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 1.44 2003/11/24 23:56:07 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>
82 #include <machine/cpufunc.h>
86 static int untimely_switch = 0;
88 static int token_debug = 0;
90 static __int64_t switch_count = 0;
91 static __int64_t preempt_hit = 0;
92 static __int64_t preempt_miss = 0;
93 static __int64_t preempt_weird = 0;
94 static __int64_t ipiq_count = 0;
95 static __int64_t ipiq_fifofull = 0;
99 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
101 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
103 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
104 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
105 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
106 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
107 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
113 * These helper procedures handle the runq, they can only be called from
114 * within a critical section.
116 * WARNING! Prior to SMP being brought up it is possible to enqueue and
117 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
118 * instead of 'mycpu' when referencing the globaldata structure. Once
119 * SMP live enqueuing and dequeueing only occurs on the current cpu.
123 _lwkt_dequeue(thread_t td)
125 if (td->td_flags & TDF_RUNQ) {
126 int nq = td->td_pri & TDPRI_MASK;
127 struct globaldata *gd = td->td_gd;
129 td->td_flags &= ~TDF_RUNQ;
130 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
131 /* runqmask is passively cleaned up by the switcher */
137 _lwkt_enqueue(thread_t td)
139 if ((td->td_flags & TDF_RUNQ) == 0) {
140 int nq = td->td_pri & TDPRI_MASK;
141 struct globaldata *gd = td->td_gd;
143 td->td_flags |= TDF_RUNQ;
144 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
145 gd->gd_runqmask |= 1 << nq;
151 _lwkt_wantresched(thread_t ntd, thread_t cur)
153 return((ntd->td_pri & TDPRI_MASK) > (cur->td_pri & TDPRI_MASK));
157 * LWKTs operate on a per-cpu basis
159 * WARNING! Called from early boot, 'mycpu' may not work yet.
162 lwkt_gdinit(struct globaldata *gd)
166 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
167 TAILQ_INIT(&gd->gd_tdrunq[i]);
169 TAILQ_INIT(&gd->gd_tdallq);
173 * Initialize a thread wait structure prior to first use.
175 * NOTE! called from low level boot code, we cannot do anything fancy!
178 lwkt_init_wait(lwkt_wait_t w)
180 TAILQ_INIT(&w->wa_waitq);
184 * Create a new thread. The thread must be associated with a process context
185 * or LWKT start address before it can be scheduled. If the target cpu is
186 * -1 the thread will be created on the current cpu.
188 * If you intend to create a thread without a process context this function
189 * does everything except load the startup and switcher function.
192 lwkt_alloc_thread(struct thread *td, int cpu)
199 if (mycpu->gd_tdfreecount > 0) {
200 --mycpu->gd_tdfreecount;
201 td = TAILQ_FIRST(&mycpu->gd_tdfreeq);
202 KASSERT(td != NULL && (td->td_flags & TDF_RUNNING) == 0,
203 ("lwkt_alloc_thread: unexpected NULL or corrupted td"));
204 TAILQ_REMOVE(&mycpu->gd_tdfreeq, td, td_threadq);
206 stack = td->td_kstack;
207 flags = td->td_flags & (TDF_ALLOCATED_STACK|TDF_ALLOCATED_THREAD);
211 td = zalloc(thread_zone);
213 td = malloc(sizeof(struct thread));
215 td->td_kstack = NULL;
216 flags |= TDF_ALLOCATED_THREAD;
219 if ((stack = td->td_kstack) == NULL) {
221 stack = (void *)kmem_alloc(kernel_map, THREAD_STACK);
223 stack = libcaps_alloc_stack(THREAD_STACK);
225 flags |= TDF_ALLOCATED_STACK;
228 lwkt_init_thread(td, stack, flags, mycpu);
230 lwkt_init_thread(td, stack, flags, globaldata_find(cpu));
235 * Initialize a preexisting thread structure. This function is used by
236 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
238 * All threads start out in a critical section at a priority of
239 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
240 * appropriate. This function may send an IPI message when the
241 * requested cpu is not the current cpu and consequently gd_tdallq may
242 * not be initialized synchronously from the point of view of the originating
245 * NOTE! we have to be careful in regards to creating threads for other cpus
246 * if SMP has not yet been activated.
