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/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <machine/cpu.h>
53 #include <sys/spinlock.h>
56 #include <sys/thread2.h>
57 #include <sys/spinlock2.h>
58 #include <sys/mplock2.h>
60 #include <sys/dsched.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_object.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_pager.h>
69 #include <vm/vm_extern.h>
71 #include <machine/stdarg.h>
72 #include <machine/smp.h>
74 #if !defined(KTR_CTXSW)
75 #define KTR_CTXSW KTR_ALL
77 KTR_INFO_MASTER(ctxsw);
78 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
79 sizeof(int) + sizeof(struct thread *));
80 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
81 sizeof(int) + sizeof(struct thread *));
82 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
83 sizeof (struct thread *) + sizeof(char *));
84 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
86 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
89 static int panic_on_cscount = 0;
91 static __int64_t switch_count = 0;
92 static __int64_t preempt_hit = 0;
93 static __int64_t preempt_miss = 0;
94 static __int64_t preempt_weird = 0;
95 static __int64_t token_contention_count __debugvar = 0;
96 static int lwkt_use_spin_port;
97 static struct objcache *thread_cache;
100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
102 static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
104 extern void cpu_heavy_restore(void);
105 extern void cpu_lwkt_restore(void);
106 extern void cpu_kthread_restore(void);
107 extern void cpu_idle_restore(void);
112 jg_tos_ok(struct thread *td)
120 KKASSERT(td->td_sp != NULL);
121 tos = ((void **)td->td_sp)[0];
123 if ((tos == cpu_heavy_restore) || (tos == cpu_lwkt_restore) ||
124 (tos == cpu_kthread_restore) || (tos == cpu_idle_restore)) {
133 * We can make all thread ports use the spin backend instead of the thread
134 * backend. This should only be set to debug the spin backend.
136 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
139 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
141 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
142 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
143 "Successful preemption events");
144 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
145 "Failed preemption events");
146 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
148 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
149 &token_contention_count, 0, "spinning due to token contention");
151 static int fairq_enable = 1;
152 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0, "");
155 * These helper procedures handle the runq, they can only be called from
156 * within a critical section.
158 * WARNING! Prior to SMP being brought up it is possible to enqueue and
159 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
160 * instead of 'mycpu' when referencing the globaldata structure. Once
161 * SMP live enqueuing and dequeueing only occurs on the current cpu.
165 _lwkt_dequeue(thread_t td)
167 if (td->td_flags & TDF_RUNQ) {
168 struct globaldata *gd = td->td_gd;
170 td->td_flags &= ~TDF_RUNQ;
171 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
172 gd->gd_fairq_total_pri -= td->td_pri;
173 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
174 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
181 * NOTE: There are a limited number of lwkt threads runnable since user
182 * processes only schedule one at a time per cpu.
186 _lwkt_enqueue(thread_t td)
190 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
191 struct globaldata *gd = td->td_gd;
193 td->td_flags |= TDF_RUNQ;
194 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
196 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
197 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
199 while (xtd && xtd->td_pri > td->td_pri)
200 xtd = TAILQ_NEXT(xtd, td_threadq);
202 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
204 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
206 gd->gd_fairq_total_pri += td->td_pri;
211 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
213 struct thread *td = (struct thread *)obj;
215 td->td_kstack = NULL;
216 td->td_kstack_size = 0;
217 td->td_flags = TDF_ALLOCATED_THREAD;
222 _lwkt_thread_dtor(void *obj, void *privdata)
224 struct thread *td = (struct thread *)obj;
226 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
227 ("_lwkt_thread_dtor: not allocated from objcache"));
228 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
229 td->td_kstack_size > 0,
230 ("_lwkt_thread_dtor: corrupted stack"));
231 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
235 * Initialize the lwkt s/system.
240 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
241 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
242 NULL, CACHE_NTHREADS/2,
243 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
247 * Schedule a thread to run. As the current thread we can always safely
248 * schedule ourselves, and a shortcut procedure is provided for that
251 * (non-blocking, self contained on a per cpu basis)
254 lwkt_schedule_self(thread_t td)
256 crit_enter_quick(td);
257 KASSERT(td != &td->td_gd->gd_idlethread,
258 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
259 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
265 * Deschedule a thread.
