2 * Copyright (c) 2003-2011 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>
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 #ifdef _KERNEL_VIRTUAL
78 #if !defined(KTR_CTXSW)
79 #define KTR_CTXSW KTR_ALL
81 KTR_INFO_MASTER(ctxsw);
82 KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p", int cpu, struct thread *td);
83 KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p", int cpu, struct thread *td);
84 KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s", struct thread *td, char *comm);
85 KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", struct thread *td);
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 int lwkt_use_spin_port;
97 static struct objcache *thread_cache;
98 int cpu_mwait_spin = 0;
100 static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
101 static void lwkt_setcpu_remote(void *arg);
104 * We can make all thread ports use the spin backend instead of the thread
105 * backend. This should only be set to debug the spin backend.
107 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
110 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
111 "Panic if attempting to switch lwkt's while mastering cpusync");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
114 "Number of switched threads");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
116 "Successful preemption events");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
118 "Failed preemption events");
119 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
120 "Number of preempted threads.");
121 static int fairq_enable = 0;
122 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
123 &fairq_enable, 0, "Turn on fairq priority accumulators");
124 static int fairq_bypass = -1;
125 SYSCTL_INT(_lwkt, OID_AUTO, fairq_bypass, CTLFLAG_RW,
126 &fairq_bypass, 0, "Allow fairq to bypass td on token failure");
127 extern int lwkt_sched_debug;
128 int lwkt_sched_debug = 0;
129 SYSCTL_INT(_lwkt, OID_AUTO, sched_debug, CTLFLAG_RW,
130 &lwkt_sched_debug, 0, "Scheduler debug");
131 static int lwkt_spin_loops = 10;
132 SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
133 &lwkt_spin_loops, 0, "Scheduler spin loops until sorted decon");
134 static int preempt_enable = 1;
135 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
136 &preempt_enable, 0, "Enable preemption");
137 static int lwkt_cache_threads = 0;
138 SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
139 &lwkt_cache_threads, 0, "thread+kstack cache");
142 * These helper procedures handle the runq, they can only be called from
143 * within a critical section.
145 * WARNING! Prior to SMP being brought up it is possible to enqueue and
146 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
147 * instead of 'mycpu' when referencing the globaldata structure. Once
148 * SMP live enqueuing and dequeueing only occurs on the current cpu.
152 _lwkt_dequeue(thread_t td)
154 if (td->td_flags & TDF_RUNQ) {
155 struct globaldata *gd = td->td_gd;
157 td->td_flags &= ~TDF_RUNQ;
158 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
159 --gd->gd_tdrunqcount;
160 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
161 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
168 * There are a limited number of lwkt threads runnable since user
169 * processes only schedule one at a time per cpu. However, there can
170 * be many user processes in kernel mode exiting from a tsleep() which
173 * NOTE: lwkt_schedulerclock() will force a round-robin based on td_pri and
174 * will ignore user priority. This is to ensure that user threads in
175 * kernel mode get cpu at some point regardless of what the user
180 _lwkt_enqueue(thread_t td)
184 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
185 struct globaldata *gd = td->td_gd;
187 td->td_flags |= TDF_RUNQ;
188 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
190 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
191 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
194 * NOTE: td_upri - higher numbers more desireable, same sense
195 * as td_pri (typically reversed from lwp_upri).
197 * In the equal priority case we want the best selection
198 * at the beginning so the less desireable selections know
199 * that they have to setrunqueue/go-to-another-cpu, even
200 * though it means switching back to the 'best' selection.
201 * This also avoids degenerate situations when many threads
202 * are runnable or waking up at the same time.
204 * If upri matches exactly place at end/round-robin.
207 (xtd->td_pri >= td->td_pri ||
208 (xtd->td_pri == td->td_pri &&
209 xtd->td_upri >= td->td_upri))) {
210 xtd = TAILQ_NEXT(xtd, td_threadq);
213 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
215 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
217 ++gd->gd_tdrunqcount;
220 * Request a LWKT reschedule if we are now at the head of the queue.
222 if (TAILQ_FIRST(&gd->gd_tdrunq) == td)
228 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
230 struct thread *td = (struct thread *)obj;
232 td->td_kstack = NULL;
233 td->td_kstack_size = 0;
234 td->td_flags = TDF_ALLOCATED_THREAD;
240 _lwkt_thread_dtor(void *obj, void *privdata)
242 struct thread *td = (struct thread *)obj;
244 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
245 ("_lwkt_thread_dtor: not allocated from objcache"));
246 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
247 td->td_kstack_size > 0,
248 ("_lwkt_thread_dtor: corrupted stack"));
249 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
250 td->td_kstack = NULL;
255 * Initialize the lwkt s/system.
