2 * Copyright (c) 2003,2004 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
34 * $DragonFly: src/sys/kern/lwkt_thread.c,v 1.114 2008/05/26 17:11:09 nth Exp $
38 * Each cpu in a system has its own self-contained light weight kernel
39 * thread scheduler, which means that generally speaking we only need
40 * to use a critical section to avoid problems. Foreign thread
41 * scheduling is queued via (async) IPIs.
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
48 #include <sys/rtprio.h>
49 #include <sys/queue.h>
50 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
55 #include <sys/spinlock.h>
58 #include <sys/thread2.h>
59 #include <sys/spinlock2.h>
62 #include <vm/vm_param.h>
63 #include <vm/vm_kern.h>
64 #include <vm/vm_object.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_pager.h>
68 #include <vm/vm_extern.h>
70 #include <machine/stdarg.h>
71 #include <machine/smp.h>
73 static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
75 static int untimely_switch = 0;
77 static int panic_on_cscount = 0;
79 static __int64_t switch_count = 0;
80 static __int64_t preempt_hit = 0;
81 static __int64_t preempt_miss = 0;
82 static __int64_t preempt_weird = 0;
83 static __int64_t token_contention_count = 0;
84 static __int64_t mplock_contention_count = 0;
85 static int lwkt_use_spin_port;
86 static struct objcache *thread_cache;
89 * We can make all thread ports use the spin backend instead of the thread
90 * backend. This should only be set to debug the spin backend.
92 TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
94 SYSCTL_INT(_lwkt, OID_AUTO, untimely_switch, CTLFLAG_RW, &untimely_switch, 0, "");
96 SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0, "");
98 SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0, "");
99 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0, "");
100 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0, "");
101 SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0, "");
103 SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
104 &token_contention_count, 0, "spinning due to token contention");
105 SYSCTL_QUAD(_lwkt, OID_AUTO, mplock_contention_count, CTLFLAG_RW,
106 &mplock_contention_count, 0, "spinning due to MPLOCK contention");
112 #if !defined(KTR_GIANT_CONTENTION)
113 #define KTR_GIANT_CONTENTION KTR_ALL
116 KTR_INFO_MASTER(giant);
117 KTR_INFO(KTR_GIANT_CONTENTION, giant, beg, 0, "thread=%p", sizeof(void *));
118 KTR_INFO(KTR_GIANT_CONTENTION, giant, end, 1, "thread=%p", sizeof(void *));
120 #define loggiant(name) KTR_LOG(giant_ ## name, curthread)
123 * These helper procedures handle the runq, they can only be called from
124 * within a critical section.
126 * WARNING! Prior to SMP being brought up it is possible to enqueue and
127 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
128 * instead of 'mycpu' when referencing the globaldata structure. Once
129 * SMP live enqueuing and dequeueing only occurs on the current cpu.
133 _lwkt_dequeue(thread_t td)
135 if (td->td_flags & TDF_RUNQ) {
136 int nq = td->td_pri & TDPRI_MASK;
137 struct globaldata *gd = td->td_gd;
139 td->td_flags &= ~TDF_RUNQ;
140 TAILQ_REMOVE(&gd->gd_tdrunq[nq], td, td_threadq);
141 /* runqmask is passively cleaned up by the switcher */
147 _lwkt_enqueue(thread_t td)
149 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_TSLEEPQ|TDF_BLOCKQ)) == 0) {
150 int nq = td->td_pri & TDPRI_MASK;
151 struct globaldata *gd = td->td_gd;
153 td->td_flags |= TDF_RUNQ;
154 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], td, td_threadq);
155 gd->gd_runqmask |= 1 << nq;
160 _lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
162 struct thread *td = (struct thread *)obj;
164 td->td_kstack = NULL;
165 td->td_kstack_size = 0;
166 td->td_flags = TDF_ALLOCATED_THREAD;
171 _lwkt_thread_dtor(void *obj, void *privdata)
173 struct thread *td = (struct thread *)obj;
175 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
176 ("_lwkt_thread_dtor: not allocated from objcache"));
177 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
178 td->td_kstack_size > 0,
179 ("_lwkt_thread_dtor: corrupted stack"));
180 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
184 * Initialize the lwkt s/system.