267 * (non-blocking, self contained on a per cpu basis)
270 lwkt_deschedule_self(thread_t td)
272 crit_enter_quick(td);
278 * LWKTs operate on a per-cpu basis
280 * WARNING! Called from early boot, 'mycpu' may not work yet.
283 lwkt_gdinit(struct globaldata *gd)
285 TAILQ_INIT(&gd->gd_tdrunq);
286 TAILQ_INIT(&gd->gd_tdallq);
290 * Create a new thread. The thread must be associated with a process context
291 * or LWKT start address before it can be scheduled. If the target cpu is
292 * -1 the thread will be created on the current cpu.
294 * If you intend to create a thread without a process context this function
295 * does everything except load the startup and switcher function.
298 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
300 globaldata_t gd = mycpu;
304 * If static thread storage is not supplied allocate a thread. Reuse
305 * a cached free thread if possible. gd_freetd is used to keep an exiting
306 * thread intact through the exit.
309 if ((td = gd->gd_freetd) != NULL)
310 gd->gd_freetd = NULL;
312 td = objcache_get(thread_cache, M_WAITOK);
313 KASSERT((td->td_flags &
314 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
315 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
316 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
320 * Try to reuse cached stack.
322 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
323 if (flags & TDF_ALLOCATED_STACK) {
324 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
329 stack = (void *)kmem_alloc(&kernel_map, stksize);
330 flags |= TDF_ALLOCATED_STACK;
333 lwkt_init_thread(td, stack, stksize, flags, gd);
335 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
340 * Initialize a preexisting thread structure. This function is used by
341 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
343 * All threads start out in a critical section at a priority of
344 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
345 * appropriate. This function may send an IPI message when the
346 * requested cpu is not the current cpu and consequently gd_tdallq may
347 * not be initialized synchronously from the point of view of the originating
350 * NOTE! we have to be careful in regards to creating threads for other cpus
351 * if SMP has not yet been activated.
356 lwkt_init_thread_remote(void *arg)
361 * Protected by critical section held by IPI dispatch
363 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
369 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
370 struct globaldata *gd)
372 globaldata_t mygd = mycpu;
374 bzero(td, sizeof(struct thread));
375 td->td_kstack = stack;
376 td->td_kstack_size = stksize;
377 td->td_flags = flags;
379 td->td_pri = TDPRI_KERN_DAEMON;
380 td->td_critcount = 1;
381 td->td_toks_stop = &td->td_toks_base;
383 if ((flags & TDF_MPSAFE) == 0)
386 if (lwkt_use_spin_port)
387 lwkt_initport_spin(&td->td_msgport);
389 lwkt_initport_thread(&td->td_msgport, td);
390 pmap_init_thread(td);
393 * Normally initializing a thread for a remote cpu requires sending an
394 * IPI. However, the idlethread is setup before the other cpus are
395 * activated so we have to treat it as a special case. XXX manipulation
396 * of gd_tdallq requires the BGL.
398 if (gd == mygd || td == &gd->gd_idlethread) {
400 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
403 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
407 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
411 dsched_new_thread(td);
415 lwkt_set_comm(thread_t td, const char *ctl, ...)
420 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
422 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
426 lwkt_hold(thread_t td)
432 lwkt_rele(thread_t td)
434 KKASSERT(td->td_refs > 0);
439 lwkt_wait_free(thread_t td)
442 tsleep(td, 0, "tdreap", hz);
446 lwkt_free_thread(thread_t td)
448 KASSERT((td->td_flags & TDF_RUNNING) == 0,
449 ("lwkt_free_thread: did not exit! %p", td));
451 if (td->td_flags & TDF_ALLOCATED_THREAD) {
452 objcache_put(thread_cache, td);
453 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
454 /* client-allocated struct with internally allocated stack */
455 KASSERT(td->td_kstack && td->td_kstack_size > 0,
456 ("lwkt_free_thread: corrupted stack"));
457 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
458 td->td_kstack = NULL;
459 td->td_kstack_size = 0;
461 KTR_LOG(ctxsw_deadtd, td);
466 * Switch to the next runnable lwkt. If no LWKTs are runnable then
467 * switch to the idlethread. Switching must occur within a critical
468 * section to avoid races with the scheduling queue.
470 * We always have full control over our cpu's run queue. Other cpus
471 * that wish to manipulate our queue must use the cpu_*msg() calls to
472 * talk to our cpu, so a critical section is all that is needed and
473 * the result is very, very fast thread switching.
