If no LWKTs are runnable then
303 * switch to the idlethread. Switching must occur within a critical
304 * section to avoid races with the scheduling queue.
306 * We always have full control over our cpu's run queue. Other cpus
307 * that wish to manipulate our queue must use the cpu_*msg() calls to
308 * talk to our cpu, so a critical section is all that is needed and
309 * the result is very, very fast thread switching.
311 * The LWKT scheduler uses a fixed priority model and round-robins at
312 * each priority level. User process scheduling is a totally
313 * different beast and LWKT priorities should not be confused with
314 * user process priorities.
316 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
317 * cleans it up. Note that the td_switch() function cannot do anything that
318 * requires the MP lock since the MP lock will have already been setup for
319 * the target thread (not the current thread). It's nice to have a scheduler
320 * that does not need the MP lock to work because it allows us to do some
321 * really cool high-performance MP lock optimizations.
327 struct globaldata *gd;
328 thread_t td = curthread;
335 * Switching from within a 'fast' (non thread switched) interrupt is
338 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
339 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
343 * Passive release (used to transition from user to kernel mode
344 * when we block or switch rather then when we enter the kernel).
345 * This function is NOT called if we are switching into a preemption
346 * or returning from a preemption. Typically this causes us to lose
347 * our P_CURPROC designation (if we have one) and become a true LWKT
348 * thread, and may also hand P_CURPROC to another process and schedule
359 * td_mpcount cannot be used to determine if we currently hold the
360 * MP lock because get_mplock() will increment it prior to attempting
361 * to get the lock, and switch out if it can't. Our ownership of
362 * the actual lock will remain stable while we are in a critical section
363 * (but, of course, another cpu may own or release the lock so the
364 * actual value of mp_lock is not stable).
366 mpheld = MP_LOCK_HELD();
368 if ((ntd = td->td_preempted) != NULL) {
370 * We had preempted another thread on this cpu, resume the preempted
371 * thread. This occurs transparently, whether the preempted thread
372 * was scheduled or not (it may have been preempted after descheduling
375 * We have to setup the MP lock for the original thread after backing
376 * out the adjustment that was made to curthread when the original
379 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
381 if (ntd->td_mpcount && mpheld == 0) {
382 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
383 td, ntd, td->td_mpcount, ntd->td_mpcount);
385 if (ntd->td_mpcount) {
386 td->td_mpcount -= ntd->td_mpcount;
387 KKASSERT(td->td_mpcount >= 0);
390 ntd->td_flags |= TDF_PREEMPT_DONE;
391 /* YYY release mp lock on switchback if original doesn't need it */
394 * Priority queue / round-robin at each priority. Note that user
395 * processes run at a fixed, low priority and the user process
396 * scheduler deals with interactions between user processes
397 * by scheduling and descheduling them from the LWKT queue as
400 * We have to adjust the MP lock for the target thread. If we
401 * need the MP lock and cannot obtain it we try to locate a
402 * thread that does not need the MP lock.
406 if (gd->gd_runqmask) {
407 int nq = bsrl(gd->gd_runqmask);
408 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
409 gd->gd_runqmask &= ~(1 << nq);
413 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
415 * Target needs MP lock and we couldn't get it, try
416 * to locate a thread which does not need the MP lock
417 * to run. If we cannot locate a thread spin in idle.
419 u_int32_t rqmask = gd->gd_runqmask;
421 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
422 if (ntd->td_mpcount == 0)
427 rqmask &= ~(1 << nq);
431 ntd = &gd->gd_idlethread;
432 ntd->td_flags |= TDF_IDLE_NOHLT;
434 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
435 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
438 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
439 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
442 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
443 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
447 * Nothing to run but we may still need the BGL to deal with
448 * pending interrupts, spin in idle if so.
450 ntd = &gd->gd_idlethread;
452 ntd->td_flags |= TDF_IDLE_NOHLT;
455 KASSERT(ntd->td_pri >= TDPRI_CRIT,
456 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
459 * Do the actual switch. If the new target does not need the MP lock
460 * and we are holding it, release the MP lock. If the new target requires
461 * the MP lock we have already acquired it for the target.
464 if (ntd->td_mpcount == 0 ) {
468 ASSERT_MP_LOCK_HELD();
479 * Switch if another thread has a higher priority. Do not switch to other
480 * threads at the same priority.