189 /* An objcache has 2 magazines per CPU so divide cache size by 2. */
190 thread_cache = objcache_create_mbacked(M_THREAD, sizeof(struct thread), 0,
191 CACHE_NTHREADS/2, _lwkt_thread_ctor, _lwkt_thread_dtor,
196 * Schedule a thread to run. As the current thread we can always safely
197 * schedule ourselves, and a shortcut procedure is provided for that
200 * (non-blocking, self contained on a per cpu basis)
203 lwkt_schedule_self(thread_t td)
205 crit_enter_quick(td);
206 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
207 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
213 * Deschedule a thread.
215 * (non-blocking, self contained on a per cpu basis)
218 lwkt_deschedule_self(thread_t td)
220 crit_enter_quick(td);
226 * LWKTs operate on a per-cpu basis
228 * WARNING! Called from early boot, 'mycpu' may not work yet.
231 lwkt_gdinit(struct globaldata *gd)
235 for (i = 0; i < sizeof(gd->gd_tdrunq)/sizeof(gd->gd_tdrunq[0]); ++i)
236 TAILQ_INIT(&gd->gd_tdrunq[i]);
238 TAILQ_INIT(&gd->gd_tdallq);
242 * Create a new thread. The thread must be associated with a process context
243 * or LWKT start address before it can be scheduled. If the target cpu is
244 * -1 the thread will be created on the current cpu.
246 * If you intend to create a thread without a process context this function
247 * does everything except load the startup and switcher function.
250 lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
255 td = objcache_get(thread_cache, M_WAITOK);
256 KASSERT((td->td_flags &
257 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
258 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
259 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
261 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
262 if (flags & TDF_ALLOCATED_STACK) {
263 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
268 stack = (void *)kmem_alloc(&kernel_map, stksize);
269 flags |= TDF_ALLOCATED_STACK;
272 lwkt_init_thread(td, stack, stksize, flags, mycpu);
274 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
279 * Initialize a preexisting thread structure. This function is used by
280 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
282 * All threads start out in a critical section at a priority of
283 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
284 * appropriate. This function may send an IPI message when the
285 * requested cpu is not the current cpu and consequently gd_tdallq may
286 * not be initialized synchronously from the point of view of the originating
289 * NOTE! we have to be careful in regards to creating threads for other cpus
290 * if SMP has not yet been activated.
295 lwkt_init_thread_remote(void *arg)
300 * Protected by critical section held by IPI dispatch
302 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
308 lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
309 struct globaldata *gd)
311 globaldata_t mygd = mycpu;
313 bzero(td, sizeof(struct thread));
314 td->td_kstack = stack;
315 td->td_kstack_size = stksize;
316 td->td_flags = flags;
318 td->td_pri = TDPRI_KERN_DAEMON + TDPRI_CRIT;
320 if ((flags & TDF_MPSAFE) == 0)
323 if (lwkt_use_spin_port)
324 lwkt_initport_spin(&td->td_msgport);
326 lwkt_initport_thread(&td->td_msgport, td);
327 pmap_init_thread(td);
330 * Normally initializing a thread for a remote cpu requires sending an
331 * IPI. However, the idlethread is setup before the other cpus are
332 * activated so we have to treat it as a special case. XXX manipulation
333 * of gd_tdallq requires the BGL.
335 if (gd == mygd || td == &gd->gd_idlethread) {
337 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
340 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
344 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
350 lwkt_set_comm(thread_t td, const char *ctl, ...)
355 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
360 lwkt_hold(thread_t td)
366 lwkt_rele(thread_t td)
368 KKASSERT(td->td_refs > 0);
373 lwkt_wait_free(thread_t td)
376 tsleep(td, 0, "tdreap", hz);
380 lwkt_free_thread(thread_t td)
382 KASSERT((td->td_flags & TDF_RUNNING) == 0,
383 ("lwkt_free_thread: did not exit! %p", td));
385 if (td->td_flags & TDF_ALLOCATED_THREAD) {
386 objcache_put(thread_cache, td);
387 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
388 /* client-allocated struct with internally allocated stack */
389 KASSERT(td->td_kstack && td->td_kstack_size > 0,
390 ("lwkt_free_thread: corrupted stack"));
391 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
392 td->td_kstack = NULL;
393 td->td_kstack_size = 0;
399 * Switch to the next runnable lwkt. If no LWKTs are runnable then
400 * switch to the idlethread. Switching must occur within a critical
401 * section to avoid races with the scheduling queue.