249 lwkt_init_thread_remote(void *arg)
253 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
257 lwkt_init_thread(thread_t td, void *stack, int flags, struct globaldata *gd)
259 bzero(td, sizeof(struct thread));
260 td->td_kstack = stack;
261 td->td_flags |= flags;
263 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
264 lwkt_initport(&td->td_msgport, td);
265 pmap_init_thread(td);
266 if (smp_active == 0 || gd == mycpu) {
268 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
271 lwkt_send_ipiq(gd->gd_cpuid, lwkt_init_thread_remote, td);
276 lwkt_set_comm(thread_t td, const char *ctl, ...)
281 vsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
286 lwkt_hold(thread_t td)
292 lwkt_rele(thread_t td)
294 KKASSERT(td->td_refs > 0);
301 lwkt_wait_free(thread_t td)
304 tsleep(td, 0, "tdreap", hz);
310 lwkt_free_thread(thread_t td)
312 struct globaldata *gd = mycpu;
314 KASSERT((td->td_flags & TDF_RUNNING) == 0,
315 ("lwkt_free_thread: did not exit! %p", td));
318 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
319 if (gd->gd_tdfreecount < CACHE_NTHREADS &&
320 (td->td_flags & TDF_ALLOCATED_THREAD)
322 ++gd->gd_tdfreecount;
323 TAILQ_INSERT_HEAD(&gd->gd_tdfreeq, td, td_threadq);
327 if (td->td_kstack && (td->td_flags & TDF_ALLOCATED_STACK)) {
329 kmem_free(kernel_map, (vm_offset_t)td->td_kstack, THREAD_STACK);
331 libcaps_free_stack(td->td_kstack, THREAD_STACK);
334 td->td_kstack = NULL;
336 if (td->td_flags & TDF_ALLOCATED_THREAD) {
338 zfree(thread_zone, td);
348 * Switch to the next runnable lwkt. If no LWKTs are runnable then
349 * switch to the idlethread. Switching must occur within a critical
350 * section to avoid races with the scheduling queue.
352 * We always have full control over our cpu's run queue. Other cpus
353 * that wish to manipulate our queue must use the cpu_*msg() calls to
354 * talk to our cpu, so a critical section is all that is needed and
355 * the result is very, very fast thread switching.
357 * The LWKT scheduler uses a fixed priority model and round-robins at
358 * each priority level. User process scheduling is a totally
359 * different beast and LWKT priorities should not be confused with
360 * user process priorities.
362 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
363 * cleans it up. Note that the td_switch() function cannot do anything that
364 * requires the MP lock since the MP lock will have already been setup for
365 * the target thread (not the current thread). It's nice to have a scheduler
366 * that does not need the MP lock to work because it allows us to do some
367 * really cool high-performance MP lock optimizations.
373 struct globaldata *gd;
374 thread_t td = curthread;
381 * Switching from within a 'fast' (non thread switched) interrupt is
384 if (mycpu->gd_intr_nesting_level && panicstr == NULL) {
385 panic("lwkt_switch: cannot switch from within a fast interrupt, yet\n");
389 * Passive release (used to transition from user to kernel mode
390 * when we block or switch rather then when we enter the kernel).
391 * This function is NOT called if we are switching into a preemption
392 * or returning from a preemption. Typically this causes us to lose
393 * our P_CURPROC designation (if we have one) and become a true LWKT
394 * thread, and may also hand P_CURPROC to another process and schedule
405 * td_mpcount cannot be used to determine if we currently hold the
406 * MP lock because get_mplock() will increment it prior to attempting
407 * to get the lock, and switch out if it can't. Our ownership of
408 * the actual lock will remain stable while we are in a critical section
409 * (but, of course, another cpu may own or release the lock so the
410 * actual value of mp_lock is not stable).