485 struct globaldata *gd = mycpu;
486 struct thread *td = gd->gd_curthread;
488 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
494 * Request that the target thread preempt the current thread. Preemption
495 * only works under a specific set of conditions:
497 * - We are not preempting ourselves
498 * - The target thread is owned by the current cpu
499 * - We are not currently being preempted
500 * - The target is not currently being preempted
501 * - We are able to satisfy the target's MP lock requirements (if any).
503 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
504 * this is called via lwkt_schedule() through the td_preemptable callback.
505 * critpri is the managed critical priority that we should ignore in order
506 * to determine whether preemption is possible (aka usually just the crit
507 * priority of lwkt_schedule() itself).
509 * XXX at the moment we run the target thread in a critical section during
510 * the preemption in order to prevent the target from taking interrupts
511 * that *WE* can't. Preemption is strictly limited to interrupt threads
512 * and interrupt-like threads, outside of a critical section, and the
513 * preempted source thread will be resumed the instant the target blocks
514 * whether or not the source is scheduled (i.e. preemption is supposed to
515 * be as transparent as possible).
517 * The target thread inherits our MP count (added to its own) for the
518 * duration of the preemption in order to preserve the atomicy of the
519 * MP lock during the preemption. Therefore, any preempting targets must be
520 * careful in regards to MP assertions. Note that the MP count may be
521 * out of sync with the physical mp_lock, but we do not have to preserve
522 * the original ownership of the lock if it was out of synch (that is, we
523 * can leave it synchronized on return).
526 lwkt_preempt(thread_t ntd, int critpri)
528 struct globaldata *gd = mycpu;
529 thread_t td = gd->gd_curthread;
536 * The caller has put us in a critical section. We can only preempt
537 * if the caller of the caller was not in a critical section (basically
538 * a local interrupt), as determined by the 'critpri' parameter. If
539 * we are unable to preempt
541 * YYY The target thread must be in a critical section (else it must
542 * inherit our critical section? I dunno yet).
544 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
547 if (!_lwkt_wantresched(ntd, td)) {
551 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
556 if (ntd->td_gd != gd) {
561 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
565 if (ntd->td_preempted) {
571 * note: an interrupt might have occured just as we were transitioning
572 * to or from the MP lock. In this case td_mpcount will be pre-disposed
573 * (non-zero) but not actually synchronized with the actual state of the
574 * lock. We can use it to imply an MP lock requirement for the
575 * preemption but we cannot use it to test whether we hold the MP lock
578 savecnt = td->td_mpcount;
579 mpheld = MP_LOCK_HELD();
580 ntd->td_mpcount += td->td_mpcount;
581 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
582 ntd->td_mpcount -= td->td_mpcount;
589 ntd->td_preempted = td;
590 td->td_flags |= TDF_PREEMPT_LOCK;
592 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
594 KKASSERT(savecnt == td->td_mpcount);
595 mpheld = MP_LOCK_HELD();
596 if (mpheld && td->td_mpcount == 0)
598 else if (mpheld == 0 && td->td_mpcount)
599 panic("lwkt_preempt(): MP lock was not held through");
601 ntd->td_preempted = NULL;
602 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
606 * Yield our thread while higher priority threads are pending. This is
607 * typically called when we leave a critical section but it can be safely
608 * called while we are in a critical section.
610 * This function will not generally yield to equal priority threads but it
611 * can occur as a side effect. Note that lwkt_switch() is called from
612 * inside the critical section to prevent its own crit_exit() from reentering
613 * lwkt_yield_quick().
615 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
616 * came along but was blocked and made pending.
618 * (self contained on a per cpu basis)
621 lwkt_yield_quick(void)
623 globaldata_t gd = mycpu;
624 thread_t td = gd->gd_curthread;
627 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
628 * it with a non-zero cpl then we might not wind up calling splz after
629 * a task switch when the critical section is exited even though the
630 * new task could accept the interrupt.
632 * XXX from crit_exit() only called after last crit section is released.
633 * If called directly will run splz() even if in a critical section.
635 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
636 * except for this special case, we MUST call splz() here to handle any
637 * pending ints, particularly after we switch, or we might accidently
638 * halt the cpu with interrupts pending.