403 * We always have full control over our cpu's run queue. Other cpus
404 * that wish to manipulate our queue must use the cpu_*msg() calls to
405 * talk to our cpu, so a critical section is all that is needed and
406 * the result is very, very fast thread switching.
408 * The LWKT scheduler uses a fixed priority model and round-robins at
409 * each priority level. User process scheduling is a totally
410 * different beast and LWKT priorities should not be confused with
411 * user process priorities.
413 * The MP lock may be out of sync with the thread's td_mpcount. lwkt_switch()
414 * cleans it up. Note that the td_switch() function cannot do anything that
415 * requires the MP lock since the MP lock will have already been setup for
416 * the target thread (not the current thread). It's nice to have a scheduler
417 * that does not need the MP lock to work because it allows us to do some
418 * really cool high-performance MP lock optimizations.
420 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
421 * is not called by the current thread in the preemption case, only when
422 * the preempting thread blocks (in order to return to the original thread).
427 globaldata_t gd = mycpu;
428 thread_t td = gd->gd_curthread;
435 * Switching from within a 'fast' (non thread switched) interrupt or IPI
436 * is illegal. However, we may have to do it anyway if we hit a fatal
437 * kernel trap or we have paniced.
439 * If this case occurs save and restore the interrupt nesting level.
441 if (gd->gd_intr_nesting_level) {
445 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
446 panic("lwkt_switch: cannot switch from within "
447 "a fast interrupt, yet, td %p\n", td);
449 savegdnest = gd->gd_intr_nesting_level;
450 savegdtrap = gd->gd_trap_nesting_level;
451 gd->gd_intr_nesting_level = 0;
452 gd->gd_trap_nesting_level = 0;
453 if ((td->td_flags & TDF_PANICWARN) == 0) {
454 td->td_flags |= TDF_PANICWARN;
455 kprintf("Warning: thread switch from interrupt or IPI, "
456 "thread %p (%s)\n", td, td->td_comm);
458 db_print_backtrace();
462 gd->gd_intr_nesting_level = savegdnest;
463 gd->gd_trap_nesting_level = savegdtrap;
469 * Passive release (used to transition from user to kernel mode
470 * when we block or switch rather then when we enter the kernel).
471 * This function is NOT called if we are switching into a preemption
472 * or returning from a preemption. Typically this causes us to lose
473 * our current process designation (if we have one) and become a true
474 * LWKT thread, and may also hand the current process designation to
475 * another process and schedule thread.
482 lwkt_relalltokens(td);
485 * We had better not be holding any spin locks, but don't get into an
486 * endless panic loop.
488 KASSERT(gd->gd_spinlock_rd == NULL || panicstr != NULL,
489 ("lwkt_switch: still holding a shared spinlock %p!",
490 gd->gd_spinlock_rd));
491 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
492 ("lwkt_switch: still holding %d exclusive spinlocks!",
493 gd->gd_spinlocks_wr));
498 * td_mpcount cannot be used to determine if we currently hold the
499 * MP lock because get_mplock() will increment it prior to attempting
500 * to get the lock, and switch out if it can't. Our ownership of
501 * the actual lock will remain stable while we are in a critical section
502 * (but, of course, another cpu may own or release the lock so the
503 * actual value of mp_lock is not stable).