257 * Nominally cache up to 32 thread + kstack structures. Cache more on
258 * systems with a lot of cpu cores.
263 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
264 if (lwkt_cache_threads == 0) {
265 lwkt_cache_threads = ncpus * 4;
266 if (lwkt_cache_threads < 32)
267 lwkt_cache_threads = 32;
269 thread_cache = objcache_create_mbacked(
270 M_THREAD, sizeof(struct thread),
271 0, lwkt_cache_threads,
272 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
274 SYSINIT(lwkt_init, SI_BOOT2_LWKT_INIT, SI_ORDER_FIRST, lwkt_init, NULL);
277 * Schedule a thread to run. As the current thread we can always safely
278 * schedule ourselves, and a shortcut procedure is provided for that
281 * (non-blocking, self contained on a per cpu basis)
284 lwkt_schedule_self(thread_t td)
286 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
287 crit_enter_quick(td);
288 KASSERT(td != &td->td_gd->gd_idlethread,
289 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
290 KKASSERT(td->td_lwp == NULL ||
291 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
297 * Deschedule a thread.
299 * (non-blocking, self contained on a per cpu basis)
302 lwkt_deschedule_self(thread_t td)
304 crit_enter_quick(td);
310 * LWKTs operate on a per-cpu basis
312 * WARNING! Called from early boot, 'mycpu' may not work yet.
315 lwkt_gdinit(struct globaldata *gd)
317 TAILQ_INIT(&gd->gd_tdrunq);
318 TAILQ_INIT(&gd->gd_tdallq);
322 * Create a new thread. The thread must be associated with a process context
323 * or LWKT start address before it can be scheduled. If the target cpu is
324 * -1 the thread will be created on the current cpu.
326 * If you intend to create a thread without a process context this function
327 * does everything except load the startup and switcher function.
330 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
332 static int cpu_rotator;
333 globaldata_t gd = mycpu;
337 * If static thread storage is not supplied allocate a thread. Reuse
338 * a cached free thread if possible. gd_freetd is used to keep an exiting
339 * thread intact through the exit.
343 if ((td = gd->gd_freetd) != NULL) {
344 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
346 gd->gd_freetd = NULL;
348 td = objcache_get(thread_cache, M_WAITOK);
349 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
353 KASSERT((td->td_flags &
354 (TDF_ALLOCATED_THREAD|TDF_RUNNING|TDF_PREEMPT_LOCK)) ==
355 TDF_ALLOCATED_THREAD,
356 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
357 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
361 * Try to reuse cached stack.
363 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
364 if (flags & TDF_ALLOCATED_STACK) {
365 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
370 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
371 flags |= TDF_ALLOCATED_STACK;
378 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
383 * Initialize a preexisting thread structure. This function is used by
384 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
386 * All threads start out in a critical section at a priority of
387 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
388 * appropriate. This function may send an IPI message when the
389 * requested cpu is not the current cpu and consequently gd_tdallq may
390 * not be initialized synchronously from the point of view of the originating
393 * NOTE! we have to be careful in regards to creating threads for other cpus
394 * if SMP has not yet been activated.
397 lwkt_init_thread_remote(void *arg)
402 * Protected by critical section held by IPI dispatch
404 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
408 * lwkt core thread structural initialization.
410 * NOTE: All threads are initialized as mpsafe threads.
413 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
414 struct globaldata *gd)
416 globaldata_t mygd = mycpu;
418 bzero(td, sizeof(struct thread));
419 td->td_kstack = stack;
420 td->td_kstack_size = stksize;
421 td->td_flags = flags;
423 td->td_type = TD_TYPE_GENERIC;
425 td->td_pri = TDPRI_KERN_DAEMON;
426 td->td_critcount = 1;
427 td->td_toks_have = NULL;
428 td->td_toks_stop = &td->td_toks_base;
429 if (lwkt_use_spin_port || (flags & TDF_FORCE_SPINPORT)) {
430 lwkt_initport_spin(&td->td_msgport, td,
431 (flags & TDF_FIXEDCPU) ? TRUE : FALSE);
433 lwkt_initport_thread(&td->td_msgport, td);
435 pmap_init_thread(td);
437 * Normally initializing a thread for a remote cpu requires sending an
438 * IPI. However, the idlethread is setup before the other cpus are
439 * activated so we have to treat it as a special case. XXX manipulation
440 * of gd_tdallq requires the BGL.
442 if (gd == mygd || td == &gd->gd_idlethread) {
444 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
447 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
449 dsched_enter_thread(td);
453 lwkt_set_comm(thread_t td, const char *ctl, ...)
