2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Each cpu in a system has its own self-contained light weight kernel
37 * thread scheduler, which means that generally speaking we only need
38 * to use a critical section to avoid problems. Foreign thread
39 * scheduling is queued via (async) IPIs.
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/kinfo.h>
48 #include <sys/queue.h>
49 #include <sys/sysctl.h>
50 #include <sys/kthread.h>
51 #include <machine/cpu.h>
54 #include <sys/spinlock.h>
57 #include <sys/thread2.h>
58 #include <sys/spinlock2.h>
59 #include <sys/mplock2.h>
61 #include <sys/dsched.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_pager.h>
70 #include <vm/vm_extern.h>
72 #include <machine/stdarg.h>
73 #include <machine/smp.h>
75 #if !defined(KTR_CTXSW)
76 #define KTR_CTXSW KTR_ALL
78 KTR_INFO_MASTER(ctxsw);
79 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
87 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
90 static int panic_on_cscount = 0;
92 static __int64_t switch_count = 0;
93 static __int64_t preempt_hit = 0;
94 static __int64_t preempt_miss = 0;
95 static __int64_t preempt_weird = 0;
96 static __int64_t token_contention_count __debugvar = 0;
97 static int lwkt_use_spin_port;
98 static struct objcache *thread_cache;
101 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
103 static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
105 extern void cpu_heavy_restore(void);
106 extern void cpu_lwkt_restore(void);
107 extern void cpu_kthread_restore(void);
108 extern void cpu_idle_restore(void);
113 jg_tos_ok(struct thread *td)
121 KKASSERT(td->td_sp != NULL);
122 tos = ((void **)td->td_sp)[0];
124 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
125 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
134 * We can make all thread ports use the spin backend instead of the thread
135 * backend. This should only be set to debug the spin backend.
137 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
140 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
142 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
143 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
144 "Successful preemption events");
145 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
146 "Failed preemption events");
147 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
149 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
150 &token_contention_count, 0, "spinning due to token contention");
152 static int fairq_enable = 1;
153 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
154 static int user_pri_sched = 0;
155 SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0, "");
158 * These helper procedures handle the runq, they can only be called from
159 * within a critical section.
161 * WARNING! Prior to SMP being brought up it is possible to enqueue and
162 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
163 * instead of 'mycpu' when referencing the globaldata structure. Once
164 * SMP live enqueuing and dequeueing only occurs on the current cpu.
168 _lwkt_dequeue(thread_t td)
170 if (td->td_flags & TDF_RUNQ) {
171 struct globaldata *gd = td->td_gd;
173 td->td_flags &= ~TDF_RUNQ;
174 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
175 gd->gd_fairq_total_pri -= td->td_pri;
176 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
177 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
184 * NOTE: There are a limited number of lwkt threads runnable since user
185 * processes only schedule one at a time per cpu.
189 _lwkt_enqueue(thread_t td)
193 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
194 struct globaldata *gd = td->td_gd;
196 td->td_flags |= TDF_RUNQ;
197 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
199 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
200 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
202 while (xtd && xtd->td_pri > td->td_pri)
203 xtd = TAILQ_NEXT(xtd, td_threadq);
205 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
207 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
209 gd->gd_fairq_total_pri += td->td_pri;
214 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
216 struct thread *td = (struct thread *)obj;
218 td->td_kstack = NULL;
219 td->td_kstack_size = 0;
220 td->td_flags = TDF_ALLOCATED_THREAD;
225 _lwkt_thread_dtor(void *obj, void *privdata)
227 struct thread *td = (struct thread *)obj;
229 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
230 ("_lwkt_thread_dtor: not allocated from objcache"));
231 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
232 td->td_kstack_size > 0,
233 ("_lwkt_thread_dtor: corrupted stack"));
234 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
238 * Initialize the lwkt s/system.
243 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
244 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
245 NULL, CACHE_NTHREADS/2,
246 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
250 * Schedule a thread to run. As the current thread we can always safely
251 * schedule ourselves, and a shortcut procedure is provided for that
254 * (non-blocking, self contained on a per cpu basis)
257 lwkt_schedule_self(thread_t td)
259 crit_enter_quick(td);
260 KASSERT(td != &td->td_gd->gd_idlethread,
261 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
262 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
268 * Deschedule a thread.
270 * (non-blocking, self contained on a per cpu basis)
273 lwkt_deschedule_self(thread_t td)
275 crit_enter_quick(td);
281 * LWKTs operate on a per-cpu basis
283 * WARNING! Called from early boot, 'mycpu' may not work yet.