505 mpheld = MP_LOCK_HELD();
507 if (td->td_cscount) {
508 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
510 if (panic_on_cscount)
511 panic("switching while mastering cpusync");
515 if ((ntd = td->td_preempted) != NULL) {
517 * We had preempted another thread on this cpu, resume the preempted
518 * thread. This occurs transparently, whether the preempted thread
519 * was scheduled or not (it may have been preempted after descheduling
522 * We have to setup the MP lock for the original thread after backing
523 * out the adjustment that was made to curthread when the original
526 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
528 if (ntd->td_mpcount && mpheld == 0) {
529 panic("MPLOCK NOT HELD ON RETURN: %p %p %d %d",
530 td, ntd, td->td_mpcount, ntd->td_mpcount);
532 if (ntd->td_mpcount) {
533 td->td_mpcount -= ntd->td_mpcount;
534 KKASSERT(td->td_mpcount >= 0);
537 ntd->td_flags |= TDF_PREEMPT_DONE;
540 * XXX. The interrupt may have woken a thread up, we need to properly
541 * set the reschedule flag if the originally interrupted thread is at
544 if (gd->gd_runqmask > (2 << (ntd->td_pri & TDPRI_MASK)) - 1)
546 /* YYY release mp lock on switchback if original doesn't need it */
549 * Priority queue / round-robin at each priority. Note that user
550 * processes run at a fixed, low priority and the user process
551 * scheduler deals with interactions between user processes
552 * by scheduling and descheduling them from the LWKT queue as
555 * We have to adjust the MP lock for the target thread. If we
556 * need the MP lock and cannot obtain it we try to locate a
557 * thread that does not need the MP lock. If we cannot, we spin
560 * A similar issue exists for the tokens held by the target thread.
561 * If we cannot obtain ownership of the tokens we cannot immediately
562 * schedule the thread.
566 * If an LWKT reschedule was requested, well that is what we are
567 * doing now so clear it.
569 clear_lwkt_resched();
571 if (gd->gd_runqmask) {
572 int nq = bsrl(gd->gd_runqmask);
573 if ((ntd = TAILQ_FIRST(&gd->gd_tdrunq[nq])) == NULL) {
574 gd->gd_runqmask &= ~(1 << nq);
579 * THREAD SELECTION FOR AN SMP MACHINE BUILD
581 * If the target needs the MP lock and we couldn't get it,
582 * or if the target is holding tokens and we could not
583 * gain ownership of the tokens, continue looking for a
584 * thread to schedule and spin instead of HLT if we can't.
586 * NOTE: the mpheld variable invalid after this conditional, it
587 * can change due to both cpu_try_mplock() returning success
588 * AND interactions in lwkt_getalltokens() due to the fact that
589 * we are trying to check the mpcount of a thread other then
590 * the current thread. Because of this, if the current thread
591 * is not holding td_mpcount, an IPI indirectly run via
592 * lwkt_getalltokens() can obtain and release the MP lock and
593 * cause the core MP lock to be released.
595 if ((ntd->td_mpcount && mpheld == 0 && !cpu_try_mplock()) ||
596 (ntd->td_toks && lwkt_getalltokens(ntd) == 0)
598 u_int32_t rqmask = gd->gd_runqmask;
600 mpheld = MP_LOCK_HELD();
603 TAILQ_FOREACH(ntd, &gd->gd_tdrunq[nq], td_threadq) {
604 if (ntd->td_mpcount && !mpheld && !cpu_try_mplock()) {
605 /* spinning due to MP lock being held */
607 ++mplock_contention_count;
609 /* mplock still not held, 'mpheld' still valid */
614 * mpheld state invalid after getalltokens call returns
615 * failure, but the variable is only needed for
618 if (ntd->td_toks && !lwkt_getalltokens(ntd)) {
619 /* spinning due to token contention */
621 ++token_contention_count;
623 mpheld = MP_LOCK_HELD();
630 rqmask &= ~(1 << nq);
634 cpu_mplock_contested();
635 ntd = &gd->gd_idlethread;
636 ntd->td_flags |= TDF_IDLE_NOHLT;
637 goto using_idle_thread;
639 ++gd->gd_cnt.v_swtch;
640 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
641 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
644 ++gd->gd_cnt.v_swtch;
645 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
646 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
650 * THREAD SELECTION FOR A UP MACHINE BUILD. We don't have to
651 * worry about tokens or the BGL. However, we still have
652 * to call lwkt_getalltokens() in order to properly detect
653 * stale tokens. This call cannot fail for a UP build!
655 lwkt_getalltokens(ntd);
656 ++gd->gd_cnt.v_swtch;
657 TAILQ_REMOVE(&gd->gd_tdrunq[nq], ntd, td_threadq);
658 TAILQ_INSERT_TAIL(&gd->gd_tdrunq[nq], ntd, td_threadq);
662 * We have nothing to run but only let the idle loop halt
663 * the cpu if there are no pending interrupts.