458 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
460 KTR_LOG(ctxsw_newtd, td, td->td_comm);
464 * Prevent the thread from getting destroyed. Note that unlike PHOLD/PRELE
465 * this does not prevent the thread from migrating to another cpu so the
466 * gd_tdallq state is not protected by this.
469 lwkt_hold(thread_t td)
471 atomic_add_int(&td->td_refs, 1);
475 lwkt_rele(thread_t td)
477 KKASSERT(td->td_refs > 0);
478 atomic_add_int(&td->td_refs, -1);
482 lwkt_free_thread(thread_t td)
484 KKASSERT(td->td_refs == 0);
485 KKASSERT((td->td_flags & (TDF_RUNNING | TDF_PREEMPT_LOCK |
486 TDF_RUNQ | TDF_TSLEEPQ)) == 0);
487 if (td->td_flags & TDF_ALLOCATED_THREAD) {
488 objcache_put(thread_cache, td);
489 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
490 /* client-allocated struct with internally allocated stack */
491 KASSERT(td->td_kstack && td->td_kstack_size > 0,
492 ("lwkt_free_thread: corrupted stack"));
493 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
494 td->td_kstack = NULL;
495 td->td_kstack_size = 0;
498 KTR_LOG(ctxsw_deadtd, td);
503 * Switch to the next runnable lwkt. If no LWKTs are runnable then
504 * switch to the idlethread. Switching must occur within a critical
505 * section to avoid races with the scheduling queue.
507 * We always have full control over our cpu's run queue. Other cpus
508 * that wish to manipulate our queue must use the cpu_*msg() calls to
509 * talk to our cpu, so a critical section is all that is needed and
510 * the result is very, very fast thread switching.
512 * The LWKT scheduler uses a fixed priority model and round-robins at
513 * each priority level. User process scheduling is a totally
514 * different beast and LWKT priorities should not be confused with
515 * user process priorities.
517 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
518 * is not called by the current thread in the preemption case, only when
519 * the preempting thread blocks (in order to return to the original thread).
521 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
522 * migration and tsleep deschedule the current lwkt thread and call
523 * lwkt_switch(). In particular, the target cpu of the migration fully
524 * expects the thread to become non-runnable and can deadlock against
525 * cpusync operations if we run any IPIs prior to switching the thread out.
527 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
528 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
533 globaldata_t gd = mycpu;
534 thread_t td = gd->gd_curthread;
538 KKASSERT(gd->gd_processing_ipiq == 0);
539 KKASSERT(td->td_flags & TDF_RUNNING);
542 * Switching from within a 'fast' (non thread switched) interrupt or IPI
543 * is illegal. However, we may have to do it anyway if we hit a fatal
544 * kernel trap or we have paniced.
546 * If this case occurs save and restore the interrupt nesting level.
548 if (gd->gd_intr_nesting_level) {
552 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
553 panic("lwkt_switch: Attempt to switch from a "
554 "fast interrupt, ipi, or hard code section, "
558 savegdnest = gd->gd_intr_nesting_level;
559 savegdtrap = gd->gd_trap_nesting_level;
560 gd->gd_intr_nesting_level = 0;
561 gd->gd_trap_nesting_level = 0;
562 if ((td->td_flags & TDF_PANICWARN) == 0) {
563 td->td_flags |= TDF_PANICWARN;
564 kprintf("Warning: thread switch from interrupt, IPI, "
565 "or hard code section.\n"
566 "thread %p (%s)\n", td, td->td_comm);
570 gd->gd_intr_nesting_level = savegdnest;
571 gd->gd_trap_nesting_level = savegdtrap;
577 * Release our current user process designation if we are blocking
578 * or if a user reschedule was requested.
580 * NOTE: This function is NOT called if we are switching into or
581 * returning from a preemption.
583 * NOTE: Releasing our current user process designation may cause
584 * it to be assigned to another thread, which in turn will
585 * cause us to block in the usched acquire code when we attempt
586 * to return to userland.
588 * NOTE: On SMP systems this can be very nasty when heavy token
589 * contention is present so we want to be careful not to
590 * release the designation gratuitously.
592 if (td->td_release &&
593 (user_resched_wanted() || (td->td_flags & TDF_RUNQ) == 0)) {
598 * Release all tokens. Once we do this we must remain in the critical
599 * section and cannot run IPIs or other interrupts until we switch away
600 * because they may implode if they try to get a token using our thread
604 if (TD_TOKS_HELD(td))
605 lwkt_relalltokens(td);
608 * We had better not be holding any spin locks, but don't get into an
609 * endless panic loop.