286 lwkt_gdinit(struct globaldata *gd)
288 TAILQ_INIT(&gd->gd_tdrunq);
289 TAILQ_INIT(&gd->gd_tdallq);
293 * Create a new thread. The thread must be associated with a process context
294 * or LWKT start address before it can be scheduled. If the target cpu is
295 * -1 the thread will be created on the current cpu.
297 * If you intend to create a thread without a process context this function
298 * does everything except load the startup and switcher function.
301 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
303 globaldata_t gd = mycpu;
307 * If static thread storage is not supplied allocate a thread. Reuse
308 * a cached free thread if possible. gd_freetd is used to keep an exiting
309 * thread intact through the exit.
312 if ((td = gd->gd_freetd) != NULL)
313 gd->gd_freetd = NULL;
315 td = objcache_get(thread_cache, M_WAITOK);
316 KASSERT((td->td_flags &
317 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
318 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
319 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
323 * Try to reuse cached stack.
325 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
326 if (flags & TDF_ALLOCATED_STACK) {
327 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
332 stack = (void *)kmem_alloc(&kernel_map, stksize);
333 flags |= TDF_ALLOCATED_STACK;
336 lwkt_init_thread(td, stack, stksize, flags, gd);
338 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
343 * Initialize a preexisting thread structure. This function is used by
344 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
346 * All threads start out in a critical section at a priority of
347 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
348 * appropriate. This function may send an IPI message when the
349 * requested cpu is not the current cpu and consequently gd_tdallq may
350 * not be initialized synchronously from the point of view of the originating
353 * NOTE! we have to be careful in regards to creating threads for other cpus
354 * if SMP has not yet been activated.
359 lwkt_init_thread_remote(void *arg)
364 * Protected by critical section held by IPI dispatch
366 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
372 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
373 struct globaldata *gd)
375 globaldata_t mygd = mycpu;
377 bzero(td, sizeof(struct thread));
378 td->td_kstack = stack;
379 td->td_kstack_size = stksize;
380 td->td_flags = flags;
382 td->td_pri = TDPRI_KERN_DAEMON;
383 td->td_critcount = 1;
384 td->td_toks_stop = &td->td_toks_base;
386 if ((flags & TDF_MPSAFE) == 0)
389 if (lwkt_use_spin_port)
390 lwkt_initport_spin(&td->td_msgport);
392 lwkt_initport_thread(&td->td_msgport, td);
393 pmap_init_thread(td);
396 * Normally initializing a thread for a remote cpu requires sending an
397 * IPI. However, the idlethread is setup before the other cpus are
398 * activated so we have to treat it as a special case. XXX manipulation
399 * of gd_tdallq requires the BGL.
401 if (gd == mygd || td == &gd->gd_idlethread) {
403 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
406 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
410 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
414 dsched_new_thread(td);
418 lwkt_set_comm(thread_t td, const char *ctl, ...)
423 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
425 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
429 lwkt_hold(thread_t td)
435 lwkt_rele(thread_t td)
437 KKASSERT(td->td_refs > 0);
442 lwkt_wait_free(thread_t td)
445 tsleep(td, 0, "tdreap", hz);
449 lwkt_free_thread(thread_t td)
451 KASSERT((td->td_flags & TDF_RUNNING) == 0,
452 ("lwkt_free_thread: did not exit! %p", td));
454 if (td->td_flags & TDF_ALLOCATED_THREAD) {
455 objcache_put(thread_cache, td);
456 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
457 /* client-allocated struct with internally allocated stack */
458 KASSERT(td->td_kstack && td->td_kstack_size > 0,
459 ("lwkt_free_thread: corrupted stack"));
460 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
461 td->td_kstack = NULL;
462 td->td_kstack_size = 0;
464 KTR_LOG(ctxsw_deadtd, td);
469 * Switch to the next runnable lwkt. If no LWKTs are runnable then
470 * switch to the idlethread. Switching must occur within a critical
471 * section to avoid races with the scheduling queue.
473 * We always have full control over our cpu's run queue. Other cpus
474 * that wish to manipulate our queue must use the cpu_*msg() calls to
475 * talk to our cpu, so a critical section is all that is needed and
476 * the result is very, very fast thread switching.
478 * The LWKT scheduler uses a fixed priority model and round-robins at
479 * each priority level. User process scheduling is a totally
480 * different beast and LWKT priorities should not be confused with
481 * user process priorities.
