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);
111 * We can make all thread ports use the spin backend instead of the thread
112 * backend. This should only be set to debug the spin backend.
114 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
117 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
118 "Panic if attempting to switch lwkt's while mastering cpusync");
120 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
121 "Number of switched threads");
122 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
123 "Successful preemption events");
124 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
125 "Failed preemption events");
126 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
127 "Number of preempted threads.");
129 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
130 &token_contention_count, 0, "spinning due to token contention");
132 static int fairq_enable = 1;
133 SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW, &fairq_enable, 0,
134 "Turn on fairq priority accumulators");
135 static int user_pri_sched = 1;
136 SYSCTL_INT(_lwkt, OID_AUTO, user_pri_sched, CTLFLAG_RW, &user_pri_sched, 0,
138 static int preempt_enable = 1;
139 SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW, &preempt_enable, 0,
140 "Enable preemption");
144 * These helper procedures handle the runq, they can only be called from
145 * within a critical section.
147 * WARNING! Prior to SMP being brought up it is possible to enqueue and
148 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
149 * instead of 'mycpu' when referencing the globaldata structure. Once
150 * SMP live enqueuing and dequeueing only occurs on the current cpu.
154 _lwkt_dequeue(thread_t td)
156 if (td->td_flags & TDF_RUNQ) {
157 struct globaldata *gd = td->td_gd;
159 td->td_flags &= ~TDF_RUNQ;
160 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
161 gd->gd_fairq_total_pri -= td->td_pri;
162 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
163 atomic_clear_int_nonlocked(&gd->gd_reqflags, RQF_RUNNING);
170 * NOTE: There are a limited number of lwkt threads runnable since user
171 * processes only schedule one at a time per cpu.
175 _lwkt_enqueue(thread_t td)
179 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
180 struct globaldata *gd = td->td_gd;
182 td->td_flags |= TDF_RUNQ;
183 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
185 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
186 atomic_set_int_nonlocked(&gd->gd_reqflags,
187 RQF_RUNNING | RQF_WAKEUP);
189 atomic_set_int_nonlocked(&gd->gd_reqflags, RQF_WAKEUP);
190 while (xtd && xtd->td_pri > td->td_pri)
191 xtd = TAILQ_NEXT(xtd, td_threadq);
193 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
195 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
197 gd->gd_fairq_total_pri += td->td_pri;
202 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
204 struct thread *td = (struct thread *)obj;
206 td->td_kstack = NULL;
207 td->td_kstack_size = 0;
208 td->td_flags = TDF_ALLOCATED_THREAD;
213 _lwkt_thread_dtor(void *obj, void *privdata)
215 struct thread *td = (struct thread *)obj;
217 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
218 ("_lwkt_thread_dtor: not allocated from objcache"));
219 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
220 td->td_kstack_size > 0,
221 ("_lwkt_thread_dtor: corrupted stack"));
222 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
226 * Initialize the lwkt s/system.
231 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
232 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread),
233 NULL, CACHE_NTHREADS/2,
234 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
238 * Schedule a thread to run. As the current thread we can always safely
239 * schedule ourselves, and a shortcut procedure is provided for that
242 * (non-blocking, self contained on a per cpu basis)
245 lwkt_schedule_self(thread_t td)
247 crit_enter_quick(td);
248 KASSERT(td != &td->td_gd->gd_idlethread,
249 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
250 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
256 * Deschedule a thread.
258 * (non-blocking, self contained on a per cpu basis)
261 lwkt_deschedule_self(thread_t td)
263 crit_enter_quick(td);
269 * LWKTs operate on a per-cpu basis
271 * WARNING! Called from early boot, 'mycpu' may not work yet.
274 lwkt_gdinit(struct globaldata *gd)
276 TAILQ_INIT(&gd->gd_tdrunq);
277 TAILQ_INIT(&gd->gd_tdallq);
281 * Create a new thread. The thread must be associated with a process context
282 * or LWKT start address before it can be scheduled. If the target cpu is
283 * -1 the thread will be created on the current cpu.