611 KASSERT(gd->gd_spinlocks == 0 || panicstr != NULL,
612 ("lwkt_switch: still holding %d exclusive spinlocks!",
616 if (td->td_cscount) {
617 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
619 if (panic_on_cscount)
620 panic("switching while mastering cpusync");
625 * If we had preempted another thread on this cpu, resume the preempted
626 * thread. This occurs transparently, whether the preempted thread
627 * was scheduled or not (it may have been preempted after descheduling
630 * We have to setup the MP lock for the original thread after backing
631 * out the adjustment that was made to curthread when the original
634 if ((ntd = td->td_preempted) != NULL) {
635 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
636 ntd->td_flags |= TDF_PREEMPT_DONE;
639 * The interrupt may have woken a thread up, we need to properly
640 * set the reschedule flag if the originally interrupted thread is
641 * at a lower priority.
643 * The interrupt may not have descheduled.
645 if (TAILQ_FIRST(&gd->gd_tdrunq) != ntd)
647 goto havethread_preempted;
651 * Figure out switch target. If we cannot switch to our desired target
652 * look for a thread that we can switch to.
654 * NOTE! The limited spin loop and related parameters are extremely
655 * important for system performance, particularly for pipes and
656 * concurrent conflicting VM faults.
658 clear_lwkt_resched();
659 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
663 if (TD_TOKS_NOT_HELD(ntd) ||
664 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops)))
668 ++gd->gd_cnt.v_lock_colls;
670 } while (ntd->td_contended < (lwkt_spin_loops >> 1));
674 * Bleh, the thread we wanted to switch to has a contended token.
675 * See if we can switch to another thread.
677 * We generally don't want to do this because it represents a
678 * priority inversion. Do not allow the case if the thread
679 * is returning to userland (not a kernel thread) AND the thread
682 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
683 if (ntd->td_pri < TDPRI_KERN_LPSCHED && upri > ntd->td_upri)
690 if (TD_TOKS_NOT_HELD(ntd) ||
691 lwkt_getalltokens(ntd, (ntd->td_contended > lwkt_spin_loops))) {
695 ++gd->gd_cnt.v_lock_colls;
699 * Fall through, switch to idle thread to get us out of the current
700 * context. Since we were contended, prevent HLT by flagging a
707 * We either contended on ntd or the runq is empty. We must switch
708 * through the idle thread to get out of the current context.
710 ntd = &gd->gd_idlethread;
711 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
712 ASSERT_NO_TOKENS_HELD(ntd);
713 cpu_time.cp_msg[0] = 0;
714 cpu_time.cp_stallpc = 0;
719 * Clear gd_idle_repeat when doing a normal switch to a non-idle
722 ntd->td_wmesg = NULL;
723 ntd->td_contended = 0;
724 ++gd->gd_cnt.v_swtch;
725 gd->gd_idle_repeat = 0;
727 havethread_preempted:
729 * If the new target does not need the MP lock and we are holding it,
730 * release the MP lock. If the new target requires the MP lock we have
731 * already acquired it for the target.
735 KASSERT(ntd->td_critcount,
736 ("priority problem in lwkt_switch %d %d",
737 td->td_critcount, ntd->td_critcount));
741 * Execute the actual thread switch operation. This function
742 * returns to the current thread and returns the previous thread
743 * (which may be different from the thread we switched to).
745 * We are responsible for marking ntd as TDF_RUNNING.
747 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
749 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
750 ntd->td_flags |= TDF_RUNNING;
751 lwkt_switch_return(td->td_switch(ntd));
752 /* ntd invalid, td_switch() can return a different thread_t */
756 * catch-all. XXX is this strictly needed?
760 /* NOTE: current cpu may have changed after switch */
765 * Called by assembly in the td_switch (thread restore path) for thread
766 * bootstrap cases which do not 'return' to lwkt_switch().
769 lwkt_switch_return(thread_t otd)
774 * Check if otd was migrating. Now that we are on ntd we can finish
775 * up the migration. This is a bit messy but it is the only place
776 * where td is known to be fully descheduled.
778 * We can only activate the migration if otd was migrating but not
779 * held on the cpu due to a preemption chain. We still have to
780 * clear TDF_RUNNING on the old thread either way.
782 * We are responsible for clearing the previously running thread's
785 if ((rgd = otd->td_migrate_gd) != NULL &&
786 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
787 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
788 (TDF_MIGRATING | TDF_RUNNING));
789 otd->td_migrate_gd = NULL;
790 otd->td_flags &= ~TDF_RUNNING;
791 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
793 otd->td_flags &= ~TDF_RUNNING;
797 * Final exit validations (see lwp_wait()). Note that otd becomes
798 * invalid the *instant* we set TDF_MP_EXITSIG.