483 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
484 * cleans it up. Note that the td_switch() function cannot do anything that
485 * requires the MP lock since the MP lock will have already been setup for
486 * the target thread (not the current thread). It's nice to have a scheduler
487 * that does not need the MP lock to work because it allows us to do some
488 * really cool high-performance MP lock optimizations.
490 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
491 * is not called by the current thread in the preemption case, only when
492 * the preempting thread blocks (in order to return to the original thread).
497 globaldata_t gd = mycpu;
498 thread_t td = gd->gd_curthread;
507 const char *lmsg; /* diagnostic - 'systat -pv 1' */
511 * Switching from within a 'fast' (non thread switched) interrupt or IPI
512 * is illegal. However, we may have to do it anyway if we hit a fatal
513 * kernel trap or we have paniced.
515 * If this case occurs save and restore the interrupt nesting level.
517 if (gd->gd_intr_nesting_level) {
521 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
522 panic("lwkt_switch: cannot switch from within "
523 "a fast interrupt, yet, td %p\n", td);
525 savegdnest = gd->gd_intr_nesting_level;
526 savegdtrap = gd->gd_trap_nesting_level;
527 gd->gd_intr_nesting_level = 0;
528 gd->gd_trap_nesting_level = 0;
529 if ((td->td_flags & TDF_PANICWARN) == 0) {
530 td->td_flags |= TDF_PANICWARN;
531 kprintf("Warning: thread switch from interrupt or IPI, "
532 "thread %p (%s)\n", td, td->td_comm);
536 gd->gd_intr_nesting_level = savegdnest;
537 gd->gd_trap_nesting_level = savegdtrap;
543 * Passive release (used to transition from user to kernel mode
544 * when we block or switch rather then when we enter the kernel).
545 * This function is NOT called if we are switching into a preemption
546 * or returning from a preemption. Typically this causes us to lose
547 * our current process designation (if we have one) and become a true
548 * LWKT thread, and may also hand the current process designation to
549 * another process and schedule thread.
555 if (TD_TOKS_HELD(td))
556 lwkt_relalltokens(td);
559 * We had better not be holding any spin locks, but don't get into an
560 * endless panic loop.
562 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
563 ("lwkt_switch: still holding a shared spinlock %p!",
564 gd->gd_spinlock_rd));
565 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
566 ("lwkt_switch: still holding %d exclusive spinlocks!",
567 gd->gd_spinlocks_wr));
572 * td_mpcount cannot be used to determine if we currently hold the
573 * MP lock because get_mplock() will increment it prior to attempting
574 * to get the lock, and switch out if it can't. Our ownership of
575 * the actual lock will remain stable while we are in a critical section
576 * (but, of course, another cpu may own or release the lock so the
577 * actual value of mp_lock is not stable).
579 mpheld = MP_LOCK_HELD(gd);
581 if (td->td_cscount) {
582 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
584 if (panic_on_cscount)
585 panic("switching while mastering cpusync");
591 * If we had preempted another thread on this cpu, resume the preempted
592 * thread. This occurs transparently, whether the preempted thread
593 * was scheduled or not (it may have been preempted after descheduling
596 * We have to setup the MP lock for the original thread after backing
597 * out the adjustment that was made to curthread when the original
600 if ((ntd = td->td_preempted) != NULL) {
601 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
603 if (ntd->td_mpcount && mpheld == 0) {
604 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
605 td, ntd, td->td_mpcount, ntd->td_mpcount);
607 if (ntd->td_mpcount) {
608 td->td_mpcount -= ntd->td_mpcount;
609 KKASSERT(td->td_mpcount >= 0);
612 ntd->td_flags |= TDF_PREEMPT_DONE;
615 * The interrupt may have woken a thread up, we need to properly
616 * set the reschedule flag if the originally interrupted thread is
617 * at a lower priority.
619 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
620 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
623 /* YYY release mp lock on switchback if original doesn't need it */
624 goto havethread_preempted;
628 * Implement round-robin fairq with priority insertion. The priority
629 * insertion is handled by _lwkt_enqueue()
631 * We have to adjust the MP lock for the target thread. If we
632 * need the MP lock and cannot obtain it we try to locate a
633 * thread that does not need the MP lock. If we cannot, we spin
636 * A similar issue exists for the tokens held by the target thread.
637 * If we cannot obtain ownership of the tokens we cannot immediately
638 * schedule the thread.
641 clear_lwkt_resched();
643 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
646 * Hotpath if we can get all necessary resources.