665 ntd = &gd->gd_idlethread;
666 if (gd->gd_reqflags & RQF_IDLECHECK_MASK)
667 ntd->td_flags |= TDF_IDLE_NOHLT;
671 * The idle thread should not be holding the MP lock unless we
672 * are trapping in the kernel or in a panic. Since we select the
673 * idle thread unconditionally when no other thread is available,
674 * if the MP lock is desired during a panic or kernel trap, we
675 * have to loop in the scheduler until we get it.
677 if (ntd->td_mpcount) {
678 mpheld = MP_LOCK_HELD();
679 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL) {
680 panic("Idle thread %p was holding the BGL!", ntd);
681 } else if (mpheld == 0) {
682 cpu_mplock_contested();
689 KASSERT(ntd->td_pri >= TDPRI_CRIT,
690 ("priority problem in lwkt_switch %d %d", td->td_pri, ntd->td_pri));
693 * Do the actual switch. If the new target does not need the MP lock
694 * and we are holding it, release the MP lock. If the new target requires
695 * the MP lock we have already acquired it for the target.
698 if (ntd->td_mpcount == 0 ) {
702 ASSERT_MP_LOCK_HELD(ntd);
709 /* NOTE: current cpu may have changed after switch */
714 * Request that the target thread preempt the current thread. Preemption
715 * only works under a specific set of conditions:
717 * - We are not preempting ourselves
718 * - The target thread is owned by the current cpu
719 * - We are not currently being preempted
720 * - The target is not currently being preempted
721 * - We are not holding any spin locks
722 * - The target thread is not holding any tokens
723 * - We are able to satisfy the target's MP lock requirements (if any).
725 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
726 * this is called via lwkt_schedule() through the td_preemptable callback.
727 * critpri is the managed critical priority that we should ignore in order
728 * to determine whether preemption is possible (aka usually just the crit
729 * priority of lwkt_schedule() itself).
731 * XXX at the moment we run the target thread in a critical section during
732 * the preemption in order to prevent the target from taking interrupts
733 * that *WE* can't. Preemption is strictly limited to interrupt threads
734 * and interrupt-like threads, outside of a critical section, and the
735 * preempted source thread will be resumed the instant the target blocks
736 * whether or not the source is scheduled (i.e. preemption is supposed to
737 * be as transparent as possible).
739 * The target thread inherits our MP count (added to its own) for the
740 * duration of the preemption in order to preserve the atomicy of the
741 * MP lock during the preemption. Therefore, any preempting targets must be
742 * careful in regards to MP assertions. Note that the MP count may be
743 * out of sync with the physical mp_lock, but we do not have to preserve
744 * the original ownership of the lock if it was out of synch (that is, we
745 * can leave it synchronized on return).
748 lwkt_preempt(thread_t ntd, int critpri)
750 struct globaldata *gd = mycpu;
758 * The caller has put us in a critical section. We can only preempt
759 * if the caller of the caller was not in a critical section (basically
760 * a local interrupt), as determined by the 'critpri' parameter. We
761 * also can't preempt if the caller is holding any spinlocks (even if
762 * he isn't in a critical section). This also handles the tokens test.
764 * YYY The target thread must be in a critical section (else it must
765 * inherit our critical section? I dunno yet).
767 * Set need_lwkt_resched() unconditionally for now YYY.
769 KASSERT(ntd->td_pri >= TDPRI_CRIT, ("BADCRIT0 %d", ntd->td_pri));
771 td = gd->gd_curthread;
772 if ((ntd->td_pri & TDPRI_MASK) <= (td->td_pri & TDPRI_MASK)) {
776 if ((td->td_pri & ~TDPRI_MASK) > critpri) {
782 if (ntd->td_gd != gd) {
789 * Take the easy way out and do not preempt if we are holding
790 * any spinlocks. We could test whether the thread(s) being
791 * preempted interlock against the target thread's tokens and whether
792 * we can get all the target thread's tokens, but this situation
793 * should not occur very often so its easier to simply not preempt.
794 * Also, plain spinlocks are impossible to figure out at this point so
795 * just don't preempt.
797 * Do not try to preempt if the target thread is holding any tokens.
798 * We could try to acquire the tokens but this case is so rare there
799 * is no need to support it.