800 while (otd->td_flags & TDF_EXITING) {
803 mpflags = otd->td_mpflags;
806 if (mpflags & TDF_MP_EXITWAIT) {
807 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
808 mpflags | TDF_MP_EXITSIG)) {
813 if (atomic_cmpset_int(&otd->td_mpflags, mpflags,
814 mpflags | TDF_MP_EXITSIG)) {
823 * Request that the target thread preempt the current thread. Preemption
824 * can only occur if our only critical section is the one that we were called
825 * with, the relative priority of the target thread is higher, and the target
826 * thread holds no tokens. This also only works if we are not holding any
827 * spinlocks (obviously).
829 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
830 * this is called via lwkt_schedule() through the td_preemptable callback.
831 * critcount is the managed critical priority that we should ignore in order
832 * to determine whether preemption is possible (aka usually just the crit
833 * priority of lwkt_schedule() itself).
835 * Preemption is typically limited to interrupt threads.
837 * Operation works in a fairly straight-forward manner. The normal
838 * scheduling code is bypassed and we switch directly to the target
839 * thread. When the target thread attempts to block or switch away
840 * code at the base of lwkt_switch() will switch directly back to our
841 * thread. Our thread is able to retain whatever tokens it holds and
842 * if the target needs one of them the target will switch back to us
843 * and reschedule itself normally.
846 lwkt_preempt(thread_t ntd, int critcount)
848 struct globaldata *gd = mycpu;
851 int save_gd_intr_nesting_level;
854 * The caller has put us in a critical section. We can only preempt
855 * if the caller of the caller was not in a critical section (basically
856 * a local interrupt), as determined by the 'critcount' parameter. We
857 * also can't preempt if the caller is holding any spinlocks (even if
858 * he isn't in a critical section). This also handles the tokens test.
860 * YYY The target thread must be in a critical section (else it must
861 * inherit our critical section? I dunno yet).
863 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
865 td = gd->gd_curthread;
866 if (preempt_enable == 0) {
870 if (ntd->td_pri <= td->td_pri) {
874 if (td->td_critcount > critcount) {
878 if (td->td_cscount) {
882 if (ntd->td_gd != gd) {
887 * We don't have to check spinlocks here as they will also bump
890 * Do not try to preempt if the target thread is holding any tokens.
891 * We could try to acquire the tokens but this case is so rare there
892 * is no need to support it.
894 KKASSERT(gd->gd_spinlocks == 0);
896 if (TD_TOKS_HELD(ntd)) {
900 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
904 if (ntd->td_preempted) {
908 KKASSERT(gd->gd_processing_ipiq == 0);
911 * Since we are able to preempt the current thread, there is no need to
912 * call need_lwkt_resched().
914 * We must temporarily clear gd_intr_nesting_level around the switch
915 * since switchouts from the target thread are allowed (they will just
916 * return to our thread), and since the target thread has its own stack.
918 * A preemption must switch back to the original thread, assert the
922 ntd->td_preempted = td;
923 td->td_flags |= TDF_PREEMPT_LOCK;
924 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
925 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
926 gd->gd_intr_nesting_level = 0;
928 KKASSERT((ntd->td_flags & TDF_RUNNING) == 0);
929 ntd->td_flags |= TDF_RUNNING;
930 xtd = td->td_switch(ntd);
931 KKASSERT(xtd == ntd);
932 lwkt_switch_return(xtd);
933 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
935 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
936 ntd->td_preempted = NULL;
937 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
941 * Conditionally call splz() if gd_reqflags indicates work is pending.
942 * This will work inside a critical section but not inside a hard code
945 * (self contained on a per cpu basis)
950 globaldata_t gd = mycpu;
951 thread_t td = gd->gd_curthread;
953 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
954 gd->gd_intr_nesting_level == 0 &&
955 td->td_nest_count < 2)
962 * This version is integrated into crit_exit, reqflags has already
963 * been tested but td_critcount has not.
965 * We only want to execute the splz() on the 1->0 transition of
966 * critcount and not in a hard code section or if too deeply nested.
968 * NOTE: gd->gd_spinlocks is implied to be 0 when td_critcount is 0.
971 lwkt_maybe_splz(thread_t td)
973 globaldata_t gd = td->td_gd;
975 if (td->td_critcount == 0 &&
976 gd->gd_intr_nesting_level == 0 &&
977 td->td_nest_count < 2)
984 * Drivers which set up processing co-threads can call this function to
985 * run the co-thread at a higher priority and to allow it to preempt
989 lwkt_set_interrupt_support_thread(void)
991 thread_t td = curthread;
993 lwkt_setpri_self(TDPRI_INT_SUPPORT);
994 td->td_flags |= TDF_INTTHREAD;
995 td->td_preemptable = lwkt_preempt;
1000 * This function is used to negotiate a passive release of the current
1001 * process/lwp designation with the user scheduler, allowing the user
1002 * scheduler to schedule another user thread. The related kernel thread
1003 * (curthread) continues running in the released state.