640 if (gd->gd_reqflags && td->td_nest_count < 2)
644 * YYY enabling will cause wakeup() to task-switch, which really
645 * confused the old 4.x code. This is a good way to simulate
646 * preemption and MP without actually doing preemption or MP, because a
647 * lot of code assumes that wakeup() does not block.
649 if (untimely_switch && td->td_nest_count == 0 &&
650 gd->gd_intr_nesting_level == 0
654 * YYY temporary hacks until we disassociate the userland scheduler
655 * from the LWKT scheduler.
657 if (td->td_flags & TDF_RUNQ) {
658 lwkt_switch(); /* will not reenter yield function */
660 lwkt_schedule_self(); /* make sure we are scheduled */
661 lwkt_switch(); /* will not reenter yield function */
662 lwkt_deschedule_self(); /* make sure we are descheduled */
664 crit_exit_noyield(td);
669 * This implements a normal yield which, unlike _quick, will yield to equal
670 * priority threads as well. Note that gd_reqflags tests will be handled by
671 * the crit_exit() call in lwkt_switch().
673 * (self contained on a per cpu basis)
678 lwkt_schedule_self();
683 * Schedule a thread to run. As the current thread we can always safely
684 * schedule ourselves, and a shortcut procedure is provided for that
687 * (non-blocking, self contained on a per cpu basis)
690 lwkt_schedule_self(void)
692 thread_t td = curthread;
695 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
697 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
698 panic("SCHED SELF PANIC");
703 * Generic schedule. Possibly schedule threads belonging to other cpus and
704 * deal with threads that might be blocked on a wait queue.
706 * YYY this is one of the best places to implement load balancing code.
707 * Load balancing can be accomplished by requesting other sorts of actions
708 * for the thread in question.
711 lwkt_schedule(thread_t td)
714 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
715 && td->td_proc->p_stat == SSLEEP
717 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
719 curthread->td_proc ? curthread->td_proc->p_pid : -1,
720 curthread->td_proc ? curthread->td_proc->p_stat : -1,
722 td->td_proc ? curthread->td_proc->p_pid : -1,
723 td->td_proc ? curthread->td_proc->p_stat : -1
725 panic("SCHED PANIC");
729 if (td == curthread) {
735 * If the thread is on a wait list we have to send our scheduling
736 * request to the owner of the wait structure. Otherwise we send
737 * the scheduling request to the cpu owning the thread. Races
738 * are ok, the target will forward the message as necessary (the
739 * message may chase the thread around before it finally gets
742 * (remember, wait structures use stable storage)
744 if ((w = td->td_wait) != NULL) {
745 if (lwkt_trytoken(&w->wa_token)) {
746 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
749 if (smp_active == 0 || td->td_gd == mycpu) {
751 if (td->td_preemptable) {
752 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
753 } else if (_lwkt_wantresched(td, curthread)) {
757 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
759 lwkt_reltoken(&w->wa_token);
761 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
765 * If the wait structure is NULL and we own the thread, there
766 * is no race (since we are in a critical section). If we
767 * do not own the thread there might be a race but the
768 * target cpu will deal with it.
770 if (smp_active == 0 || td->td_gd == mycpu) {
772 if (td->td_preemptable) {
773 td->td_preemptable(td, TDPRI_CRIT);
774 } else if (_lwkt_wantresched(td, curthread)) {
778 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
786 * Managed acquisition. This code assumes that the MP lock is held for
787 * the tdallq operation and that the thread has been descheduled from its
788 * original cpu. We also have to wait for the thread to be entirely switched
789 * out on its original cpu (this is usually fast enough that we never loop)
790 * since the LWKT system does not have to hold the MP lock while switching
791 * and the target may have released it before switching.
794 lwkt_acquire(thread_t td)
796 struct globaldata *gd;
799 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
800 while (td->td_flags & TDF_RUNNING) /* XXX spin */
804 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
807 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
813 * Deschedule a thread.
815 * (non-blocking, self contained on a per cpu basis)
818 lwkt_deschedule_self(void)
820 thread_t td = curthread;
823 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
829 * Generic deschedule. Descheduling threads other then your own should be
830 * done only in carefully controlled circumstances. Descheduling is
833 * This function may block if the cpu has run out of messages.
836 lwkt_deschedule(thread_t td)
839 if (td == curthread) {
842 if (td->td_gd == mycpu) {
845 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
852 * Set the target thread's priority. This routine does not automatically
853 * switch to a higher priority thread, LWKT threads are not designed for
854 * continuous priority changes. Yield if you want to switch.