412 mpheld = MP_LOCK_HELD();
414 if ((ntd = td->td_preempted) != NULL) {
416 * We had preempted another thread on this cpu, resume the preempted
417 * thread. This occurs transparently, whether the preempted thread
418 * was scheduled or not (it may have been preempted after descheduling
421 * We have to setup the MP lock for the original thread after backing
422 * out the adjustment that was made to curthread when the original
425 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
427 if (ntd->td_mpcount && mpheld == 0) {
428 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d\n",
429 td, ntd, td->td_mpcount, ntd->td_mpcount);
431 if (ntd->td_mpcount) {
432 td->td_mpcount -= ntd->td_mpcount;
433 KKASSERT(td->td_mpcount >= 0);
436 ntd->td_flags |= TDF_PREEMPT_DONE;
437 /* YYY release mp lock on switchback if original doesn't need it */
440 * Priority queue / round-robin at each priority. Note that user
441 * processes run at a fixed, low priority and the user process
442 * scheduler deals with interactions between user processes
443 * by scheduling and descheduling them from the LWKT queue as
446 * We have to adjust the MP lock for the target thread. If we
447 * need the MP lock and cannot obtain it we try to locate a
448 * thread that does not need the MP lock.
452 if (gd->gd_runqmask) {
453 int nq = bsrl(gd->gd_runqmask);
454 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
455 gd->gd_runqmask &= ~(1 << nq);
459 if (ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) {
461 * Target needs MP lock and we couldn't get it, try
462 * to locate a thread which does not need the MP lock
463 * to run. If we cannot locate a thread spin in idle.
465 u_int32_t rqmask = gd->gd_runqmask;
467 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
468 if (ntd->td_mpcount == 0)
473 rqmask &= ~(1 << nq);
477 ntd = &gd->gd_idlethread;
478 ntd->td_flags |= TDF_IDLE_NOHLT;
480 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
481 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
484 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
485 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
488 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
489 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
493 * Nothing to run but we may still need the BGL to deal with
494 * pending interrupts, spin in idle if so.
496 ntd = &gd->gd_idlethread;
498 ntd->td_flags |= TDF_IDLE_NOHLT;
501 KASSERT(ntd->td_pri >= TDPRI_CRIT,
502 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
505 * Do the actual switch. If the new target does not need the MP lock
506 * and we are holding it, release the MP lock. If the new target requires
507 * the MP lock we have already acquired it for the target.
510 if (ntd->td_mpcount == 0 ) {
514 ASSERT_MP_LOCK_HELD();
525 * Switch if another thread has a higher priority. Do not switch to other
526 * threads at the same priority.
531 struct globaldata *gd = mycpu;
532 struct thread *td = gd->gd_curthread;
534 if ((td->td_pri & TDPRI_MASK) < bsrl(gd->gd_runqmask)) {
540 * Request that the target thread preempt the current thread. Preemption
541 * only works under a specific set of conditions:
543 * - We are not preempting ourselves
544 * - The target thread is owned by the current cpu
545 * - We are not currently being preempted
546 * - The target is not currently being preempted
547 * - We are able to satisfy the target's MP lock requirements (if any).
549 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
550 * this is called via lwkt_schedule() through the td_preemptable callback.
551 * critpri is the managed critical priority that we should ignore in order
552 * to determine whether preemption is possible (aka usually just the crit
553 * priority of lwkt_schedule() itself).
555 * XXX at the moment we run the target thread in a critical section during
556 * the preemption in order to prevent the target from taking interrupts
557 * that *WE* can't. Preemption is strictly limited to interrupt threads
558 * and interrupt-like threads, outside of a critical section, and the
559 * preempted source thread will be resumed the instant the target blocks
560 * whether or not the source is scheduled (i.e. preemption is supposed to
561 * be as transparent as possible).
563 * The target thread inherits our MP count (added to its own) for the
564 * duration of the preemption in order to preserve the atomicy of the
565 * MP lock during the preemption. Therefore, any preempting targets must be
566 * careful in regards to MP assertions. Note that the MP count may be
567 * out of sync with the physical mp_lock, but we do not have to preserve
568 * the original ownership of the lock if it was out of synch (that is, we
569 * can leave it synchronized on return).
572 lwkt_preempt(thread_t ntd, int critpri)
574 struct globaldata *gd = mycpu;
575 thread_t td = gd->gd_curthread;
582 * The caller has put us in a critical section. We can only preempt
583 * if the caller of the caller was not in a critical section (basically
584 * a local interrupt), as determined by the 'critpri' parameter. If
585 * we are unable to preempt
587 * YYY The target thread must be in a critical section (else it must
588 * inherit our critical section? I dunno yet).