1006 lwkt_passive_release(struct thread *td)
1008 struct lwp *lp = td->td_lwp;
1010 td->td_release = NULL;
1011 lwkt_setpri_self(TDPRI_KERN_USER);
1013 lp->lwp_proc->p_usched->release_curproc(lp);
1018 * This implements a LWKT yield, allowing a kernel thread to yield to other
1019 * kernel threads at the same or higher priority. This function can be
1020 * called in a tight loop and will typically only yield once per tick.
1022 * Most kernel threads run at the same priority in order to allow equal
1025 * (self contained on a per cpu basis)
1030 globaldata_t gd = mycpu;
1031 thread_t td = gd->gd_curthread;
1033 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1035 if (lwkt_resched_wanted()) {
1036 lwkt_schedule_self(curthread);
1042 * The quick version processes pending interrupts and higher-priority
1043 * LWKT threads but will not round-robin same-priority LWKT threads.
1045 * When called while attempting to return to userland the only same-pri
1046 * threads are the ones which have already tried to become the current
1050 lwkt_yield_quick(void)
1052 globaldata_t gd = mycpu;
1053 thread_t td = gd->gd_curthread;
1055 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1057 if (lwkt_resched_wanted()) {
1059 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1060 clear_lwkt_resched();
1062 lwkt_schedule_self(curthread);
1070 * This yield is designed for kernel threads with a user context.
1072 * The kernel acting on behalf of the user is potentially cpu-bound,
1073 * this function will efficiently allow other threads to run and also
1074 * switch to other processes by releasing.
1076 * The lwkt_user_yield() function is designed to have very low overhead
1077 * if no yield is determined to be needed.
1080 lwkt_user_yield(void)
1082 globaldata_t gd = mycpu;
1083 thread_t td = gd->gd_curthread;
1086 * Always run any pending interrupts in case we are in a critical
1089 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1093 * Switch (which forces a release) if another kernel thread needs
1094 * the cpu, if userland wants us to resched, or if our kernel
1095 * quantum has run out.
1097 if (lwkt_resched_wanted() ||
1098 user_resched_wanted())
1105 * Reacquire the current process if we are released.
1107 * XXX not implemented atm. The kernel may be holding locks and such,
1108 * so we want the thread to continue to receive cpu.
1110 if (td->td_release == NULL && lp) {
1111 lp->lwp_proc->p_usched->acquire_curproc(lp);
1112 td->td_release = lwkt_passive_release;
1113 lwkt_setpri_self(TDPRI_USER_NORM);
1119 * Generic schedule. Possibly schedule threads belonging to other cpus and
1120 * deal with threads that might be blocked on a wait queue.
1122 * We have a little helper inline function which does additional work after
1123 * the thread has been enqueued, including dealing with preemption and
1124 * setting need_lwkt_resched() (which prevents the kernel from returning
1125 * to userland until it has processed higher priority threads).
1127 * It is possible for this routine to be called after a failed _enqueue
1128 * (due to the target thread migrating, sleeping, or otherwise blocked).
1129 * We have to check that the thread is actually on the run queue!
1133 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount)
1135 if (ntd->td_flags & TDF_RUNQ) {
1136 if (ntd->td_preemptable) {
1137 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1144 _lwkt_schedule(thread_t td)
1146 globaldata_t mygd = mycpu;
1148 KASSERT(td != &td->td_gd->gd_idlethread,
1149 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1150 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1151 crit_enter_gd(mygd);
1152 KKASSERT(td->td_lwp == NULL ||
1153 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1155 if (td == mygd->gd_curthread) {
1159 * If we own the thread, there is no race (since we are in a
1160 * critical section). If we do not own the thread there might
1161 * be a race but the target cpu will deal with it.
1163 if (td->td_gd == mygd) {
1165 _lwkt_schedule_post(mygd, td, 1);
1167 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1174 lwkt_schedule(thread_t td)
1180 lwkt_schedule_noresched(thread_t td) /* XXX not impl */
1186 * When scheduled remotely if frame != NULL the IPIQ is being
1187 * run via doreti or an interrupt then preemption can be allowed.
1189 * To allow preemption we have to drop the critical section so only
1190 * one is present in _lwkt_schedule_post.