648 * If nothing is runnable switch to the idle thread
651 ntd = &gd->gd_idlethread;
652 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
653 ntd->td_flags |= TDF_IDLE_NOHLT;
655 if (ntd->td_mpcount) {
656 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
657 panic("Idle thread %p was holding the BGL!", ntd);
659 set_cpu_contention_mask(gd);
660 handle_cpu_contention_mask();
662 mpheld = MP_LOCK_HELD(gd);
667 clr_cpu_contention_mask(gd);
669 cpu_time.cp_msg[0] = 0;
670 cpu_time.cp_stallpc = 0;
677 * NOTE: For UP there is no mplock and lwkt_getalltokens()
680 if (ntd->td_fairq_accum >= 0 &&
682 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
684 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
687 clr_cpu_contention_mask(gd);
696 if (ntd->td_fairq_accum >= 0)
697 set_cpu_contention_mask(gd);
698 /* Reload mpheld (it become stale after mplock/token ops) */
699 mpheld = MP_LOCK_HELD(gd);
700 if (ntd->td_mpcount && mpheld == 0) {
702 laddr = ntd->td_mplock_stallpc;
707 * Coldpath - unable to schedule ntd, continue looking for threads
708 * to schedule. This is only allowed of the (presumably) kernel
709 * thread exhausted its fair share. A kernel thread stuck on
710 * resources does not currently allow a user thread to get in
714 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
715 (ntd->td_fairq_accum < 0));
723 * If the fair-share scheduler ran out ntd gets moved to the
724 * end and its accumulator will be bumped, if it didn't we
725 * maintain the same queue position.
727 * nlast keeps track of the last element prior to any moves.
729 if (ntd->td_fairq_accum < 0) {
730 lwkt_fairq_accumulate(gd, ntd);
736 xtd = TAILQ_NEXT(ntd, td_threadq);
737 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
738 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
741 * Set terminal element (nlast)
750 ntd = TAILQ_NEXT(ntd, td_threadq);
754 * If we exhausted the run list switch to the idle thread.
755 * Since one or more threads had resource acquisition issues
756 * we do not allow the idle thread to halt.
758 * NOTE: nlast can be NULL.
762 ntd = &gd->gd_idlethread;
763 ntd->td_flags |= TDF_IDLE_NOHLT;
765 if (ntd->td_mpcount) {
766 mpheld = MP_LOCK_HELD(gd);
767 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
768 panic("Idle thread %p was holding the BGL!", ntd);
770 set_cpu_contention_mask(gd);
771 handle_cpu_contention_mask();
773 mpheld = MP_LOCK_HELD(gd);
775 break; /* try again from the top, almost */
781 * If fairq accumulations occured we do not schedule the
782 * idle thread. This will cause us to try again from
786 break; /* try again from the top, almost */
788 strlcpy(cpu_time.cp_msg, lmsg, sizeof(cpu_time.cp_msg));
789 cpu_time.cp_stallpc = (uintptr_t)laddr;
794 * Try to switch to this thread.
796 * NOTE: For UP there is no mplock and lwkt_getalltokens()
799 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok ||
800 user_pri_sched) && ntd->td_fairq_accum >= 0 &&
802 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
804 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd, &lmsg, &laddr))
807 clr_cpu_contention_mask(gd);
812 if (ntd->td_fairq_accum >= 0)
813 set_cpu_contention_mask(gd);
815 * Reload mpheld (it become stale after mplock/token ops).
817 mpheld = MP_LOCK_HELD(gd);
818 if (ntd->td_mpcount && mpheld == 0) {
820 laddr = ntd->td_mplock_stallpc;
822 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
828 * All threads exhausted but we can loop due to a negative
831 * While we are looping in the scheduler be sure to service
832 * any interrupts which were made pending due to our critical
833 * section, otherwise we could livelock (e.g.) IPIs.
835 * NOTE: splz can enter and exit the mplock so mpheld is
836 * stale after this call.
842 * Our mplock can be cached and cause other cpus to livelock
843 * if we loop due to e.g. not being able to acquire tokens.
845 if (MP_LOCK_HELD(gd))
846 cpu_rel_mplock(gd->gd_cpuid);
852 * Do the actual switch. WARNING: mpheld is stale here.
854 * We must always decrement td_fairq_accum on non-idle threads just
855 * in case a thread never gets a tick due to being in a continuous
856 * critical section. The page-zeroing code does that.
858 * If the thread we came up with is a higher or equal priority verses
859 * the thread at the head of the queue we move our thread to the
860 * front. This way we can always check the front of the queue.