856 * We have to retain the critical section count which uses the high bits
857 * of the td_pri field. The specified priority may also indicate zero or
858 * more critical sections by adding TDPRI_CRIT*N.
861 lwkt_setpri(thread_t td, int pri)
864 KKASSERT(td->td_gd == mycpu);
866 if (td->td_flags & TDF_RUNQ) {
868 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
871 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
877 lwkt_setpri_self(int pri)
879 thread_t td = curthread;
881 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
883 if (td->td_flags & TDF_RUNQ) {
885 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
888 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
894 lwkt_preempted_proc(void)
896 thread_t td = curthread;
897 while (td->td_preempted)
898 td = td->td_preempted;
902 typedef struct lwkt_gettoken_req {
910 * This function deschedules the current thread and blocks on the specified
911 * wait queue. We obtain ownership of the wait queue in order to block
912 * on it. A generation number is used to interlock the wait queue in case
913 * it gets signalled while we are blocked waiting on the token.
915 * Note: alternatively we could dequeue our thread and then message the
916 * target cpu owning the wait queue. YYY implement as sysctl.
918 * Note: wait queue signals normally ping-pong the cpu as an optimization.
922 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
924 thread_t td = curthread;
926 lwkt_gettoken(&w->wa_token);
927 if (w->wa_gen == *gen) {
929 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
932 td->td_wmesg = wmesg;
935 lwkt_regettoken(&w->wa_token);
936 if (td->td_wmesg != NULL) {
941 /* token might be lost, doesn't matter for gen update */
943 lwkt_reltoken(&w->wa_token);
947 * Signal a wait queue. We gain ownership of the wait queue in order to
948 * signal it. Once a thread is removed from the wait queue we have to
949 * deal with the cpu owning the thread.
951 * Note: alternatively we could message the target cpu owning the wait
952 * queue. YYY implement as sysctl.
955 lwkt_signal(lwkt_wait_t w, int count)
960 lwkt_gettoken(&w->wa_token);
964 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
967 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
970 if (td->td_gd == mycpu) {
973 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
975 lwkt_regettoken(&w->wa_token);
977 lwkt_reltoken(&w->wa_token);
983 * Acquire ownership of a token
985 * Acquire ownership of a token. The token may have spl and/or critical
986 * section side effects, depending on its purpose. These side effects
987 * guarentee that you will maintain ownership of the token as long as you
988 * do not block. If you block you may lose access to the token (but you
989 * must still release it even if you lose your access to it).
991 * YYY for now we use a critical section to prevent IPIs from taking away
992 * a token, but do we really only need to disable IPIs ?
994 * YYY certain tokens could be made to act like mutexes when performance
995 * would be better (e.g. t_cpu == -1). This is not yet implemented.
997 * YYY the tokens replace 4.x's simplelocks for the most part, but this
998 * means that 4.x does not expect a switch so for now we cannot switch
999 * when waiting for an IPI to be returned.
1001 * YYY If the token is owned by another cpu we may have to send an IPI to
1002 * it and then block. The IPI causes the token to be given away to the
1003 * requesting cpu, unless it has already changed hands. Since only the
1004 * current cpu can give away a token it owns we do not need a memory barrier.
1005 * This needs serious optimization.
1012 lwkt_gettoken_remote(void *arg)
1014 lwkt_gettoken_req *req = arg;
1015 if (req->tok->t_cpu == mycpu->gd_cpuid) {
1018 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
1020 req->tok->t_cpu = req->cpu;
1021 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1022 /* else set reqcpu to point to current cpu for release */
1029 lwkt_gettoken(lwkt_token_t tok)
1032 * Prevent preemption so the token can't be taken away from us once
1033 * we gain ownership of it. Use a synchronous request which might
1034 * block. The request will be forwarded as necessary playing catchup
1040 if (curthread->td_pri > 1800) {
1041 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1042 tok, ((int **)&tok)[-1]);
1044 if (curthread->td_pri > 2000) {
1045 curthread->td_pri = 1000;
1050 while (tok->t_cpu != mycpu->gd_cpuid) {
1051 struct lwkt_gettoken_req req;
1055 req.cpu = mycpu->gd_cpuid;
1057 dcpu = (volatile int)tok->t_cpu;
1058 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1061 printf("REQT%d ", dcpu);
1063 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1064 lwkt_wait_ipiq(dcpu, seq);
1067 printf("REQR%d ", tok->t_cpu);
1072 * leave us in a critical section on return. This will be undone
1073 * by lwkt_reltoken(). Bump the generation number.