285 * If you intend to create a thread without a process context this function
286 * does everything except load the startup and switcher function.
289 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
291 globaldata_t gd = mycpu;
295 * If static thread storage is not supplied allocate a thread. Reuse
296 * a cached free thread if possible. gd_freetd is used to keep an exiting
297 * thread intact through the exit.
301 if ((td = gd->gd_freetd) != NULL) {
302 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
304 gd->gd_freetd = NULL;
306 td = objcache_get(thread_cache, M_WAITOK);
307 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
311 KASSERT((td->td_flags &
312 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
313 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
314 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
318 * Try to reuse cached stack.
320 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
321 if (flags & TDF_ALLOCATED_STACK) {
322 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
327 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
328 flags |= TDF_ALLOCATED_STACK;
331 lwkt_init_thread(td, stack, stksize, flags, gd);
333 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
338 * Initialize a preexisting thread structure. This function is used by
339 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
341 * All threads start out in a critical section at a priority of
342 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
343 * appropriate. This function may send an IPI message when the
344 * requested cpu is not the current cpu and consequently gd_tdallq may
345 * not be initialized synchronously from the point of view of the originating
348 * NOTE! we have to be careful in regards to creating threads for other cpus
349 * if SMP has not yet been activated.
354 lwkt_init_thread_remote(void *arg)
359 * Protected by critical section held by IPI dispatch
361 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
367 * lwkt core thread structural initialization.
369 * NOTE: All threads are initialized as mpsafe threads.
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;
385 if (lwkt_use_spin_port)
386 lwkt_initport_spin(&td->td_msgport);
388 lwkt_initport_thread(&td->td_msgport, td);
389 pmap_init_thread(td);
392 * Normally initializing a thread for a remote cpu requires sending an
393 * IPI. However, the idlethread is setup before the other cpus are
394 * activated so we have to treat it as a special case. XXX manipulation
395 * of gd_tdallq requires the BGL.
397 if (gd == mygd || td == &gd->gd_idlethread) {
399 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
402 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
406 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
410 dsched_new_thread(td);
414 lwkt_set_comm(thread_t td, const char *ctl, ...)
419 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
421 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
425 lwkt_hold(thread_t td)
431 lwkt_rele(thread_t td)
433 KKASSERT(td->td_refs > 0);
438 lwkt_wait_free(thread_t td)
441 tsleep(td, 0, "tdreap", hz);
445 lwkt_free_thread(thread_t td)
447 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
448 if (td->td_flags & TDF_ALLOCATED_THREAD) {
449 objcache_put(thread_cache, td);
450 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
451 /* client-allocated struct with internally allocated stack */
452 KASSERT(td->td_kstack && td->td_kstack_size > 0,
453 ("lwkt_free_thread: corrupted stack"));
454 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
455 td->td_kstack = NULL;
456 td->td_kstack_size = 0;
458 KTR_LOG(ctxsw_deadtd, td);
463 * Switch to the next runnable lwkt. If no LWKTs are runnable then
464 * switch to the idlethread. Switching must occur within a critical
465 * section to avoid races with the scheduling queue.
467 * We always have full control over our cpu's run queue. Other cpus
468 * that wish to manipulate our queue must use the cpu_*msg() calls to
469 * talk to our cpu, so a critical section is all that is needed and
470 * the result is very, very fast thread switching.
472 * The LWKT scheduler uses a fixed priority model and round-robins at
473 * each priority level. User process scheduling is a totally
474 * different beast and LWKT priorities should not be confused with
475 * user process priorities.
477 * Note that the td_switch() function cannot do anything that requires
478 * the MP lock since the MP lock will have already been setup for
479 * the target thread (not the current thread). It's nice to have a scheduler
480 * that does not need the MP lock to work because it allows us to do some
481 * really cool high-performance MP lock optimizations.