475 * The LWKT scheduler uses a fixed priority model and round-robins at
476 * each priority level. User process scheduling is a totally
477 * different beast and LWKT priorities should not be confused with
478 * user process priorities.
480 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
481 * cleans it up. Note that the td_switch() function cannot do anything that
482 * requires the MP lock since the MP lock will have already been setup for
483 * the target thread (not the current thread). It's nice to have a scheduler
484 * that does not need the MP lock to work because it allows us to do some
485 * really cool high-performance MP lock optimizations.
487 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
488 * is not called by the current thread in the preemption case, only when
489 * the preempting thread blocks (in order to return to the original thread).
494 globaldata_t gd = mycpu;
495 thread_t td = gd->gd_curthread;
506 * Switching from within a 'fast' (non thread switched) interrupt or IPI
507 * is illegal. However, we may have to do it anyway if we hit a fatal
508 * kernel trap or we have paniced.
510 * If this case occurs save and restore the interrupt nesting level.
512 if (gd->gd_intr_nesting_level) {
516 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
517 panic("lwkt_switch: cannot switch from within "
518 "a fast interrupt, yet, td %p\n", td);
520 savegdnest = gd->gd_intr_nesting_level;
521 savegdtrap = gd->gd_trap_nesting_level;
522 gd->gd_intr_nesting_level = 0;
523 gd->gd_trap_nesting_level = 0;
524 if ((td->td_flags & TDF_PANICWARN) == 0) {
525 td->td_flags |= TDF_PANICWARN;
526 kprintf("Warning: thread switch from interrupt or IPI, "
527 "thread %p (%s)\n", td, td->td_comm);
531 gd->gd_intr_nesting_level = savegdnest;
532 gd->gd_trap_nesting_level = savegdtrap;
538 * Passive release (used to transition from user to kernel mode
539 * when we block or switch rather then when we enter the kernel).
540 * This function is NOT called if we are switching into a preemption
541 * or returning from a preemption. Typically this causes us to lose
542 * our current process designation (if we have one) and become a true
543 * LWKT thread, and may also hand the current process designation to
544 * another process and schedule thread.
550 if (TD_TOKS_HELD(td))
551 lwkt_relalltokens(td);
554 * We had better not be holding any spin locks, but don't get into an
555 * endless panic loop.
557 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
558 ("lwkt_switch: still holding a shared spinlock %p!",
559 gd->gd_spinlock_rd));
560 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
561 ("lwkt_switch: still holding %d exclusive spinlocks!",
562 gd->gd_spinlocks_wr));
567 * td_mpcount cannot be used to determine if we currently hold the
568 * MP lock because get_mplock() will increment it prior to attempting
569 * to get the lock, and switch out if it can't. Our ownership of
570 * the actual lock will remain stable while we are in a critical section
571 * (but, of course, another cpu may own or release the lock so the
572 * actual value of mp_lock is not stable).
574 mpheld = MP_LOCK_HELD();
576 if (td->td_cscount) {
577 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
579 if (panic_on_cscount)
580 panic("switching while mastering cpusync");
586 * If we had preempted another thread on this cpu, resume the preempted
587 * thread. This occurs transparently, whether the preempted thread
588 * was scheduled or not (it may have been preempted after descheduling
591 * We have to setup the MP lock for the original thread after backing
592 * out the adjustment that was made to curthread when the original
595 if ((ntd = td->td_preempted) != NULL) {
596 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
598 if (ntd->td_mpcount && mpheld == 0) {
599 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
600 td, ntd, td->td_mpcount, ntd->td_mpcount);
602 if (ntd->td_mpcount) {
603 td->td_mpcount -= ntd->td_mpcount;
604 KKASSERT(td->td_mpcount >= 0);
607 ntd->td_flags |= TDF_PREEMPT_DONE;
610 * The interrupt may have woken a thread up, we need to properly
611 * set the reschedule flag if the originally interrupted thread is
612 * at a lower priority.
614 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
615 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
618 /* YYY release mp lock on switchback if original doesn't need it */
619 goto havethread_preempted;
623 * Implement round-robin fairq with priority insertion. The priority
624 * insertion is handled by _lwkt_enqueue()
626 * We have to adjust the MP lock for the target thread. If we
627 * need the MP lock and cannot obtain it we try to locate a
628 * thread that does not need the MP lock. If we cannot, we spin
631 * A similar issue exists for the tokens held by the target thread.