801 if (gd->gd_spinlock_rd || gd->gd_spinlocks_wr) {
811 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
816 if (ntd->td_preempted) {
823 * note: an interrupt might have occured just as we were transitioning
824 * to or from the MP lock. In this case td_mpcount will be pre-disposed
825 * (non-zero) but not actually synchronized with the actual state of the
826 * lock. We can use it to imply an MP lock requirement for the
827 * preemption but we cannot use it to test whether we hold the MP lock
830 savecnt = td->td_mpcount;
831 mpheld = MP_LOCK_HELD();
832 ntd->td_mpcount += td->td_mpcount;
833 if (mpheld == 0 && ntd->td_mpcount && !cpu_try_mplock()) {
834 ntd->td_mpcount -= td->td_mpcount;
842 * Since we are able to preempt the current thread, there is no need to
843 * call need_lwkt_resched().
846 ntd->td_preempted = td;
847 td->td_flags |= TDF_PREEMPT_LOCK;
849 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
851 KKASSERT(savecnt == td->td_mpcount);
852 mpheld = MP_LOCK_HELD();
853 if (mpheld && td->td_mpcount == 0)
855 else if (mpheld == 0 && td->td_mpcount)
856 panic("lwkt_preempt(): MP lock was not held through");
858 ntd->td_preempted = NULL;
859 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
863 * Yield our thread while higher priority threads are pending. This is
864 * typically called when we leave a critical section but it can be safely
865 * called while we are in a critical section.
867 * This function will not generally yield to equal priority threads but it
868 * can occur as a side effect. Note that lwkt_switch() is called from
869 * inside the critical section to prevent its own crit_exit() from reentering
870 * lwkt_yield_quick().
872 * gd_reqflags indicates that *something* changed, e.g. an interrupt or softint
873 * came along but was blocked and made pending.
875 * (self contained on a per cpu basis)
878 lwkt_yield_quick(void)
880 globaldata_t gd = mycpu;
881 thread_t td = gd->gd_curthread;
884 * gd_reqflags is cleared in splz if the cpl is 0. If we were to clear
885 * it with a non-zero cpl then we might not wind up calling splz after
886 * a task switch when the critical section is exited even though the
887 * new task could accept the interrupt.
889 * XXX from crit_exit() only called after last crit section is released.
890 * If called directly will run splz() even if in a critical section.
892 * td_nest_count prevent deep nesting via splz() or doreti(). Note that
893 * except for this special case, we MUST call splz() here to handle any
894 * pending ints, particularly after we switch, or we might accidently
895 * halt the cpu with interrupts pending.
897 if (gd->gd_reqflags && td->td_nest_count < 2)
901 * YYY enabling will cause wakeup() to task-switch, which really
902 * confused the old 4.x code. This is a good way to simulate
903 * preemption and MP without actually doing preemption or MP, because a
904 * lot of code assumes that wakeup() does not block.
906 if (untimely_switch && td->td_nest_count == 0 &&
907 gd->gd_intr_nesting_level == 0
909 crit_enter_quick(td);
911 * YYY temporary hacks until we disassociate the userland scheduler
912 * from the LWKT scheduler.
914 if (td->td_flags & TDF_RUNQ) {
915 lwkt_switch(); /* will not reenter yield function */
917 lwkt_schedule_self(td); /* make sure we are scheduled */
918 lwkt_switch(); /* will not reenter yield function */
919 lwkt_deschedule_self(td); /* make sure we are descheduled */
921 crit_exit_noyield(td);
926 * This implements a normal yield which, unlike _quick, will yield to equal
927 * priority threads as well. Note that gd_reqflags tests will be handled by
928 * the crit_exit() call in lwkt_switch().
930 * (self contained on a per cpu basis)
935 lwkt_schedule_self(curthread);
940 * Generic schedule. Possibly schedule threads belonging to other cpus and
941 * deal with threads that might be blocked on a wait queue.
943 * We have a little helper inline function which does additional work after
944 * the thread has been enqueued, including dealing with preemption and
945 * setting need_lwkt_resched() (which prevents the kernel from returning
946 * to userland until it has processed higher priority threads).
948 * It is possible for this routine to be called after a failed _enqueue
949 * (due to the target thread migrating, sleeping, or otherwise blocked).