483 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
484 * is not called by the current thread in the preemption case, only when
485 * the preempting thread blocks (in order to return to the original thread).
490 globaldata_t gd = mycpu;
491 thread_t td = gd->gd_curthread;
498 * Switching from within a 'fast' (non thread switched) interrupt or IPI
499 * is illegal. However, we may have to do it anyway if we hit a fatal
500 * kernel trap or we have paniced.
502 * If this case occurs save and restore the interrupt nesting level.
504 if (gd->gd_intr_nesting_level) {
508 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
509 panic("lwkt_switch: Attempt to switch from a "
510 "a fast interrupt, ipi, or hard code section, "
514 savegdnest = gd->gd_intr_nesting_level;
515 savegdtrap = gd->gd_trap_nesting_level;
516 gd->gd_intr_nesting_level = 0;
517 gd->gd_trap_nesting_level = 0;
518 if ((td->td_flags & TDF_PANICWARN) == 0) {
519 td->td_flags |= TDF_PANICWARN;
520 kprintf("Warning: thread switch from interrupt, IPI, "
521 "or hard code section.\n"
522 "thread %p (%s)\n", td, td->td_comm);
526 gd->gd_intr_nesting_level = savegdnest;
527 gd->gd_trap_nesting_level = savegdtrap;
533 * Passive release (used to transition from user to kernel mode
534 * when we block or switch rather then when we enter the kernel).
535 * This function is NOT called if we are switching into a preemption
536 * or returning from a preemption. Typically this causes us to lose
537 * our current process designation (if we have one) and become a true
538 * LWKT thread, and may also hand the current process designation to
539 * another process and schedule thread.
545 if (TD_TOKS_HELD(td))
546 lwkt_relalltokens(td);
549 * We had better not be holding any spin locks, but don't get into an
550 * endless panic loop.
552 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
553 ("lwkt_switch: still holding %d exclusive spinlocks!",
554 gd->gd_spinlocks_wr));
559 if (td->td_cscount) {
560 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
562 if (panic_on_cscount)
563 panic("switching while mastering cpusync");
569 * If we had preempted another thread on this cpu, resume the preempted
570 * thread. This occurs transparently, whether the preempted thread
571 * was scheduled or not (it may have been preempted after descheduling
574 * We have to setup the MP lock for the original thread after backing
575 * out the adjustment that was made to curthread when the original
578 if ((ntd = td->td_preempted) != NULL) {
579 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
580 ntd->td_flags |= TDF_PREEMPT_DONE;
583 * The interrupt may have woken a thread up, we need to properly
584 * set the reschedule flag if the originally interrupted thread is
585 * at a lower priority.
587 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
588 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
591 /* YYY release mp lock on switchback if original doesn't need it */
592 goto havethread_preempted;
596 * Implement round-robin fairq with priority insertion. The priority
597 * insertion is handled by _lwkt_enqueue()
599 * We have to adjust the MP lock for the target thread. If we
600 * need the MP lock and cannot obtain it we try to locate a
601 * thread that does not need the MP lock. If we cannot, we spin
604 * A similar issue exists for the tokens held by the target thread.
605 * If we cannot obtain ownership of the tokens we cannot immediately
606 * schedule the thread.
609 atomic_clear_int_nonlocked(&mycpu->gd_reqflags, RQF_WAKEUP);
610 clear_lwkt_resched();
612 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
615 * Hotpath if we can get all necessary resources.
617 * If nothing is runnable switch to the idle thread
620 ntd = &gd->gd_idlethread;
622 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
623 ASSERT_NO_TOKENS_HELD(ntd);
625 cpu_time.cp_msg[0] = 0;
626 cpu_time.cp_stallpc = 0;
633 * NOTE: For UP there is no mplock and lwkt_getalltokens()
636 if (ntd->td_fairq_accum >= 0 &&
637 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
643 * Coldpath - unable to schedule ntd, continue looking for threads
650 * If the fair-share scheduler ran out ntd gets moved to the
651 * end and its accumulator will be bumped, if it didn't we
652 * maintain the same queue position.