863 ++gd->gd_cnt.v_swtch;
864 --ntd->td_fairq_accum;
865 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
866 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
867 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
868 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
870 havethread_preempted:
873 * If the new target does not need the MP lock and we are holding it,
874 * release the MP lock. If the new target requires the MP lock we have
875 * already acquired it for the target.
877 * WARNING: mpheld is stale here.
880 KASSERT(ntd->td_critcount,
881 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
883 if (ntd->td_mpcount == 0 ) {
884 if (MP_LOCK_HELD(gd))
885 cpu_rel_mplock(gd->gd_cpuid);
887 ASSERT_MP_LOCK_HELD(ntd);
894 int tos_ok __debugvar = jg_tos_ok(ntd);
898 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
901 /* NOTE: current cpu may have changed after switch */
906 * Request that the target thread preempt the current thread. Preemption
907 * only works under a specific set of conditions:
909 * - We are not preempting ourselves
910 * - The target thread is owned by the current cpu
911 * - We are not currently being preempted
912 * - The target is not currently being preempted
913 * - We are not holding any spin locks
914 * - The target thread is not holding any tokens
915 * - We are able to satisfy the target's MP lock requirements (if any).
917 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
918 * this is called via lwkt_schedule() through the td_preemptable callback.
919 * critcount is the managed critical priority that we should ignore in order
920 * to determine whether preemption is possible (aka usually just the crit
921 * priority of lwkt_schedule() itself).
923 * XXX at the moment we run the target thread in a critical section during
924 * the preemption in order to prevent the target from taking interrupts
925 * that *WE* can't. Preemption is strictly limited to interrupt threads
926 * and interrupt-like threads, outside of a critical section, and the
927 * preempted source thread will be resumed the instant the target blocks
928 * whether or not the source is scheduled (i.e. preemption is supposed to
929 * be as transparent as possible).
931 * The target thread inherits our MP count (added to its own) for the
932 * duration of the preemption in order to preserve the atomicy of the
933 * MP lock during the preemption. Therefore, any preempting targets must be
934 * careful in regards to MP assertions. Note that the MP count may be
935 * out of sync with the physical mp_lock, but we do not have to preserve
936 * the original ownership of the lock if it was out of synch (that is, we
937 * can leave it synchronized on return).
940 lwkt_preempt(thread_t ntd, int critcount)
942 struct globaldata *gd = mycpu;
950 * The caller has put us in a critical section. We can only preempt
951 * if the caller of the caller was not in a critical section (basically
952 * a local interrupt), as determined by the 'critcount' parameter. We
953 * also can't preempt if the caller is holding any spinlocks (even if
954 * he isn't in a critical section). This also handles the tokens test.
956 * YYY The target thread must be in a critical section (else it must
957 * inherit our critical section? I dunno yet).
959 * Set need_lwkt_resched() unconditionally for now YYY.
961 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
963 td = gd->gd_curthread;
964 if (ntd->td_pri <= td->td_pri) {
968 if (td->td_critcount > critcount) {
974 if (ntd->td_gd != gd) {
981 * We don't have to check spinlocks here as they will also bump
984 * Do not try to preempt if the target thread is holding any tokens.
985 * We could try to acquire the tokens but this case is so rare there
986 * is no need to support it.
988 KKASSERT(gd->gd_spinlock_rd == NULL);
989 KKASSERT(gd->gd_spinlocks_wr == 0);
991 if (TD_TOKS_HELD(ntd)) {
996 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1001 if (ntd->td_preempted) {
1003 need_lwkt_resched();
1008 * note: an interrupt might have occured just as we were transitioning
1009 * to or from the MP lock. In this case td_mpcount will be pre-disposed
1010 * (non-zero) but not actually synchronized with the actual state of the
1011 * lock. We can use it to imply an MP lock requirement for the
1012 * preemption but we cannot use it to test whether we hold the MP lock
1015 savecnt = td->td_mpcount;
1016 mpheld = MP_LOCK_HELD(gd);
1017 ntd->td_mpcount += td->td_mpcount;
1018 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
1019 ntd->td_mpcount -= td->td_mpcount;
1021 need_lwkt_resched();
1027 * Since we are able to preempt the current thread, there is no need to
1028 * call need_lwkt_resched().