590 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
593 if (!_lwkt_wantresched(ntd, td)) {
597 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
602 if (ntd->td_gd != gd) {
607 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
611 if (ntd->td_preempted) {
617 * note: an interrupt might have occured just as we were transitioning
618 * to or from the MP lock. In this case td_mpcount will be pre-disposed
619 * (non-zero) but not actually synchronized with the actual state of the
620 * lock. We can use it to imply an MP lock requirement for the
621 * preemption but we cannot use it to test whether we hold the MP lock
624 savecnt = td->td_mpcount;
625 mpheld = MP_LOCK_HELD();
626 ntd->td_mpcount += td->td_mpcount;
627 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
628 ntd->td_mpcount -= td->td_mpcount;
635 ntd->td_preempted = td;
636 td->td_flags |= TDF_PREEMPT_LOCK;
638 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
640 KKASSERT(savecnt == td->td_mpcount);
641 mpheld = MP_LOCK_HELD();
642 if (mpheld && td->td_mpcount == 0)
644 else if (mpheld == 0 && td->td_mpcount)
645 panic("lwkt_preempt(): MP lock was not held through");
647 ntd->td_preempted = NULL;
648 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
652 * Yield our thread while higher priority threads are pending. This is
653 * typically called when we leave a critical section but it can be safely
654 * called while we are in a critical section.
656 * This function will not generally yield to equal priority threads but it
657 * can occur as a side effect. Note that lwkt_switch() is called from
658 * inside the critical section to prevent its own crit_exit() from reentering
659 * lwkt_yield_quick().
661 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
662 * came along but was blocked and made pending.
664 * (self contained on a per cpu basis)
667 lwkt_yield_quick(void)
669 globaldata_t gd = mycpu;
670 thread_t td = gd->gd_curthread;
673 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
674 * it with a non-zero cpl then we might not wind up calling splz after
675 * a task switch when the critical section is exited even though the
676 * new task could accept the interrupt.
678 * XXX from crit_exit() only called after last crit section is released.
679 * If called directly will run splz() even if in a critical section.
681 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
682 * except for this special case, we MUST call splz() here to handle any
683 * pending ints, particularly after we switch, or we might accidently
684 * halt the cpu with interrupts pending.
686 if (gd->gd_reqflags && td->td_nest_count < 2)
690 * YYY enabling will cause wakeup() to task-switch, which really
691 * confused the old 4.x code. This is a good way to simulate
692 * preemption and MP without actually doing preemption or MP, because a
693 * lot of code assumes that wakeup() does not block.
695 if (untimely_switch && td->td_nest_count == 0 &&
696 gd->gd_intr_nesting_level == 0
700 * YYY temporary hacks until we disassociate the userland scheduler
701 * from the LWKT scheduler.
703 if (td->td_flags & TDF_RUNQ) {
704 lwkt_switch(); /* will not reenter yield function */
706 lwkt_schedule_self(); /* make sure we are scheduled */
707 lwkt_switch(); /* will not reenter yield function */
708 lwkt_deschedule_self(); /* make sure we are descheduled */
710 crit_exit_noyield(td);
715 * This implements a normal yield which, unlike _quick, will yield to equal
716 * priority threads as well. Note that gd_reqflags tests will be handled by
717 * the crit_exit() call in lwkt_switch().
719 * (self contained on a per cpu basis)
724 lwkt_schedule_self();
729 * Schedule a thread to run. As the current thread we can always safely
730 * schedule ourselves, and a shortcut procedure is provided for that
733 * (non-blocking, self contained on a per cpu basis)
736 lwkt_schedule_self(void)
738 thread_t td = curthread;
741 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
744 if (td->td_proc && td->td_proc->p_stat == SSLEEP)
745 panic("SCHED SELF PANIC");
751 * Generic schedule. Possibly schedule threads belonging to other cpus and
752 * deal with threads that might be blocked on a wait queue.
754 * YYY this is one of the best places to implement load balancing code.
755 * Load balancing can be accomplished by requesting other sorts of actions
756 * for the thread in question.