654 * nlast keeps track of the last element prior to any moves.
656 if (ntd->td_fairq_accum < 0) {
657 lwkt_fairq_accumulate(gd, ntd);
663 xtd = TAILQ_NEXT(ntd, td_threadq);
664 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
665 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
668 * Set terminal element (nlast)
677 ntd = TAILQ_NEXT(ntd, td_threadq);
681 * If we exhausted the run list switch to the idle thread.
683 * NOTE: nlast can be NULL.
687 ntd = &gd->gd_idlethread;
689 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
690 ASSERT_NO_TOKENS_HELD(ntd);
691 /* contention case, do not clear contention mask */
695 * If fairq accumulations occured we do not schedule the
696 * idle thread. This will cause us to try again from
700 break; /* try again from the top, almost */
705 * Try to switch to this thread.
707 * NOTE: For UP there is no mplock and lwkt_getalltokens()
710 if (ntd->td_fairq_accum >= 0 &&
711 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
717 * Thread was runnable but we were unable to get the required
718 * resources (tokens and/or mplock), continue the scan.
724 * All threads exhausted but we can loop due to a negative
727 * While we are looping in the scheduler be sure to service
728 * any interrupts which were made pending due to our critical
729 * section, otherwise we could livelock (e.g.) IPIs.
735 * We must always decrement td_fairq_accum on non-idle threads just
736 * in case a thread never gets a tick due to being in a continuous
737 * critical section. The page-zeroing code does that.
739 * If the thread we came up with is a higher or equal priority verses
740 * the thread at the head of the queue we move our thread to the
741 * front. This way we can always check the front of the queue.
744 ++gd->gd_cnt.v_swtch;
745 --ntd->td_fairq_accum;
746 ntd->td_wmesg = NULL;
747 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
748 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
749 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
750 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
752 havethread_preempted:
755 * If the new target does not need the MP lock and we are holding it,
756 * release the MP lock. If the new target requires the MP lock we have
757 * already acquired it for the target.
760 KASSERT(ntd->td_critcount,
761 ("priority problem in lwkt_switch %d %d",
762 td->td_critcount, ntd->td_critcount));
766 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
769 /* NOTE: current cpu may have changed after switch */
774 * Request that the target thread preempt the current thread. Preemption
775 * only works under a specific set of conditions:
777 * - We are not preempting ourselves
778 * - The target thread is owned by the current cpu
779 * - We are not currently being preempted
780 * - The target is not currently being preempted
781 * - We are not holding any spin locks
782 * - The target thread is not holding any tokens
783 * - We are able to satisfy the target's MP lock requirements (if any).
785 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
786 * this is called via lwkt_schedule() through the td_preemptable callback.
787 * critcount is the managed critical priority that we should ignore in order
788 * to determine whether preemption is possible (aka usually just the crit
789 * priority of lwkt_schedule() itself).
791 * XXX at the moment we run the target thread in a critical section during
792 * the preemption in order to prevent the target from taking interrupts
793 * that *WE* can't. Preemption is strictly limited to interrupt threads
794 * and interrupt-like threads, outside of a critical section, and the
795 * preempted source thread will be resumed the instant the target blocks
796 * whether or not the source is scheduled (i.e. preemption is supposed to
797 * be as transparent as possible).
800 lwkt_preempt(thread_t ntd, int critcount)
802 struct globaldata *gd = mycpu;
804 int save_gd_intr_nesting_level;
807 * The caller has put us in a critical section. We can only preempt
808 * if the caller of the caller was not in a critical section (basically
809 * a local interrupt), as determined by the 'critcount' parameter. We
810 * also can't preempt if the caller is holding any spinlocks (even if
811 * he isn't in a critical section). This also handles the tokens test.