1031 ntd->td_preempted = td;
1032 td->td_flags |= TDF_PREEMPT_LOCK;
1033 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1036 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1038 KKASSERT(savecnt == td->td_mpcount);
1039 mpheld = MP_LOCK_HELD(gd);
1040 if (mpheld && td->td_mpcount == 0)
1041 cpu_rel_mplock(gd->gd_cpuid);
1042 else if (mpheld == 0 && td->td_mpcount)
1043 panic("lwkt_preempt(): MP lock was not held through");
1045 ntd->td_preempted = NULL;
1046 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1050 * Conditionally call splz() if gd_reqflags indicates work is pending.
1052 * td_nest_count prevents deep nesting via splz() or doreti() which
1053 * might otherwise blow out the kernel stack. Note that except for
1054 * this special case, we MUST call splz() here to handle any
1055 * pending ints, particularly after we switch, or we might accidently
1056 * halt the cpu with interrupts pending.
1058 * (self contained on a per cpu basis)
1063 globaldata_t gd = mycpu;
1064 thread_t td = gd->gd_curthread;
1066 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1071 * This function is used to negotiate a passive release of the current
1072 * process/lwp designation with the user scheduler, allowing the user
1073 * scheduler to schedule another user thread. The related kernel thread
1074 * (curthread) continues running in the released state.
1077 lwkt_passive_release(struct thread *td)
1079 struct lwp *lp = td->td_lwp;
1081 td->td_release = NULL;
1082 lwkt_setpri_self(TDPRI_KERN_USER);
1083 lp->lwp_proc->p_usched->release_curproc(lp);
1088 * This implements a normal yield. This routine is virtually a nop if
1089 * there is nothing to yield to but it will always run any pending interrupts
1090 * if called from a critical section.
1092 * This yield is designed for kernel threads without a user context.
1094 * (self contained on a per cpu basis)
1099 globaldata_t gd = mycpu;
1100 thread_t td = gd->gd_curthread;
1103 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1105 if (td->td_fairq_accum < 0) {
1106 lwkt_schedule_self(curthread);
1109 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1110 if (xtd && xtd->td_pri > td->td_pri) {
1111 lwkt_schedule_self(curthread);
1118 * This yield is designed for kernel threads with a user context.
1120 * The kernel acting on behalf of the user is potentially cpu-bound,
1121 * this function will efficiently allow other threads to run and also
1122 * switch to other processes by releasing.
1124 * The lwkt_user_yield() function is designed to have very low overhead
1125 * if no yield is determined to be needed.
1128 lwkt_user_yield(void)
1130 globaldata_t gd = mycpu;
1131 thread_t td = gd->gd_curthread;
1134 * Always run any pending interrupts in case we are in a critical
1137 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1142 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1143 * kernel can prevent other cpus from servicing interrupt threads
1144 * which still require the MP lock (which is a lot of them). This
1145 * has a chaining effect since if the interrupt is blocked, so is
1146 * the event, so normal scheduling will not pick up on the problem.
1148 if (cpu_contention_mask && td->td_mpcount) {
1154 * Switch (which forces a release) if another kernel thread needs
1155 * the cpu, if userland wants us to resched, or if our kernel
1156 * quantum has run out.
1158 if (lwkt_resched_wanted() ||
1159 user_resched_wanted() ||
1160 td->td_fairq_accum < 0)
1167 * Reacquire the current process if we are released.
1169 * XXX not implemented atm. The kernel may be holding locks and such,
1170 * so we want the thread to continue to receive cpu.
1172 if (td->td_release == NULL && lp) {
1173 lp->lwp_proc->p_usched->acquire_curproc(lp);
1174 td->td_release = lwkt_passive_release;
1175 lwkt_setpri_self(TDPRI_USER_NORM);
1181 * Generic schedule. Possibly schedule threads belonging to other cpus and
1182 * deal with threads that might be blocked on a wait queue.
1184 * We have a little helper inline function which does additional work after
1185 * the thread has been enqueued, including dealing with preemption and
1186 * setting need_lwkt_resched() (which prevents the kernel from returning
1187 * to userland until it has processed higher priority threads).
1189 * It is possible for this routine to be called after a failed _enqueue
1190 * (due to the target thread migrating, sleeping, or otherwise blocked).
1191 * We have to check that the thread is actually on the run queue!
1193 * reschedok is an optimized constant propagated from lwkt_schedule() or
1194 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1195 * reschedule to be requested if the target thread has a higher priority.
1196 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1197 * be 0, prevented undesired reschedules.
1201 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1205 if (ntd->td_flags & TDF_RUNQ) {
1206 if (ntd->td_preemptable && reschedok) {
1207 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1208 } else if (reschedok) {
1210 if (ntd->td_pri > otd->td_pri)
1211 need_lwkt_resched();
1215 * Give the thread a little fair share scheduler bump if it
1216 * has been asleep for a while. This is primarily to avoid
1217 * a degenerate case for interrupt threads where accumulator
1218 * crosses into negative territory unnecessarily.