950 * We have to check that the thread is actually on the run queue!
954 _lwkt_schedule_post(globaldata_t gd, thread_t ntd, int cpri)
956 if (ntd->td_flags & TDF_RUNQ) {
957 if (ntd->td_preemptable) {
958 ntd->td_preemptable(ntd, cpri); /* YYY +token */
959 } else if ((ntd->td_flags & TDF_NORESCHED) == 0 &&
960 (ntd->td_pri & TDPRI_MASK) > (gd->gd_curthread->td_pri & TDPRI_MASK)
968 lwkt_schedule(thread_t td)
970 globaldata_t mygd = mycpu;
972 KASSERT(td != &td->td_gd->gd_idlethread, ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
974 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
975 if (td == mygd->gd_curthread) {
979 * If we own the thread, there is no race (since we are in a
980 * critical section). If we do not own the thread there might
981 * be a race but the target cpu will deal with it.
984 if (td->td_gd == mygd) {
986 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
988 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_schedule, td);
992 _lwkt_schedule_post(mygd, td, TDPRI_CRIT);
1001 * Thread migration using a 'Pull' method. The thread may or may not be
1002 * the current thread. It MUST be descheduled and in a stable state.
1003 * lwkt_giveaway() must be called on the cpu owning the thread.
1005 * At any point after lwkt_giveaway() is called, the target cpu may
1006 * 'pull' the thread by calling lwkt_acquire().
1008 * MPSAFE - must be called under very specific conditions.
1011 lwkt_giveaway(thread_t td)
1013 globaldata_t gd = mycpu;
1016 KKASSERT(td->td_gd == gd);
1017 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1018 td->td_flags |= TDF_MIGRATING;
1023 lwkt_acquire(thread_t td)
1028 KKASSERT(td->td_flags & TDF_MIGRATING);
1033 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1034 crit_enter_gd(mygd);
1035 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1038 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1039 td->td_flags &= ~TDF_MIGRATING;
1042 crit_enter_gd(mygd);
1043 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1044 td->td_flags &= ~TDF_MIGRATING;
1052 * Generic deschedule. Descheduling threads other then your own should be
1053 * done only in carefully controlled circumstances. Descheduling is
1056 * This function may block if the cpu has run out of messages.
1059 lwkt_deschedule(thread_t td)
1063 if (td == curthread) {
1066 if (td->td_gd == mycpu) {
1069 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1079 * Set the target thread's priority. This routine does not automatically
1080 * switch to a higher priority thread, LWKT threads are not designed for
1081 * continuous priority changes. Yield if you want to switch.
1083 * We have to retain the critical section count which uses the high bits
1084 * of the td_pri field. The specified priority may also indicate zero or
1085 * more critical sections by adding TDPRI_CRIT*N.
1087 * Note that we requeue the thread whether it winds up on a different runq
1088 * or not. uio_yield() depends on this and the routine is not normally
1089 * called with the same priority otherwise.
1092 lwkt_setpri(thread_t td, int pri)
1095 KKASSERT(td->td_gd == mycpu);
1097 if (td->td_flags & TDF_RUNQ) {
1099 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1102 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1108 lwkt_setpri_self(int pri)
1110 thread_t td = curthread;
1112 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1114 if (td->td_flags & TDF_RUNQ) {
1116 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1119 td->td_pri = (td->td_pri & ~TDPRI_MASK) + pri;
1125 * Determine if there is a runnable thread at a higher priority then
1126 * the current thread. lwkt_setpri() does not check this automatically.
1127 * Return 1 if there is, 0 if there isn't.
1129 * Example: if bit 31 of runqmask is set and the current thread is priority
1130 * 30, then we wind up checking the mask: 0x80000000 against 0x7fffffff.
1132 * If nq reaches 31 the shift operation will overflow to 0 and we will wind
1133 * up comparing against 0xffffffff, a comparison that will always be false.
1136 lwkt_checkpri_self(void)
1138 globaldata_t gd = mycpu;
1139 thread_t td = gd->gd_curthread;
1140 int nq = td->td_pri & TDPRI_MASK;
1142 while (gd->gd_runqmask > (__uint32_t)(2 << nq) - 1) {
1143 if (TAILQ_FIRST(&gd->gd_tdrunq[nq + 1]))
1151 * Migrate the current thread to the specified cpu.