632 * If we cannot obtain ownership of the tokens we cannot immediately
633 * schedule the thread.
636 clear_lwkt_resched();
638 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
641 * Hotpath if we can get all necessary resources.
643 * If nothing is runnable switch to the idle thread
646 ntd = &gd->gd_idlethread;
647 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
648 ntd->td_flags |= TDF_IDLE_NOHLT;
649 if (ntd->td_mpcount) {
650 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
651 panic("Idle thread %p was holding the BGL!", ntd);
663 if (ntd->td_fairq_accum >= 0 &&
665 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
667 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd))
670 clr_mplock_contention_mask(gd);
676 /* Reload mpheld (it become stale after mplock/token ops) */
677 mpheld = MP_LOCK_HELD();
681 * Coldpath - unable to schedule ntd, continue looking for threads
682 * to schedule. This is only allowed of the (presumably) kernel
683 * thread exhausted its fair share. A kernel thread stuck on
684 * resources does not currently allow a user thread to get in
688 nquserok = ((ntd->td_pri < TDPRI_KERN_LPSCHED) ||
689 (ntd->td_fairq_accum < 0));
695 * If the fair-share scheduler ran out ntd gets moved to the
696 * end and its accumulator will be bumped, if it didn't we
697 * maintain the same queue position.
699 * nlast keeps track of the last element prior to any moves.
701 if (ntd->td_fairq_accum < 0) {
702 xtd = TAILQ_NEXT(ntd, td_threadq);
703 lwkt_fairq_accumulate(gd, ntd);
705 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
706 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
714 ntd = TAILQ_NEXT(ntd, td_threadq);
718 * If we exhausted the run list switch to the idle thread.
719 * Since one or more threads had resource acquisition issues
720 * we do not allow the idle thread to halt.
722 * NOTE: nlast can be NULL.
726 ntd = &gd->gd_idlethread;
727 ntd->td_flags |= TDF_IDLE_NOHLT;
728 set_mplock_contention_mask(gd);
729 cpu_mplock_contested();
730 if (ntd->td_mpcount) {
731 mpheld = MP_LOCK_HELD();
732 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
733 panic("Idle thread %p was holding the BGL!", ntd);
736 break; /* try again from the top, almost */
741 * If fairq accumulations occured we do not schedule the
742 * idle thread. This will cause us to try again from
751 * Try to switch to this thread.
753 if ((ntd->td_pri >= TDPRI_KERN_LPSCHED || nquserok) &&
754 ntd->td_fairq_accum >= 0 &&
756 (ntd->td_mpcount == 0 || mpheld || cpu_try_mplock()) &&
758 (!TD_TOKS_HELD(ntd) || lwkt_getalltokens(ntd))
761 clr_mplock_contention_mask(gd);
766 /* Reload mpheld (it become stale after mplock/token ops) */
767 mpheld = MP_LOCK_HELD();
768 if (ntd->td_pri >= TDPRI_KERN_LPSCHED && ntd->td_fairq_accum >= 0)
775 * Do the actual switch. WARNING: mpheld is stale here.
777 * We must always decrement td_fairq_accum on non-idle threads just
778 * in case a thread never gets a tick due to being in a continuous
779 * critical section. The page-zeroing code does that.
781 * If the thread we came up with is a higher or equal priority verses
782 * the thread at the head of the queue we move our thread to the
783 * front. This way we can always check the front of the queue.
786 ++gd->gd_cnt.v_swtch;
787 --ntd->td_fairq_accum;
788 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
789 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
790 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
791 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
793 havethread_preempted:
796 * If the new target does not need the MP lock and we are holding it,
797 * release the MP lock. If the new target requires the MP lock we have
798 * already acquired it for the target.
800 * WARNING: mpheld is stale here.
803 KASSERT(ntd->td_critcount,
804 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
806 if (ntd->td_mpcount == 0 ) {
810 ASSERT_MP_LOCK_HELD(ntd);
817 int tos_ok __debugvar = jg_tos_ok(ntd);
821 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
824 /* NOTE: current cpu may have changed after switch */
829 * Request that the target thread preempt the current thread. Preemption
830 * only works under a specific set of conditions:
832 * - We are not preempting ourselves
833 * - The target thread is owned by the current cpu
834 * - We are not currently being preempted
835 * - The target is not currently being preempted
836 * - We are not holding any spin locks
837 * - The target thread is not holding any tokens
838 * - We are able to satisfy the target's MP lock requirements (if any).