1220 if (ntd->td_fairq_lticks != ticks) {
1221 ntd->td_fairq_lticks = ticks;
1222 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1223 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1224 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1231 _lwkt_schedule(thread_t td, int reschedok)
1233 globaldata_t mygd = mycpu;
1235 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1236 crit_enter_gd(mygd);
1237 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1238 if (td == mygd->gd_curthread) {
1242 * If we own the thread, there is no race (since we are in a
1243 * critical section). If we do not own the thread there might
1244 * be a race but the target cpu will deal with it.
1247 if (td->td_gd == mygd) {
1249 _lwkt_schedule_post(mygd, td, 1, reschedok);
1251 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1255 _lwkt_schedule_post(mygd, td, 1, reschedok);
1262 lwkt_schedule(thread_t td)
1264 _lwkt_schedule(td, 1);
1268 lwkt_schedule_noresched(thread_t td)
1270 _lwkt_schedule(td, 0);
1276 * When scheduled remotely if frame != NULL the IPIQ is being
1277 * run via doreti or an interrupt then preemption can be allowed.
1279 * To allow preemption we have to drop the critical section so only
1280 * one is present in _lwkt_schedule_post.
1283 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1285 thread_t td = curthread;
1288 if (frame && ntd->td_preemptable) {
1289 crit_exit_noyield(td);
1290 _lwkt_schedule(ntd, 1);
1291 crit_enter_quick(td);
1293 _lwkt_schedule(ntd, 1);
1298 * Thread migration using a 'Pull' method. The thread may or may not be
1299 * the current thread. It MUST be descheduled and in a stable state.
1300 * lwkt_giveaway() must be called on the cpu owning the thread.
1302 * At any point after lwkt_giveaway() is called, the target cpu may
1303 * 'pull' the thread by calling lwkt_acquire().
1305 * We have to make sure the thread is not sitting on a per-cpu tsleep
1306 * queue or it will blow up when it moves to another cpu.
1308 * MPSAFE - must be called under very specific conditions.
1311 lwkt_giveaway(thread_t td)
1313 globaldata_t gd = mycpu;
1316 if (td->td_flags & TDF_TSLEEPQ)
1318 KKASSERT(td->td_gd == gd);
1319 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1320 td->td_flags |= TDF_MIGRATING;
1325 lwkt_acquire(thread_t td)
1330 KKASSERT(td->td_flags & TDF_MIGRATING);
1335 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1336 crit_enter_gd(mygd);
1337 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1339 lwkt_process_ipiq();
1344 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1345 td->td_flags &= ~TDF_MIGRATING;
1348 crit_enter_gd(mygd);
1349 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1350 td->td_flags &= ~TDF_MIGRATING;
1358 * Generic deschedule. Descheduling threads other then your own should be
1359 * done only in carefully controlled circumstances. Descheduling is
1362 * This function may block if the cpu has run out of messages.
1365 lwkt_deschedule(thread_t td)
1369 if (td == curthread) {
1372 if (td->td_gd == mycpu) {
1375 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1385 * Set the target thread's priority. This routine does not automatically
1386 * switch to a higher priority thread, LWKT threads are not designed for
1387 * continuous priority changes. Yield if you want to switch.
1390 lwkt_setpri(thread_t td, int pri)
1392 KKASSERT(td->td_gd == mycpu);
1393 if (td->td_pri != pri) {
1396 if (td->td_flags & TDF_RUNQ) {
1408 * Set the initial priority for a thread prior to it being scheduled for
1409 * the first time. The thread MUST NOT be scheduled before or during
1410 * this call. The thread may be assigned to a cpu other then the current
1413 * Typically used after a thread has been created with TDF_STOPPREQ,
1414 * and before the thread is initially scheduled.
1417 lwkt_setpri_initial(thread_t td, int pri)
1420 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1425 lwkt_setpri_self(int pri)
1427 thread_t td = curthread;
1429 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1431 if (td->td_flags & TDF_RUNQ) {
1442 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1444 * Example: two competing threads, same priority N. decrement by (2*N)
1445 * increment by N*8, each thread will get 4 ticks.