1193 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1195 thread_t td = curthread;
1198 if (frame && ntd->td_preemptable) {
1199 crit_exit_noyield(td);
1200 _lwkt_schedule(ntd);
1201 crit_enter_quick(td);
1203 _lwkt_schedule(ntd);
1208 * Thread migration using a 'Pull' method. The thread may or may not be
1209 * the current thread. It MUST be descheduled and in a stable state.
1210 * lwkt_giveaway() must be called on the cpu owning the thread.
1212 * At any point after lwkt_giveaway() is called, the target cpu may
1213 * 'pull' the thread by calling lwkt_acquire().
1215 * We have to make sure the thread is not sitting on a per-cpu tsleep
1216 * queue or it will blow up when it moves to another cpu.
1218 * MPSAFE - must be called under very specific conditions.
1221 lwkt_giveaway(thread_t td)
1223 globaldata_t gd = mycpu;
1226 if (td->td_flags & TDF_TSLEEPQ)
1228 KKASSERT(td->td_gd == gd);
1229 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1230 td->td_flags |= TDF_MIGRATING;
1235 lwkt_acquire(thread_t td)
1239 int retry = 10000000;
1241 KKASSERT(td->td_flags & TDF_MIGRATING);
1246 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1247 crit_enter_gd(mygd);
1248 DEBUG_PUSH_INFO("lwkt_acquire");
1249 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1250 lwkt_process_ipiq();
1253 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1257 #ifdef _KERNEL_VIRTUAL
1264 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1265 td->td_flags &= ~TDF_MIGRATING;
1268 crit_enter_gd(mygd);
1269 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1270 td->td_flags &= ~TDF_MIGRATING;
1276 * Generic deschedule. Descheduling threads other then your own should be
1277 * done only in carefully controlled circumstances. Descheduling is
1280 * This function may block if the cpu has run out of messages.
1283 lwkt_deschedule(thread_t td)
1286 if (td == curthread) {
1289 if (td->td_gd == mycpu) {
1292 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1299 * Set the target thread's priority. This routine does not automatically
1300 * switch to a higher priority thread, LWKT threads are not designed for
1301 * continuous priority changes. Yield if you want to switch.
1304 lwkt_setpri(thread_t td, int pri)
1306 if (td->td_pri != pri) {
1309 if (td->td_flags & TDF_RUNQ) {
1310 KKASSERT(td->td_gd == mycpu);
1322 * Set the initial priority for a thread prior to it being scheduled for
1323 * the first time. The thread MUST NOT be scheduled before or during
1324 * this call. The thread may be assigned to a cpu other then the current
1327 * Typically used after a thread has been created with TDF_STOPPREQ,
1328 * and before the thread is initially scheduled.
1331 lwkt_setpri_initial(thread_t td, int pri)
1334 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1339 lwkt_setpri_self(int pri)
1341 thread_t td = curthread;
1343 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1345 if (td->td_flags & TDF_RUNQ) {
1356 * hz tick scheduler clock for LWKT threads
1359 lwkt_schedulerclock(thread_t td)
1361 globaldata_t gd = td->td_gd;
1364 if (TAILQ_FIRST(&gd->gd_tdrunq) == td) {
1366 * If the current thread is at the head of the runq shift it to the
1367 * end of any equal-priority threads and request a LWKT reschedule
1370 * Ignore upri in this situation. There will only be one user thread
1371 * in user mode, all others will be user threads running in kernel
1372 * mode and we have to make sure they get some cpu.
1374 xtd = TAILQ_NEXT(td, td_threadq);
1375 if (xtd && xtd->td_pri == td->td_pri) {
1376 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
1377 while (xtd && xtd->td_pri == td->td_pri)
1378 xtd = TAILQ_NEXT(xtd, td_threadq);
1380 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
1382 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
1383 need_lwkt_resched();
1387 * If we scheduled a thread other than the one at the head of the
1388 * queue always request a reschedule every tick.
1390 need_lwkt_resched();
1395 * Migrate the current thread to the specified cpu.
1397 * This is accomplished by descheduling ourselves from the current cpu
1398 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1399 * 'old' thread wants to migrate after it has been completely switched out
1400 * and will complete the migration.
1402 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1404 * We must be sure to release our current process designation (if a user
1405 * process) before clearing out any tsleepq we are on because the release
1406 * code may re-add us.
1408 * We must be sure to remove ourselves from the current cpu's tsleepq
1409 * before potentially moving to another queue. The thread can be on
1410 * a tsleepq due to a left-over tsleep_interlock().