759 lwkt_schedule(thread_t td)
762 if ((td->td_flags & TDF_PREEMPT_LOCK) == 0 && td->td_proc
763 && td->td_proc->p_stat == SSLEEP
765 printf("PANIC schedule curtd = %p (%d %d) target %p (%d %d)\n",
767 curthread->td_proc ? curthread->td_proc->p_pid : -1,
768 curthread->td_proc ? curthread->td_proc->p_stat : -1,
770 td->td_proc ? curthread->td_proc->p_pid : -1,
771 td->td_proc ? curthread->td_proc->p_stat : -1
773 panic("SCHED PANIC");
777 if (td == curthread) {
783 * If the thread is on a wait list we have to send our scheduling
784 * request to the owner of the wait structure. Otherwise we send
785 * the scheduling request to the cpu owning the thread. Races
786 * are ok, the target will forward the message as necessary (the
787 * message may chase the thread around before it finally gets
790 * (remember, wait structures use stable storage)
792 if ((w = td->td_wait) != NULL) {
793 if (lwkt_trytoken(&w->wa_token)) {
794 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
797 if (smp_active == 0 || td->td_gd == mycpu) {
799 if (td->td_preemptable) {
800 td->td_preemptable(td, TDPRI_CRIT*2); /* YYY +token */
801 } else if (_lwkt_wantresched(td, curthread)) {
805 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
807 lwkt_reltoken(&w->wa_token);
809 lwkt_send_ipiq(w->wa_token.t_cpu, (ipifunc_t)lwkt_schedule, td);
813 * If the wait structure is NULL and we own the thread, there
814 * is no race (since we are in a critical section). If we
815 * do not own the thread there might be a race but the
816 * target cpu will deal with it.
818 if (smp_active == 0 || td->td_gd == mycpu) {
820 if (td->td_preemptable) {
821 td->td_preemptable(td, TDPRI_CRIT);
822 } else if (_lwkt_wantresched(td, curthread)) {
826 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
834 * Managed acquisition. This code assumes that the MP lock is held for
835 * the tdallq operation and that the thread has been descheduled from its
836 * original cpu. We also have to wait for the thread to be entirely switched
837 * out on its original cpu (this is usually fast enough that we never loop)
838 * since the LWKT system does not have to hold the MP lock while switching
839 * and the target may have released it before switching.
842 lwkt_acquire(thread_t td)
844 struct globaldata *gd;
847 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
848 while (td->td_flags & TDF_RUNNING) /* XXX spin */
852 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
855 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq); /* protected by BGL */
861 * Deschedule a thread.
863 * (non-blocking, self contained on a per cpu basis)
866 lwkt_deschedule_self(void)
868 thread_t td = curthread;
871 KASSERT(td->td_wait == NULL, ("lwkt_schedule_self(): td_wait not NULL!"));
877 * Generic deschedule. Descheduling threads other then your own should be
878 * done only in carefully controlled circumstances. Descheduling is
881 * This function may block if the cpu has run out of messages.
884 lwkt_deschedule(thread_t td)
887 if (td == curthread) {
890 if (td->td_gd == mycpu) {
893 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_deschedule, td);
900 * Set the target thread's priority. This routine does not automatically
901 * switch to a higher priority thread, LWKT threads are not designed for
902 * continuous priority changes. Yield if you want to switch.
904 * We have to retain the critical section count which uses the high bits
905 * of the td_pri field. The specified priority may also indicate zero or
906 * more critical sections by adding TDPRI_CRIT*N.
909 lwkt_setpri(thread_t td, int pri)
912 KKASSERT(td->td_gd == mycpu);
914 if (td->td_flags & TDF_RUNQ) {
916 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
919 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
925 lwkt_setpri_self(int pri)
927 thread_t td = curthread;
929 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
931 if (td->td_flags & TDF_RUNQ) {
933 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
936 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
942 lwkt_preempted_proc(void)
944 thread_t td = curthread;
945 while (td->td_preempted)
946 td = td->td_preempted;
950 typedef struct lwkt_gettoken_req {
958 * This function deschedules the current thread and blocks on the specified
959 * wait queue. We obtain ownership of the wait queue in order to block
960 * on it. A generation number is used to interlock the wait queue in case
961 * it gets signalled while we are blocked waiting on the token.
963 * Note: alternatively we could dequeue our thread and then message the
964 * target cpu owning the wait queue. YYY implement as sysctl.
966 * Note: wait queue signals normally ping-pong the cpu as an optimization.
970 lwkt_block(lwkt_wait_t w, const char *wmesg, int *gen)
972 thread_t td = curthread;
974 lwkt_gettoken(&w->wa_token);
975 if (w->wa_gen == *gen) {
977 TAILQ_INSERT_TAIL(&w->wa_waitq, td, td_threadq);
980 td->td_wmesg = wmesg;
983 lwkt_regettoken(&w->wa_token);
984 if (td->td_wmesg != NULL) {
989 /* token might be lost, doesn't matter for gen update */
991 lwkt_reltoken(&w->wa_token);
995 * Signal a wait queue. We gain ownership of the wait queue in order to
996 * signal it. Once a thread is removed from the wait queue we have to
997 * deal with the cpu owning the thread.