813 * YYY The target thread must be in a critical section (else it must
814 * inherit our critical section? I dunno yet).
816 * Set need_lwkt_resched() unconditionally for now YYY.
818 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
820 if (preempt_enable == 0) {
825 td = gd->gd_curthread;
826 if (ntd->td_pri <= td->td_pri) {
830 if (td->td_critcount > critcount) {
836 if (ntd->td_gd != gd) {
843 * We don't have to check spinlocks here as they will also bump
846 * Do not try to preempt if the target thread is holding any tokens.
847 * We could try to acquire the tokens but this case is so rare there
848 * is no need to support it.
850 KKASSERT(gd->gd_spinlocks_wr == 0);
852 if (TD_TOKS_HELD(ntd)) {
857 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
862 if (ntd->td_preempted) {
869 * Since we are able to preempt the current thread, there is no need to
870 * call need_lwkt_resched().
872 * We must temporarily clear gd_intr_nesting_level around the switch
873 * since switchouts from the target thread are allowed (they will just
874 * return to our thread), and since the target thread has its own stack.
877 ntd->td_preempted = td;
878 td->td_flags |= TDF_PREEMPT_LOCK;
879 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
880 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
881 gd->gd_intr_nesting_level = 0;
883 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
885 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
886 ntd->td_preempted = NULL;
887 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
891 * Conditionally call splz() if gd_reqflags indicates work is pending.
892 * This will work inside a critical section but not inside a hard code
895 * (self contained on a per cpu basis)
900 globaldata_t gd = mycpu;
901 thread_t td = gd->gd_curthread;
903 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
904 gd->gd_intr_nesting_level == 0 &&
905 td->td_nest_count < 2)
912 * This version is integrated into crit_exit, reqflags has already
913 * been tested but td_critcount has not.
915 * We only want to execute the splz() on the 1->0 transition of
916 * critcount and not in a hard code section or if too deeply nested.
919 lwkt_maybe_splz(thread_t td)
921 globaldata_t gd = td->td_gd;
923 if (td->td_critcount == 0 &&
924 gd->gd_intr_nesting_level == 0 &&
925 td->td_nest_count < 2)
932 * This function is used to negotiate a passive release of the current
933 * process/lwp designation with the user scheduler, allowing the user
934 * scheduler to schedule another user thread. The related kernel thread
935 * (curthread) continues running in the released state.
938 lwkt_passive_release(struct thread *td)
940 struct lwp *lp = td->td_lwp;
942 td->td_release = NULL;
943 lwkt_setpri_self(TDPRI_KERN_USER);
944 lp->lwp_proc->p_usched->release_curproc(lp);
949 * This implements a normal yield. This routine is virtually a nop if
950 * there is nothing to yield to but it will always run any pending interrupts
951 * if called from a critical section.
953 * This yield is designed for kernel threads without a user context.
955 * (self contained on a per cpu basis)
960 globaldata_t gd = mycpu;
961 thread_t td = gd->gd_curthread;
964 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
966 if (td->td_fairq_accum < 0) {
967 lwkt_schedule_self(curthread);
970 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
971 if (xtd && xtd->td_pri > td->td_pri) {
972 lwkt_schedule_self(curthread);
979 * This yield is designed for kernel threads with a user context.
981 * The kernel acting on behalf of the user is potentially cpu-bound,
982 * this function will efficiently allow other threads to run and also
983 * switch to other processes by releasing.
985 * The lwkt_user_yield() function is designed to have very low overhead
986 * if no yield is determined to be needed.
989 lwkt_user_yield(void)
991 globaldata_t gd = mycpu;
992 thread_t td = gd->gd_curthread;
995 * Always run any pending interrupts in case we are in a critical
998 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1002 * Switch (which forces a release) if another kernel thread needs
1003 * the cpu, if userland wants us to resched, or if our kernel
1004 * quantum has run out.