840 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
841 * this is called via lwkt_schedule() through the td_preemptable callback.
842 * critcount is the managed critical priority that we should ignore in order
843 * to determine whether preemption is possible (aka usually just the crit
844 * priority of lwkt_schedule() itself).
846 * XXX at the moment we run the target thread in a critical section during
847 * the preemption in order to prevent the target from taking interrupts
848 * that *WE* can't. Preemption is strictly limited to interrupt threads
849 * and interrupt-like threads, outside of a critical section, and the
850 * preempted source thread will be resumed the instant the target blocks
851 * whether or not the source is scheduled (i.e. preemption is supposed to
852 * be as transparent as possible).
854 * The target thread inherits our MP count (added to its own) for the
855 * duration of the preemption in order to preserve the atomicy of the
856 * MP lock during the preemption. Therefore, any preempting targets must be
857 * careful in regards to MP assertions. Note that the MP count may be
858 * out of sync with the physical mp_lock, but we do not have to preserve
859 * the original ownership of the lock if it was out of synch (that is, we
860 * can leave it synchronized on return).
863 lwkt_preempt(thread_t ntd, int critcount)
865 struct globaldata *gd = mycpu;
873 * The caller has put us in a critical section. We can only preempt
874 * if the caller of the caller was not in a critical section (basically
875 * a local interrupt), as determined by the 'critcount' parameter. We
876 * also can't preempt if the caller is holding any spinlocks (even if
877 * he isn't in a critical section). This also handles the tokens test.
879 * YYY The target thread must be in a critical section (else it must
880 * inherit our critical section? I dunno yet).
882 * Set need_lwkt_resched() unconditionally for now YYY.
884 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
886 td = gd->gd_curthread;
887 if (ntd->td_pri <= td->td_pri) {
891 if (td->td_critcount > critcount) {
897 if (ntd->td_gd != gd) {
904 * Take the easy way out and do not preempt if we are holding
905 * any spinlocks. We could test whether the thread(s) being
906 * preempted interlock against the target thread's tokens and whether
907 * we can get all the target thread's tokens, but this situation
908 * should not occur very often so its easier to simply not preempt.
909 * Also, plain spinlocks are impossible to figure out at this point so
910 * just don't preempt.
912 * Do not try to preempt if the target thread is holding any tokens.
913 * We could try to acquire the tokens but this case is so rare there
914 * is no need to support it.
916 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
921 if (TD_TOKS_HELD(ntd)) {
926 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
931 if (ntd->td_preempted) {
938 * note: an interrupt might have occured just as we were transitioning
939 * to or from the MP lock. In this case td_mpcount will be pre-disposed
940 * (non-zero) but not actually synchronized with the actual state of the
941 * lock. We can use it to imply an MP lock requirement for the
942 * preemption but we cannot use it to test whether we hold the MP lock
945 savecnt = td->td_mpcount;
946 mpheld = MP_LOCK_HELD();
947 ntd->td_mpcount += td->td_mpcount;
948 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
949 ntd->td_mpcount -= td->td_mpcount;
957 * Since we are able to preempt the current thread, there is no need to
958 * call need_lwkt_resched().
961 ntd->td_preempted = td;
962 td->td_flags |= TDF_PREEMPT_LOCK;
963 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
966 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
968 KKASSERT(savecnt == td->td_mpcount);
969 mpheld = MP_LOCK_HELD();
970 if (mpheld && td->td_mpcount == 0)
972 else if (mpheld == 0 && td->td_mpcount)
973 panic("lwkt_preempt(): MP lock was not held through");
975 ntd->td_preempted = NULL;
976 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
980 * Conditionally call splz() if gd_reqflags indicates work is pending.
982 * td_nest_count prevents deep nesting via splz() or doreti() which
983 * might otherwise blow out the kernel stack. Note that except for
984 * this special case, we MUST call splz() here to handle any
985 * pending ints, particularly after we switch, or we might accidently
986 * halt the cpu with interrupts pending.