999 * Note: alternatively we could message the target cpu owning the wait
1000 * queue. YYY implement as sysctl.
1003 lwkt_signal(lwkt_wait_t w, int count)
1008 lwkt_gettoken(&w->wa_token);
1011 count = w->wa_count;
1012 while ((td = TAILQ_FIRST(&w->wa_waitq)) != NULL && count) {
1015 TAILQ_REMOVE(&w->wa_waitq, td, td_threadq);
1017 td->td_wmesg = NULL;
1018 if (td->td_gd == mycpu) {
1021 lwkt_send_ipiq(td->td_gd->gd_cpuid, (ipifunc_t)lwkt_schedule, td);
1023 lwkt_regettoken(&w->wa_token);
1025 lwkt_reltoken(&w->wa_token);
1031 * Acquire ownership of a token
1033 * Acquire ownership of a token. The token may have spl and/or critical
1034 * section side effects, depending on its purpose. These side effects
1035 * guarentee that you will maintain ownership of the token as long as you
1036 * do not block. If you block you may lose access to the token (but you
1037 * must still release it even if you lose your access to it).
1039 * YYY for now we use a critical section to prevent IPIs from taking away
1040 * a token, but do we really only need to disable IPIs ?
1042 * YYY certain tokens could be made to act like mutexes when performance
1043 * would be better (e.g. t_cpu == -1). This is not yet implemented.
1045 * YYY the tokens replace 4.x's simplelocks for the most part, but this
1046 * means that 4.x does not expect a switch so for now we cannot switch
1047 * when waiting for an IPI to be returned.
1049 * YYY If the token is owned by another cpu we may have to send an IPI to
1050 * it and then block. The IPI causes the token to be given away to the
1051 * requesting cpu, unless it has already changed hands. Since only the
1052 * current cpu can give away a token it owns we do not need a memory barrier.
1053 * This needs serious optimization.
1060 lwkt_gettoken_remote(void *arg)
1062 lwkt_gettoken_req *req = arg;
1063 if (req->tok->t_cpu == mycpu->gd_cpuid) {
1066 printf("GT(%d,%d) ", req->tok->t_cpu, req->cpu);
1068 req->tok->t_cpu = req->cpu;
1069 req->tok->t_reqcpu = req->cpu; /* YYY leave owned by target cpu */
1070 /* else set reqcpu to point to current cpu for release */
1077 lwkt_gettoken(lwkt_token_t tok)
1080 * Prevent preemption so the token can't be taken away from us once
1081 * we gain ownership of it. Use a synchronous request which might
1082 * block. The request will be forwarded as necessary playing catchup
1088 if (curthread->td_pri > 1800) {
1089 printf("lwkt_gettoken: %p called from %p: crit sect nesting warning\n",
1090 tok, ((int **)&tok)[-1]);
1092 if (curthread->td_pri > 2000) {
1093 curthread->td_pri = 1000;
1098 while (tok->t_cpu != mycpu->gd_cpuid) {
1099 struct lwkt_gettoken_req req;
1103 req.cpu = mycpu->gd_cpuid;
1105 dcpu = (volatile int)tok->t_cpu;
1106 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1109 printf("REQT%d ", dcpu);
1111 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1112 lwkt_wait_ipiq(dcpu, seq);
1115 printf("REQR%d ", tok->t_cpu);
1120 * leave us in a critical section on return. This will be undone
1121 * by lwkt_reltoken(). Bump the generation number.
1123 return(++tok->t_gen);
1127 * Attempt to acquire ownership of a token. Returns 1 on success, 0 on
1131 lwkt_trytoken(lwkt_token_t tok)
1135 if (tok->t_cpu != mycpu->gd_cpuid) {
1140 /* leave us in the critical section */
1146 * Release your ownership of a token. Releases must occur in reverse
1147 * order to aquisitions, eventually so priorities can be unwound properly
1148 * like SPLs. At the moment the actual implemention doesn't care.