1006 if (lwkt_resched_wanted() ||
1007 user_resched_wanted() ||
1008 td->td_fairq_accum < 0)
1015 * Reacquire the current process if we are released.
1017 * XXX not implemented atm. The kernel may be holding locks and such,
1018 * so we want the thread to continue to receive cpu.
1020 if (td->td_release == NULL && lp) {
1021 lp->lwp_proc->p_usched->acquire_curproc(lp);
1022 td->td_release = lwkt_passive_release;
1023 lwkt_setpri_self(TDPRI_USER_NORM);
1029 * Generic schedule. Possibly schedule threads belonging to other cpus and
1030 * deal with threads that might be blocked on a wait queue.
1032 * We have a little helper inline function which does additional work after
1033 * the thread has been enqueued, including dealing with preemption and
1034 * setting need_lwkt_resched() (which prevents the kernel from returning
1035 * to userland until it has processed higher priority threads).
1037 * It is possible for this routine to be called after a failed _enqueue
1038 * (due to the target thread migrating, sleeping, or otherwise blocked).
1039 * We have to check that the thread is actually on the run queue!
1041 * reschedok is an optimized constant propagated from lwkt_schedule() or
1042 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1043 * reschedule to be requested if the target thread has a higher priority.
1044 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1045 * be 0, prevented undesired reschedules.
1049 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1053 if (ntd->td_flags & TDF_RUNQ) {
1054 if (ntd->td_preemptable && reschedok) {
1055 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1056 } else if (reschedok) {
1058 if (ntd->td_pri > otd->td_pri)
1059 need_lwkt_resched();
1063 * Give the thread a little fair share scheduler bump if it
1064 * has been asleep for a while. This is primarily to avoid
1065 * a degenerate case for interrupt threads where accumulator
1066 * crosses into negative territory unnecessarily.
1068 if (ntd->td_fairq_lticks != ticks) {
1069 ntd->td_fairq_lticks = ticks;
1070 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1071 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1072 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1079 _lwkt_schedule(thread_t td, int reschedok)
1081 globaldata_t mygd = mycpu;
1083 KASSERT(td != &td->td_gd->gd_idlethread,
1084 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1085 crit_enter_gd(mygd);
1086 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1087 if (td == mygd->gd_curthread) {
1091 * If we own the thread, there is no race (since we are in a
1092 * critical section). If we do not own the thread there might
1093 * be a race but the target cpu will deal with it.
1096 if (td->td_gd == mygd) {
1098 _lwkt_schedule_post(mygd, td, 1, reschedok);
1100 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1104 _lwkt_schedule_post(mygd, td, 1, reschedok);
1111 lwkt_schedule(thread_t td)
1113 _lwkt_schedule(td, 1);
1117 lwkt_schedule_noresched(thread_t td)
1119 _lwkt_schedule(td, 0);
1125 * When scheduled remotely if frame != NULL the IPIQ is being
1126 * run via doreti or an interrupt then preemption can be allowed.
1128 * To allow preemption we have to drop the critical section so only
1129 * one is present in _lwkt_schedule_post.
1132 lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1134 thread_t td = curthread;
1137 if (frame && ntd->td_preemptable) {
1138 crit_exit_noyield(td);
1139 _lwkt_schedule(ntd, 1);
1140 crit_enter_quick(td);
1142 _lwkt_schedule(ntd, 1);
1147 * Thread migration using a 'Pull' method. The thread may or may not be
1148 * the current thread. It MUST be descheduled and in a stable state.
1149 * lwkt_giveaway() must be called on the cpu owning the thread.
1151 * At any point after lwkt_giveaway() is called, the target cpu may
1152 * 'pull' the thread by calling lwkt_acquire().
1154 * We have to make sure the thread is not sitting on a per-cpu tsleep
1155 * queue or it will blow up when it moves to another cpu.