988 * (self contained on a per cpu basis)
993 globaldata_t gd = mycpu;
994 thread_t td = gd->gd_curthread;
996 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1001 * This function is used to negotiate a passive release of the current
1002 * process/lwp designation with the user scheduler, allowing the user
1003 * scheduler to schedule another user thread. The related kernel thread
1004 * (curthread) continues running in the released state.
1007 lwkt_passive_release(struct thread *td)
1009 struct lwp *lp = td->td_lwp;
1011 td->td_release = NULL;
1012 lwkt_setpri_self(TDPRI_KERN_USER);
1013 lp->lwp_proc->p_usched->release_curproc(lp);
1018 * This implements a normal yield. This routine is virtually a nop if
1019 * there is nothing to yield to but it will always run any pending interrupts
1020 * if called from a critical section.
1022 * This yield is designed for kernel threads without a user context.
1024 * (self contained on a per cpu basis)
1029 globaldata_t gd = mycpu;
1030 thread_t td = gd->gd_curthread;
1033 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1035 if (td->td_fairq_accum < 0) {
1036 lwkt_schedule_self(curthread);
1039 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1040 if (xtd && xtd->td_pri > td->td_pri) {
1041 lwkt_schedule_self(curthread);
1048 * This yield is designed for kernel threads with a user context.
1050 * The kernel acting on behalf of the user is potentially cpu-bound,
1051 * this function will efficiently allow other threads to run and also
1052 * switch to other processes by releasing.
1054 * The lwkt_user_yield() function is designed to have very low overhead
1055 * if no yield is determined to be needed.
1058 lwkt_user_yield(void)
1060 globaldata_t gd = mycpu;
1061 thread_t td = gd->gd_curthread;
1064 * Always run any pending interrupts in case we are in a critical
1067 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1072 * XXX SEVERE TEMPORARY HACK. A cpu-bound operation running in the
1073 * kernel can prevent other cpus from servicing interrupt threads
1074 * which still require the MP lock (which is a lot of them). This
1075 * has a chaining effect since if the interrupt is blocked, so is
1076 * the event, so normal scheduling will not pick up on the problem.
1078 if (mp_lock_contention_mask && td->td_mpcount) {
1084 * Switch (which forces a release) if another kernel thread needs
1085 * the cpu, if userland wants us to resched, or if our kernel
1086 * quantum has run out.
1088 if (lwkt_resched_wanted() ||
1089 user_resched_wanted() ||
1090 td->td_fairq_accum < 0)
1097 * Reacquire the current process if we are released.
1099 * XXX not implemented atm. The kernel may be holding locks and such,
1100 * so we want the thread to continue to receive cpu.
1102 if (td->td_release == NULL && lp) {
1103 lp->lwp_proc->p_usched->acquire_curproc(lp);
1104 td->td_release = lwkt_passive_release;
1105 lwkt_setpri_self(TDPRI_USER_NORM);
1111 * Generic schedule. Possibly schedule threads belonging to other cpus and
1112 * deal with threads that might be blocked on a wait queue.
1114 * We have a little helper inline function which does additional work after
1115 * the thread has been enqueued, including dealing with preemption and
1116 * setting need_lwkt_resched() (which prevents the kernel from returning
1117 * to userland until it has processed higher priority threads).
1119 * It is possible for this routine to be called after a failed _enqueue
1120 * (due to the target thread migrating, sleeping, or otherwise blocked).
1121 * We have to check that the thread is actually on the run queue!
1123 * reschedok is an optimized constant propagated from lwkt_schedule() or
1124 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1125 * reschedule to be requested if the target thread has a higher priority.
1126 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1127 * be 0, prevented undesired reschedules.
1131 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1135 if (ntd->td_flags & TDF_RUNQ) {
1136 if (ntd->td_preemptable && reschedok) {
1137 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1138 } else if (reschedok) {
1140 if (ntd->td_pri > otd->td_pri)
1141 need_lwkt_resched();
1145 * Give the thread a little fair share scheduler bump if it
1146 * has been asleep for a while. This is primarily to avoid
1147 * a degenerate case for interrupt threads where accumulator
1148 * crosses into negative territory unnecessarily.