1414 lwkt_setcpu_self(globaldata_t rgd)
1416 thread_t td = curthread;
1418 if (td->td_gd != rgd) {
1419 crit_enter_quick(td);
1423 if (td->td_flags & TDF_TSLEEPQ)
1427 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1428 * trying to deschedule ourselves and switch away, then deschedule
1429 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1430 * call lwkt_switch() to complete the operation.
1432 td->td_flags |= TDF_MIGRATING;
1433 lwkt_deschedule_self(td);
1434 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1435 td->td_migrate_gd = rgd;
1439 * We are now on the target cpu
1441 KKASSERT(rgd == mycpu);
1442 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1443 crit_exit_quick(td);
1448 lwkt_migratecpu(int cpuid)
1452 rgd = globaldata_find(cpuid);
1453 lwkt_setcpu_self(rgd);
1457 * Remote IPI for cpu migration (called while in a critical section so we
1458 * do not have to enter another one).
1460 * The thread (td) has already been completely descheduled from the
1461 * originating cpu and we can simply assert the case. The thread is
1462 * assigned to the new cpu and enqueued.
1464 * The thread will re-add itself to tdallq when it resumes execution.
1467 lwkt_setcpu_remote(void *arg)
1470 globaldata_t gd = mycpu;
1472 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1475 td->td_flags &= ~TDF_MIGRATING;
1476 KKASSERT(td->td_migrate_gd == NULL);
1477 KKASSERT(td->td_lwp == NULL ||
1478 (td->td_lwp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1483 lwkt_preempted_proc(void)
1485 thread_t td = curthread;
1486 while (td->td_preempted)
1487 td = td->td_preempted;
1492 * Create a kernel process/thread/whatever. It shares it's address space
1493 * with proc0 - ie: kernel only.
1495 * If the cpu is not specified one will be selected. In the future
1496 * specifying a cpu of -1 will enable kernel thread migration between
1500 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1501 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1506 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1510 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1513 * Set up arg0 for 'ps' etc
1515 __va_start(ap, fmt);
1516 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1520 * Schedule the thread to run
1522 if (td->td_flags & TDF_NOSTART)
1523 td->td_flags &= ~TDF_NOSTART;
1530 * Destroy an LWKT thread. Warning! This function is not called when
1531 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1532 * uses a different reaping mechanism.
1537 thread_t td = curthread;
1542 * Do any cleanup that might block here
1544 if (td->td_flags & TDF_VERBOSE)
1545 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1547 dsched_exit_thread(td);
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.
1556 * We are the current thread so of course our own TDF_RUNNING bit will
1557 * be set, so unlike the lwp reap code we don't wait for it to clear.
1560 crit_enter_quick(td);
1563 tsleep(td, 0, "tdreap", 1);
1566 if ((std = gd->gd_freetd) != NULL) {
1567 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1568 gd->gd_freetd = NULL;
1569 objcache_put(thread_cache, std);
1576 * Remove thread resources from kernel lists and deschedule us for
1577 * the last time. We cannot block after this point or we may end
1578 * up with a stale td on the tsleepq.
1580 * None of this may block, the critical section is the only thing
1581 * protecting tdallq and the only thing preventing new lwkt_hold()
1584 if (td->td_flags & TDF_TSLEEPQ)
1586 lwkt_deschedule_self(td);
1587 lwkt_remove_tdallq(td);
1588 KKASSERT(td->td_refs == 0);
1593 KKASSERT(gd->gd_freetd == NULL);
1594 if (td->td_flags & TDF_ALLOCATED_THREAD)
1600 lwkt_remove_tdallq(thread_t td)
1602 KKASSERT(td->td_gd == mycpu);
1603 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1607 * Code reduction and branch prediction improvements. Call/return
1608 * overhead on modern cpus often degenerates into 0 cycles due to
1609 * the cpu's branch prediction hardware and return pc cache. We
1610 * can take advantage of this by not inlining medium-complexity
1611 * functions and we can also reduce the branch prediction impact
1612 * by collapsing perfectly predictable branches into a single
1613 * procedure instead of duplicating it.
1615 * Is any of this noticeable? Probably not, so I'll take the
1616 * smaller code size.
1619 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1621 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1627 thread_t td = curthread;
1628 int lcrit = td->td_critcount;
1630 td->td_critcount = 0;
1631 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1636 * Called from debugger/panic on cpus which have been stopped. We must still
1637 * process the IPIQ while stopped.
1639 * If we are dumping also try to process any pending interrupts. This may
1640 * or may not work depending on the state of the cpu at the point it was
1644 lwkt_smp_stopped(void)
1646 globaldata_t gd = mycpu;
1649 lwkt_process_ipiq();
1650 --gd->gd_intr_nesting_level;
1652 ++gd->gd_intr_nesting_level;
1654 lwkt_process_ipiq();