1150 * We can safely hand a token that we own to another cpu without notifying
1151 * it, but once we do we can't get it back without requesting it (unless
1152 * the other cpu hands it back to us before we check).
1154 * We might have lost the token, so check that.
1156 * Return the token's generation number. The number is useful to callers
1157 * who may want to know if the token was stolen during potential blockages.
1160 lwkt_reltoken(lwkt_token_t tok)
1164 if (tok->t_cpu == mycpu->gd_cpuid) {
1165 tok->t_cpu = tok->t_reqcpu;
1173 * Reacquire a token that might have been lost. 0 is returned if the
1174 * generation has not changed (nobody stole the token from us), -1 is
1175 * returned otherwise. The token is reacquired regardless but the
1176 * generation number is not bumped further if we already own the token.
1178 * For efficiency we inline the best-case situation for lwkt_regettoken()
1179 * (i.e .we still own the token).
1182 lwkt_gentoken(lwkt_token_t tok, int *gen)
1184 if (tok->t_cpu == mycpu->gd_cpuid && tok->t_gen == *gen)
1186 *gen = lwkt_regettoken(tok);
1191 * Re-acquire a token that might have been lost. The generation number
1192 * is bumped and returned regardless of whether the token had been lost
1193 * or not (because we only have cpu granularity we have to bump the token
1197 lwkt_regettoken(lwkt_token_t tok)
1199 /* assert we are in a critical section */
1200 if (tok->t_cpu != mycpu->gd_cpuid) {
1202 while (tok->t_cpu != mycpu->gd_cpuid) {
1203 struct lwkt_gettoken_req req;
1207 req.cpu = mycpu->gd_cpuid;
1209 dcpu = (volatile int)tok->t_cpu;
1210 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1213 printf("REQT%d ", dcpu);
1215 seq = lwkt_send_ipiq(dcpu, lwkt_gettoken_remote, &req);
1216 lwkt_wait_ipiq(dcpu, seq);
1219 printf("REQR%d ", tok->t_cpu);
1229 lwkt_inittoken(lwkt_token_t tok)
1232 * Zero structure and set cpu owner and reqcpu to cpu 0.
1234 bzero(tok, sizeof(*tok));
1238 * Create a kernel process/thread/whatever. It shares it's address space
1239 * with proc0 - ie: kernel only.
1241 * NOTE! By default new threads are created with the MP lock held. A
1242 * thread which does not require the MP lock should release it by calling
1243 * rel_mplock() at the start of the new thread.
1246 lwkt_create(void (*func)(void *), void *arg,
1247 struct thread **tdp, thread_t template, int tdflags, int cpu,
1248 const char *fmt, ...)
1253 td = lwkt_alloc_thread(template, cpu);
1256 cpu_set_thread_handler(td, kthread_exit, func, arg);
1257 td->td_flags |= TDF_VERBOSE | tdflags;
1263 * Set up arg0 for 'ps' etc
1265 __va_start(ap, fmt);
1266 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1270 * Schedule the thread to run
1272 if ((td->td_flags & TDF_STOPREQ) == 0)
1275 td->td_flags &= ~TDF_STOPREQ;
1280 * Destroy an LWKT thread. Warning! This function is not called when
1281 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1282 * uses a different reaping mechanism.
1287 thread_t td = curthread;
1289 if (td->td_flags & TDF_VERBOSE)
1290 printf("kthread %p %s has exited\n", td, td->td_comm);
1292 lwkt_deschedule_self();
1293 ++mycpu->gd_tdfreecount;
1294 TAILQ_INSERT_TAIL(&mycpu->gd_tdfreeq, td, td_threadq);
1300 * Create a kernel process/thread/whatever. It shares it's address space
1301 * with proc0 - ie: kernel only. 5.x compatible.
1303 * NOTE! By default kthreads are created with the MP lock held. A
1304 * thread which does not require the MP lock should release it by calling
1305 * rel_mplock() at the start of the new thread.
1308 kthread_create(void (*func)(void *), void *arg,
1309 struct thread **tdp, const char *fmt, ...)