1150 if (ntd->td_fairq_lticks != ticks) {
1151 ntd->td_fairq_lticks = ticks;
1152 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1153 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1154 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1161 _lwkt_schedule(thread_t td, int reschedok)
1163 globaldata_t mygd = mycpu;
1165 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1166 crit_enter_gd(mygd);
1167 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1168 if (td == mygd->gd_curthread) {
1172 * If we own the thread, there is no race (since we are in a
1173 * critical section). If we do not own the thread there might
1174 * be a race but the target cpu will deal with it.
1177 if (td->td_gd == mygd) {
1179 _lwkt_schedule_post(mygd, td, 1, reschedok);
1181 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1185 _lwkt_schedule_post(mygd, td, 1, reschedok);
1192 lwkt_schedule(thread_t td)
1194 _lwkt_schedule(td, 1);
1198 lwkt_schedule_noresched(thread_t td)
1200 _lwkt_schedule(td, 0);
1206 * When scheduled remotely if frame != NULL the IPIQ is being
1207 * run via doreti or an interrupt then preemption can be allowed.
1209 * To allow preemption we have to drop the critical section so only
1210 * one is present in _lwkt_schedule_post.
1213 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1215 thread_t td = curthread;
1218 if (frame && ntd->td_preemptable) {
1219 crit_exit_noyield(td);
1220 _lwkt_schedule(ntd, 1);
1221 crit_enter_quick(td);
1223 _lwkt_schedule(ntd, 1);
1228 * Thread migration using a 'Pull' method. The thread may or may not be
1229 * the current thread. It MUST be descheduled and in a stable state.
1230 * lwkt_giveaway() must be called on the cpu owning the thread.
1232 * At any point after lwkt_giveaway() is called, the target cpu may
1233 * 'pull' the thread by calling lwkt_acquire().
1235 * We have to make sure the thread is not sitting on a per-cpu tsleep
1236 * queue or it will blow up when it moves to another cpu.
1238 * MPSAFE - must be called under very specific conditions.
1241 lwkt_giveaway(thread_t td)
1243 globaldata_t gd = mycpu;
1246 if (td->td_flags & TDF_TSLEEPQ)
1248 KKASSERT(td->td_gd == gd);
1249 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1250 td->td_flags |= TDF_MIGRATING;
1255 lwkt_acquire(thread_t td)
1260 KKASSERT(td->td_flags & TDF_MIGRATING);
1265 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1266 crit_enter_gd(mygd);
1267 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1269 lwkt_process_ipiq();
1274 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1275 td->td_flags &= ~TDF_MIGRATING;
1278 crit_enter_gd(mygd);
1279 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1280 td->td_flags &= ~TDF_MIGRATING;
1288 * Generic deschedule. Descheduling threads other then your own should be
1289 * done only in carefully controlled circumstances. Descheduling is
1292 * This function may block if the cpu has run out of messages.
1295 lwkt_deschedule(thread_t td)
1299 if (td == curthread) {
1302 if (td->td_gd == mycpu) {
1305 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1315 * Set the target thread's priority. This routine does not automatically
1316 * switch to a higher priority thread, LWKT threads are not designed for
1317 * continuous priority changes. Yield if you want to switch.
1320 lwkt_setpri(thread_t td, int pri)
1322 KKASSERT(td->td_gd == mycpu);
1323 if (td->td_pri != pri) {
1326 if (td->td_flags & TDF_RUNQ) {
1338 * Set the initial priority for a thread prior to it being scheduled for
1339 * the first time. The thread MUST NOT be scheduled before or during
1340 * this call. The thread may be assigned to a cpu other then the current
1343 * Typically used after a thread has been created with TDF_STOPPREQ,
1344 * and before the thread is initially scheduled.
1347 lwkt_setpri_initial(thread_t td, int pri)
1350 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1355 lwkt_setpri_self(int pri)
1357 thread_t td = curthread;
1359 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1361 if (td->td_flags & TDF_RUNQ) {
1372 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1374 * Example: two competing threads, same priority N. decrement by (2*N)
1375 * increment by N*8, each thread will get 4 ticks.
1378 lwkt_fairq_schedulerclock(thread_t td)
1382 if (td != &td->td_gd->gd_idlethread) {
1383 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1384 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1385 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1386 if (td->td_fairq_accum < 0)
1387 need_lwkt_resched();
1388 td->td_fairq_lticks = ticks;
1390 td = td->td_preempted;
1396 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1398 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1399 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1400 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1404 * Migrate the current thread to the specified cpu.