1075 return(++tok->t_gen);
1079 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1083 lwkt_trytoken(lwkt_token_t tok)
1087 if (tok->t_cpu != mycpu->gd_cpuid) {
1092 /* leave us in the critical section */
1098 * Release your ownership of a token. Releases must occur in reverse
1099 * order to aquisitions, eventually so priorities can be unwound properly
1100 * like SPLs. At the moment the actual implemention doesn't care.
1102 * We can safely hand a token that we own to another cpu without notifying
1103 * it, but once we do we can't get it back without requesting it (unless
1104 * the other cpu hands it back to us before we check).
1106 * We might have lost the token, so check that.
1108 * Return the token's generation number. The number is useful to callers
1109 * who may want to know if the token was stolen during potential blockages.
1112 lwkt_reltoken(lwkt_token_t tok)
1116 if (tok->t_cpu == mycpu->gd_cpuid) {
1117 tok->t_cpu = tok->t_reqcpu;
1125 * Reacquire a token that might have been lost. 0 is returned if the
1126 * generation has not changed (nobody stole the token from us), -1 is
1127 * returned otherwise. The token is reacquired regardless but the
1128 * generation number is not bumped further if we already own the token.
1130 * For efficiency we inline the best-case situation for lwkt_regettoken()
1131 * (i.e .we still own the token).
1134 lwkt_gentoken(lwkt_token_t tok, int *gen)
1136 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
1138 *gen = lwkt_regettoken(tok);
1143 * Re-acquire a token that might have been lost. The generation number
1144 * is bumped and returned regardless of whether the token had been lost
1145 * or not (because we only have cpu granularity we have to bump the token
1149 lwkt_regettoken(lwkt_token_t tok)
1151 /* assert we are in a critical section */
1152 if (tok->t_cpu != mycpu->gd_cpuid) {
1154 while (tok->t_cpu != mycpu->gd_cpuid) {
1155 struct lwkt_gettoken_req req;
1159 req.cpu = mycpu->gd_cpuid;
1161 dcpu = (volatile int)tok->t_cpu;
1162 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1165 printf("REQT%d ", dcpu);
1167 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1168 lwkt_wait_ipiq(dcpu, seq);
1171 printf("REQR%d ", tok->t_cpu);
1181 lwkt_inittoken(lwkt_token_t tok)
1184 * Zero structure and set cpu owner and reqcpu to cpu 0.
1186 bzero(tok, sizeof(*tok));
1190 * Create a kernel process/thread/whatever. It shares it's address space
1191 * with proc0 - ie: kernel only.
1193 * NOTE! By default new threads are created with the MP lock held. A
1194 * thread which does not require the MP lock should release it by calling
1195 * rel_mplock() at the start of the new thread.
1198 lwkt_create(void (*func)(void *), void *arg,
1199 struct thread **tdp, thread_t template, int tdflags, int cpu,
1200 const char *fmt, ...)
1205 td = lwkt_alloc_thread(template, cpu);
1208 cpu_set_thread_handler(td, kthread_exit, func, arg);
1209 td->td_flags |= TDF_VERBOSE | tdflags;
1215 * Set up arg0 for 'ps' etc
1217 __va_start(ap, fmt);
1218 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1222 * Schedule the thread to run
1224 if ((td->td_flags & TDF_STOPREQ) == 0)
1227 td->td_flags &= ~TDF_STOPREQ;
1232 * Destroy an LWKT thread. Warning! This function is not called when
1233 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1234 * uses a different reaping mechanism.
1239 thread_t td = curthread;
1241 if (td->td_flags & TDF_VERBOSE)
1242 printf("kthread %p %s has exited\n", td, td->td_comm);
1244 lwkt_deschedule_self();
1245 ++mycpu->gd_tdfreecount;
1246 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1251 * Create a kernel process/thread/whatever. It shares it's address space
1252 * with proc0 - ie: kernel only. 5.x compatible.