1157 * MPSAFE - must be called under very specific conditions.
1160 lwkt_giveaway(thread_t td)
1162 globaldata_t gd = mycpu;
1165 if (td->td_flags & TDF_TSLEEPQ)
1167 KKASSERT(td->td_gd == gd);
1168 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1169 td->td_flags |= TDF_MIGRATING;
1174 lwkt_acquire(thread_t td)
1179 KKASSERT(td->td_flags & TDF_MIGRATING);
1184 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1185 crit_enter_gd(mygd);
1186 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1188 lwkt_process_ipiq();
1194 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1195 td->td_flags &= ~TDF_MIGRATING;
1198 crit_enter_gd(mygd);
1199 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1200 td->td_flags &= ~TDF_MIGRATING;
1208 * Generic deschedule. Descheduling threads other then your own should be
1209 * done only in carefully controlled circumstances. Descheduling is
1212 * This function may block if the cpu has run out of messages.
1215 lwkt_deschedule(thread_t td)
1219 if (td == curthread) {
1222 if (td->td_gd == mycpu) {
1225 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1235 * Set the target thread's priority. This routine does not automatically
1236 * switch to a higher priority thread, LWKT threads are not designed for
1237 * continuous priority changes. Yield if you want to switch.
1240 lwkt_setpri(thread_t td, int pri)
1242 KKASSERT(td->td_gd == mycpu);
1243 if (td->td_pri != pri) {
1246 if (td->td_flags & TDF_RUNQ) {
1258 * Set the initial priority for a thread prior to it being scheduled for
1259 * the first time. The thread MUST NOT be scheduled before or during
1260 * this call. The thread may be assigned to a cpu other then the current
1263 * Typically used after a thread has been created with TDF_STOPPREQ,
1264 * and before the thread is initially scheduled.
1267 lwkt_setpri_initial(thread_t td, int pri)
1270 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1275 lwkt_setpri_self(int pri)
1277 thread_t td = curthread;
1279 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1281 if (td->td_flags & TDF_RUNQ) {
1292 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1294 * Example: two competing threads, same priority N. decrement by (2*N)
1295 * increment by N*8, each thread will get 4 ticks.
1298 lwkt_fairq_schedulerclock(thread_t td)
1302 if (td != &td->td_gd->gd_idlethread) {
1303 td->td_fairq_accum -= td->td_gd->gd_fairq_total_pri;
1304 if (td->td_fairq_accum < -TDFAIRQ_MAX(td->td_gd))
1305 td->td_fairq_accum = -TDFAIRQ_MAX(td->td_gd);
1306 if (td->td_fairq_accum < 0)
1307 need_lwkt_resched();
1308 td->td_fairq_lticks = ticks;
1310 td = td->td_preempted;
1316 lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1318 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1319 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1320 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1324 * Migrate the current thread to the specified cpu.
1326 * This is accomplished by descheduling ourselves from the current cpu,
1327 * moving our thread to the tdallq of the target cpu, IPI messaging the
1328 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1329 * races while the thread is being migrated.
1331 * We must be sure to remove ourselves from the current cpu's tsleepq
1332 * before potentially moving to another queue. The thread can be on
1333 * a tsleepq due to a left-over tsleep_interlock().
1336 static void lwkt_setcpu_remote(void *arg);
1340 lwkt_setcpu_self(globaldata_t rgd)
1343 thread_t td = curthread;
1345 if (td->td_gd != rgd) {
1346 crit_enter_quick(td);
1347 if (td->td_flags & TDF_TSLEEPQ)
1349 td->td_flags |= TDF_MIGRATING;
1350 lwkt_deschedule_self(td);
1351 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1352 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1354 /* we are now on the target cpu */
1355 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1356 crit_exit_quick(td);
1362 lwkt_migratecpu(int cpuid)
1367 rgd = globaldata_find(cpuid);
1368 lwkt_setcpu_self(rgd);
1373 * Remote IPI for cpu migration (called while in a critical section so we
1374 * do not have to enter another one). The thread has already been moved to
1375 * our cpu's allq, but we must wait for the thread to be completely switched
1376 * out on the originating cpu before we schedule it on ours or the stack
1377 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1378 * change to main memory.