1406 * This is accomplished by descheduling ourselves from the current cpu,
1407 * moving our thread to the tdallq of the target cpu, IPI messaging the
1408 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1409 * races while the thread is being migrated.
1411 * We must be sure to remove ourselves from the current cpu's tsleepq
1412 * before potentially moving to another queue. The thread can be on
1413 * a tsleepq due to a left-over tsleep_interlock().
1416 static void lwkt_setcpu_remote(void *arg);
1420 lwkt_setcpu_self(globaldata_t rgd)
1423 thread_t td = curthread;
1425 if (td->td_gd != rgd) {
1426 crit_enter_quick(td);
1427 if (td->td_flags & TDF_TSLEEPQ)
1429 td->td_flags |= TDF_MIGRATING;
1430 lwkt_deschedule_self(td);
1431 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1432 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1434 /* we are now on the target cpu */
1435 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1436 crit_exit_quick(td);
1442 lwkt_migratecpu(int cpuid)
1447 rgd = globaldata_find(cpuid);
1448 lwkt_setcpu_self(rgd);
1453 * Remote IPI for cpu migration (called while in a critical section so we
1454 * do not have to enter another one). The thread has already been moved to
1455 * our cpu's allq, but we must wait for the thread to be completely switched
1456 * out on the originating cpu before we schedule it on ours or the stack
1457 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1458 * change to main memory.
1460 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1461 * against wakeups. It is best if this interface is used only when there
1462 * are no pending events that might try to schedule the thread.
1466 lwkt_setcpu_remote(void *arg)
1469 globaldata_t gd = mycpu;
1471 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1473 lwkt_process_ipiq();
1479 td->td_flags &= ~TDF_MIGRATING;
1480 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1486 lwkt_preempted_proc(void)
1488 thread_t td = curthread;
1489 while (td->td_preempted)
1490 td = td->td_preempted;
1495 * Create a kernel process/thread/whatever. It shares it's address space
1496 * with proc0 - ie: kernel only.
1498 * NOTE! By default new threads are created with the MP lock held. A
1499 * thread which does not require the MP lock should release it by calling
1500 * rel_mplock() at the start of the new thread.
1503 lwkt_create(void (*func)(void *), void *arg,
1504 struct thread **tdp, thread_t template, int tdflags, int cpu,
1505 const char *fmt, ...)
1510 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1514 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1517 * Set up arg0 for 'ps' etc
1519 __va_start(ap, fmt);
1520 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1524 * Schedule the thread to run
1526 if ((td->td_flags & TDF_STOPREQ) == 0)
1529 td->td_flags &= ~TDF_STOPREQ;
1534 * Destroy an LWKT thread. Warning! This function is not called when
1535 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1536 * uses a different reaping mechanism.
1541 thread_t td = curthread;
1545 if (td->td_flags & TDF_VERBOSE)
1546 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1550 * Get us into a critical section to interlock gd_freetd and loop
1551 * until we can get it freed.
1553 * We have to cache the current td in gd_freetd because objcache_put()ing
1554 * it would rip it out from under us while our thread is still active.
1557 crit_enter_quick(td);
1558 while ((std = gd->gd_freetd) != NULL) {
1559 gd->gd_freetd = NULL;
1560 objcache_put(thread_cache, std);
1564 * Remove thread resources from kernel lists and deschedule us for
1567 if (td->td_flags & TDF_TSLEEPQ)
1570 dsched_exit_thread(td);
1571 lwkt_deschedule_self(td);
1572 lwkt_remove_tdallq(td);
1573 if (td->td_flags & TDF_ALLOCATED_THREAD)
1579 lwkt_remove_tdallq(thread_t td)
1581 KKASSERT(td->td_gd == mycpu);
1582 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1588 thread_t td = curthread;
1589 int lpri = td->td_pri;
1592 panic("td_pri is/would-go negative! %p %d", td, lpri);
1598 * Called from debugger/panic on cpus which have been stopped. We must still
1599 * process the IPIQ while stopped, even if we were stopped while in a critical
1602 * If we are dumping also try to process any pending interrupts. This may
1603 * or may not work depending on the state of the cpu at the point it was
1607 lwkt_smp_stopped(void)
1609 globaldata_t gd = mycpu;
1613 lwkt_process_ipiq();
1616 lwkt_process_ipiq();