1314 td = lwkt_alloc_thread(NULL, -1);
1317 cpu_set_thread_handler(td, kthread_exit, func, arg);
1318 td->td_flags |= TDF_VERBOSE;
1324 * Set up arg0 for 'ps' etc
1326 __va_start(ap, fmt);
1327 vsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1331 * Schedule the thread to run
1342 thread_t td = curthread;
1343 int lpri = td->td_pri;
1346 panic("td_pri is/would-go negative! %p %d", td, lpri);
1350 * Destroy an LWKT thread. Warning! This function is not called when
1351 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1352 * uses a different reaping mechanism.
1354 * XXX duplicates lwkt_exit()
1365 * Send a function execution request to another cpu. The request is queued
1366 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
1367 * possible target cpu. The FIFO can be written.
1369 * YYY If the FIFO fills up we have to enable interrupts and process the
1370 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
1371 * Create a CPU_*() function to do this!
1373 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
1374 * end will take care of any pending interrupts.
1376 * Must be called from a critical section.
1379 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1383 struct globaldata *gd = mycpu;
1385 if (dcpu == gd->gd_cpuid) {
1390 ++gd->gd_intr_nesting_level;
1392 if (gd->gd_intr_nesting_level > 20)
1393 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
1395 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
1396 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1398 ip = &gd->gd_ipiq[dcpu];
1401 * We always drain before the FIFO becomes full so it should never
1402 * become full. We need to leave enough entries to deal with
1405 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
1406 windex = ip->ip_windex & MAXCPUFIFO_MASK;
1407 ip->ip_func[windex] = func;
1408 ip->ip_arg[windex] = arg;
1409 /* YYY memory barrier */
1411 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
1412 unsigned int eflags = read_eflags();
1415 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
1416 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
1417 lwkt_process_ipiq();
1419 write_eflags(eflags);
1421 --gd->gd_intr_nesting_level;
1422 cpu_send_ipiq(dcpu); /* issues memory barrier if appropriate */
1424 return(ip->ip_windex);
1428 * Send a message to several target cpus. Typically used for scheduling.
1429 * The message will not be sent to stopped cpus.
1432 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
1436 mask &= ~stopped_cpus;
1439 lwkt_send_ipiq(cpuid, func, arg);
1440 mask &= ~(1 << cpuid);
1445 * Wait for the remote cpu to finish processing a function.
1447 * YYY we have to enable interrupts and process the IPIQ while waiting
1448 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
1449 * function to do this! YYY we really should 'block' here.
1451 * Must be called from a critical section. Thsi routine may be called
1452 * from an interrupt (for example, if an interrupt wakes a foreign thread
1456 lwkt_wait_ipiq(int dcpu, int seq)
1459 int maxc = 100000000;
1461 if (dcpu != mycpu->gd_cpuid) {
1462 KKASSERT(dcpu >= 0 && dcpu < ncpus);
1463 ip = &mycpu->gd_ipiq[dcpu];
1464 if ((int)(ip->ip_xindex - seq) < 0) {
1465 unsigned int eflags = read_eflags();
1467 while ((int)(ip->ip_xindex - seq) < 0) {
1468 lwkt_process_ipiq();
1470 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, dcpu, ip->ip_xindex - seq);
1471 if (maxc < -1000000)
1472 panic("LWKT_WAIT_IPIQ");
1474 write_eflags(eflags);
1480 * Called from IPI interrupt (like a fast interrupt), which has placed
1481 * us in a critical section. The MP lock may or may not be held.
1482 * May also be called from doreti or splz, or be reentrantly called
1483 * indirectly through the ip_func[] we run.
1486 lwkt_process_ipiq(void)
1489 int cpuid = mycpu->gd_cpuid;
1491 for (n = 0; n < ncpus; ++n) {
1497 ip = globaldata_find(n)->gd_ipiq;
1503 * Note: xindex is only updated after we are sure the function has
1504 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
1505 * function may send an IPI which may block/drain.
1507 while (ip->ip_rindex != ip->ip_windex) {
1508 ri = ip->ip_rindex & MAXCPUFIFO_MASK;
1510 ip->ip_func[ri](ip->ip_arg[ri]);
1511 /* YYY memory barrier */
1512 ip->ip_xindex = ip->ip_rindex;
1520 lwkt_send_ipiq(int dcpu, ipifunc_t func, void *arg)
1522 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", dcpu, func, arg);
1523 return(0); /* NOT REACHED */
1527 lwkt_wait_ipiq(int dcpu, int seq)
1529 panic("lwkt_wait_ipiq: UP box! (%d,%d)", dcpu, seq);