1380 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1381 * against wakeups. It is best if this interface is used only when there
1382 * are no pending events that might try to schedule the thread.
1386 lwkt_setcpu_remote(void *arg)
1389 globaldata_t gd = mycpu;
1391 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1393 lwkt_process_ipiq();
1400 td->td_flags &= ~TDF_MIGRATING;
1401 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1407 lwkt_preempted_proc(void)
1409 thread_t td = curthread;
1410 while (td->td_preempted)
1411 td = td->td_preempted;
1416 * Create a kernel process/thread/whatever. It shares it's address space
1417 * with proc0 - ie: kernel only.
1419 * NOTE! By default new threads are created with the MP lock held. A
1420 * thread which does not require the MP lock should release it by calling
1421 * rel_mplock() at the start of the new thread.
1424 lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1425 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1430 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1434 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1437 * Set up arg0 for 'ps' etc
1439 __va_start(ap, fmt);
1440 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1444 * Schedule the thread to run
1446 if ((td->td_flags & TDF_STOPREQ) == 0)
1449 td->td_flags &= ~TDF_STOPREQ;
1454 * Destroy an LWKT thread. Warning! This function is not called when
1455 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1456 * uses a different reaping mechanism.
1461 thread_t td = curthread;
1466 * Do any cleanup that might block here
1468 if (td->td_flags & TDF_VERBOSE)
1469 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1472 dsched_exit_thread(td);
1475 * Get us into a critical section to interlock gd_freetd and loop
1476 * until we can get it freed.
1478 * We have to cache the current td in gd_freetd because objcache_put()ing
1479 * it would rip it out from under us while our thread is still active.
1482 crit_enter_quick(td);
1483 while ((std = gd->gd_freetd) != NULL) {
1484 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1485 gd->gd_freetd = NULL;
1486 objcache_put(thread_cache, std);
1490 * Remove thread resources from kernel lists and deschedule us for
1491 * the last time. We cannot block after this point or we may end
1492 * up with a stale td on the tsleepq.
1494 if (td->td_flags & TDF_TSLEEPQ)
1496 lwkt_deschedule_self(td);
1497 lwkt_remove_tdallq(td);
1502 KKASSERT(gd->gd_freetd == NULL);
1503 if (td->td_flags & TDF_ALLOCATED_THREAD)
1509 lwkt_remove_tdallq(thread_t td)
1511 KKASSERT(td->td_gd == mycpu);
1512 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1516 * Code reduction and branch prediction improvements. Call/return
1517 * overhead on modern cpus often degenerates into 0 cycles due to
1518 * the cpu's branch prediction hardware and return pc cache. We
1519 * can take advantage of this by not inlining medium-complexity
1520 * functions and we can also reduce the branch prediction impact
1521 * by collapsing perfectly predictable branches into a single
1522 * procedure instead of duplicating it.
1524 * Is any of this noticeable? Probably not, so I'll take the
1525 * smaller code size.
1528 crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1530 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1536 thread_t td = curthread;
1537 int lcrit = td->td_critcount;
1539 td->td_critcount = 0;
1540 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1547 * Called from debugger/panic on cpus which have been stopped. We must still
1548 * process the IPIQ while stopped, even if we were stopped while in a critical
1551 * If we are dumping also try to process any pending interrupts. This may
1552 * or may not work depending on the state of the cpu at the point it was
1556 lwkt_smp_stopped(void)
1558 globaldata_t gd = mycpu;
1562 lwkt_process_ipiq();
1565 lwkt_process_ipiq();