1448 lwkt_fairq_schedulerclock(thread_t td)
1452 if (td != &td->td_gd->gd_idlethread) {
1453 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1454 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1455 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1456 if (td->td_fairq_accum < 0)
1457 need_lwkt_resched();
1458 td->td_fairq_lticks = ticks;
1460 td = td->td_preempted;
1466 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1468 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1469 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1470 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1474 * Migrate the current thread to the specified cpu.
1476 * This is accomplished by descheduling ourselves from the current cpu,
1477 * moving our thread to the tdallq of the target cpu, IPI messaging the
1478 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1479 * races while the thread is being migrated.
1481 * We must be sure to remove ourselves from the current cpu's tsleepq
1482 * before potentially moving to another queue. The thread can be on
1483 * a tsleepq due to a left-over tsleep_interlock().
1486 static void lwkt_setcpu_remote(void *arg);
1490 lwkt_setcpu_self(globaldata_t rgd)
1493 thread_t td = curthread;
1495 if (td->td_gd != rgd) {
1496 crit_enter_quick(td);
1497 if (td->td_flags & TDF_TSLEEPQ)
1499 td->td_flags |= TDF_MIGRATING;
1500 lwkt_deschedule_self(td);
1501 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1502 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1504 /* we are now on the target cpu */
1505 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1506 crit_exit_quick(td);
1512 lwkt_migratecpu(int cpuid)
1517 rgd = globaldata_find(cpuid);
1518 lwkt_setcpu_self(rgd);
1523 * Remote IPI for cpu migration (called while in a critical section so we
1524 * do not have to enter another one). The thread has already been moved to
1525 * our cpu's allq, but we must wait for the thread to be completely switched
1526 * out on the originating cpu before we schedule it on ours or the stack
1527 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1528 * change to main memory.
1530 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1531 * against wakeups. It is best if this interface is used only when there
1532 * are no pending events that might try to schedule the thread.
1536 lwkt_setcpu_remote(void *arg)
1539 globaldata_t gd = mycpu;
1541 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1543 lwkt_process_ipiq();
1549 td->td_flags &= ~TDF_MIGRATING;
1550 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1556 lwkt_preempted_proc(void)
1558 thread_t td = curthread;
1559 while (td->td_preempted)
1560 td = td->td_preempted;
1565 * Create a kernel process/thread/whatever. It shares it's address space
1566 * with proc0 - ie: kernel only.
1568 * NOTE! By default new threads are created with the MP lock held. A
1569 * thread which does not require the MP lock should release it by calling
1570 * rel_mplock() at the start of the new thread.
1573 lwkt_create(void (*func)(void *), void *arg,
1574 struct thread **tdp, thread_t template, int tdflags, int cpu,
1575 const char *fmt, ...)
1580 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1584 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1587 * Set up arg0 for 'ps' etc
1589 __va_start(ap, fmt);
1590 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1594 * Schedule the thread to run
1596 if ((td->td_flags & TDF_STOPREQ) == 0)
1599 td->td_flags &= ~TDF_STOPREQ;
1604 * Destroy an LWKT thread. Warning! This function is not called when
1605 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1606 * uses a different reaping mechanism.
1611 thread_t td = curthread;
1615 if (td->td_flags & TDF_VERBOSE)
1616 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1620 * Get us into a critical section to interlock gd_freetd and loop
1621 * until we can get it freed.
1623 * We have to cache the current td in gd_freetd because objcache_put()ing
1624 * it would rip it out from under us while our thread is still active.
1627 crit_enter_quick(td);
1628 while ((std = gd->gd_freetd) != NULL) {
1629 gd->gd_freetd = NULL;
1630 objcache_put(thread_cache, std);
1634 * Remove thread resources from kernel lists and deschedule us for
1637 if (td->td_flags & TDF_TSLEEPQ)
1640 dsched_exit_thread(td);
1641 lwkt_deschedule_self(td);
1642 lwkt_remove_tdallq(td);
1643 if (td->td_flags & TDF_ALLOCATED_THREAD)
1649 lwkt_remove_tdallq(thread_t td)
1651 KKASSERT(td->td_gd == mycpu);
1652 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1658 thread_t td = curthread;
1659 int lcrit = td->td_critcount;
1661 td->td_critcount = 0;
1662 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1668 * Called from debugger/panic on cpus which have been stopped. We must still
1669 * process the IPIQ while stopped, even if we were stopped while in a critical
1672 * If we are dumping also try to process any pending interrupts. This may
1673 * or may not work depending on the state of the cpu at the point it was
1677 lwkt_smp_stopped(void)
1679 globaldata_t gd = mycpu;
1683 lwkt_process_ipiq();
1686 lwkt_process_ipiq();