1254 * NOTE! By default kthreads are created with the MP lock held. A
1255 * thread which does not require the MP lock should release it by calling
1256 * rel_mplock() at the start of the new thread.
1259 kthread_create(void (*func)(void *), void *arg,
1260 struct thread **tdp, const char *fmt, ...)
1265 td = lwkt_alloc_thread(NULL, -1);
1268 cpu_set_thread_handler(td, kthread_exit, func, arg);
1269 td->td_flags |= TDF_VERBOSE;
1275 * Set up arg0 for 'ps' etc
1277 __va_start(ap, fmt);
1278 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1282 * Schedule the thread to run
1291 thread_t td = curthread;
1292 int lpri = td->td_pri;
1295 panic("td_pri is/would-go negative! %p %d", td, lpri);
1299 * Destroy an LWKT thread. Warning! This function is not called when
1300 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1301 * uses a different reaping mechanism.
1303 * XXX duplicates lwkt_exit()
1314 * Send a function execution request to another cpu. The request is queued
1315 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1316 * possible target cpu. The FIFO can be written.
1318 * YYY If the FIFO fills up we have to enable interrupts and process the
1319 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1320 * Create a CPU_*() function to do this!
1322 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1323 * end will take care of any pending interrupts.
1325 * Must be called from a critical section.
1328 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1332 struct globaldata *gd = mycpu;
1334 if (dcpu == gd->gd_cpuid) {
1339 ++gd->gd_intr_nesting_level;
1341 if (gd->gd_intr_nesting_level > 20)
1342 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1344 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1345 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1347 ip = &gd->gd_ipiq[dcpu];
1350 * We always drain before the FIFO becomes full so it should never
1351 * become full. We need to leave enough entries to deal with
1354 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1355 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1356 ip->ip_func[windex] = func;
1357 ip->ip_arg[windex] = arg;
1358 /* YYY memory barrier */
1360 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1361 unsigned int eflags = read_eflags();
1364 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1365 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1366 lwkt_process_ipiq();
1368 write_eflags(eflags);
1370 --gd->gd_intr_nesting_level;
1371 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1373 return(ip->ip_windex);
1377 * Send a message to several target cpus. Typically used for scheduling.
1378 * The message will not be sent to stopped cpus.
1381 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1385 mask &= ~stopped_cpus;
1388 lwkt_send_ipiq(cpuid, func, arg);
1389 mask &= ~(1 << cpuid);
1394 * Wait for the remote cpu to finish processing a function.
1396 * YYY we have to enable interrupts and process the IPIQ while waiting
1397 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1398 * function to do this! YYY we really should 'block' here.
1400 * Must be called from a critical section. Thsi routine may be called
1401 * from an interrupt (for example, if an interrupt wakes a foreign thread
1405 lwkt_wait_ipiq(int dcpu, int seq)
1408 int maxc = 100000000;
1410 if (dcpu != mycpu->gd_cpuid) {
1411 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1412 ip = &mycpu->gd_ipiq[dcpu];
1413 if ((int)(ip->ip_xindex - seq) < 0) {
1414 unsigned int eflags = read_eflags();
1416 while ((int)(ip->ip_xindex - seq) < 0) {
1417 lwkt_process_ipiq();
1419 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1420 if (maxc < -1000000)
1421 panic("LWKT_WAIT_IPIQ");
1423 write_eflags(eflags);
1429 * Called from IPI interrupt (like a fast interrupt), which has placed
1430 * us in a critical section. The MP lock may or may not be held.
1431 * May also be called from doreti or splz, or be reentrantly called
1432 * indirectly through the ip_func[] we run.
1435 lwkt_process_ipiq(void)
1438 int cpuid = mycpu->gd_cpuid;
1440 for (n = 0; n < ncpus; ++n) {
1446 ip = globaldata_find(n)->gd_ipiq;
1452 * Note: xindex is only updated after we are sure the function has
1453 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1454 * function may send an IPI which may block/drain.
1456 while (ip->ip_rindex != ip->ip_windex) {
1457 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1459 ip->ip_func[ri](ip->ip_arg[ri]);
1460 /* YYY memory barrier */
1461 ip->ip_xindex = ip->ip_rindex;
1469 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1471 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1472 return(0); /* NOT REACHED */
1476 lwkt_wait_ipiq(int dcpu, int seq)
1478 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);