1153 * This is accomplished by descheduling ourselves from the current cpu,
1154 * moving our thread to the tdallq of the target cpu, IPI messaging the
1155 * target cpu, and switching out. TDF_MIGRATING prevents scheduling
1156 * races while the thread is being migrated.
1159 static void lwkt_setcpu_remote(void *arg);
1163 lwkt_setcpu_self(globaldata_t rgd)
1166 thread_t td = curthread;
1168 if (td->td_gd != rgd) {
1169 crit_enter_quick(td);
1170 td->td_flags |= TDF_MIGRATING;
1171 lwkt_deschedule_self(td);
1172 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1173 lwkt_send_ipiq(rgd, (ipifunc1_t)lwkt_setcpu_remote, td);
1175 /* we are now on the target cpu */
1176 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1177 crit_exit_quick(td);
1183 lwkt_migratecpu(int cpuid)
1188 rgd = globaldata_find(cpuid);
1189 lwkt_setcpu_self(rgd);
1194 * Remote IPI for cpu migration (called while in a critical section so we
1195 * do not have to enter another one). The thread has already been moved to
1196 * our cpu's allq, but we must wait for the thread to be completely switched
1197 * out on the originating cpu before we schedule it on ours or the stack
1198 * state may be corrupt. We clear TDF_MIGRATING after flushing the GD
1199 * change to main memory.
1201 * XXX The use of TDF_MIGRATING might not be sufficient to avoid races
1202 * against wakeups. It is best if this interface is used only when there
1203 * are no pending events that might try to schedule the thread.
1207 lwkt_setcpu_remote(void *arg)
1210 globaldata_t gd = mycpu;
1212 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK))
1216 td->td_flags &= ~TDF_MIGRATING;
1217 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1223 lwkt_preempted_proc(void)
1225 thread_t td = curthread;
1226 while (td->td_preempted)
1227 td = td->td_preempted;
1232 * Create a kernel process/thread/whatever. It shares it's address space
1233 * with proc0 - ie: kernel only.
1235 * NOTE! By default new threads are created with the MP lock held. A
1236 * thread which does not require the MP lock should release it by calling
1237 * rel_mplock() at the start of the new thread.
1240 lwkt_create(void (*func)(void *), void *arg,
1241 struct thread **tdp, thread_t template, int tdflags, int cpu,
1242 const char *fmt, ...)
1247 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1251 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1254 * Set up arg0 for 'ps' etc
1256 __va_start(ap, fmt);
1257 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1261 * Schedule the thread to run
1263 if ((td->td_flags & TDF_STOPREQ) == 0)
1266 td->td_flags &= ~TDF_STOPREQ;
1271 * Destroy an LWKT thread. Warning! This function is not called when
1272 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1273 * uses a different reaping mechanism.
1278 thread_t td = curthread;
1281 if (td->td_flags & TDF_VERBOSE)
1282 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1284 crit_enter_quick(td);
1285 lwkt_deschedule_self(td);
1287 lwkt_remove_tdallq(td);
1288 if (td->td_flags & TDF_ALLOCATED_THREAD) {
1289 objcache_put(thread_cache, td);
1295 lwkt_remove_tdallq(thread_t td)
1297 KKASSERT(td->td_gd == mycpu);
1298 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1304 thread_t td = curthread;
1305 int lpri = td->td_pri;
1308 panic("td_pri is/would-go negative! %p %d", td, lpri);
1314 * Called from debugger/panic on cpus which have been stopped. We must still
1315 * process the IPIQ while stopped, even if we were stopped while in a critical
1318 * If we are dumping also try to process any pending interrupts. This may
1319 * or may not work depending on the state of the cpu at the point it was
1323 lwkt_smp_stopped(void)
1325 globaldata_t gd = mycpu;
1329 lwkt_process_ipiq();
1332 lwkt_process_ipiq();
1338 * get_mplock() calls this routine if it is unable to obtain the MP lock.
1339 * get_mplock() has already incremented td_mpcount. We must block and
1340 * not return until giant is held.
1342 * All we have to do is lwkt_switch() away. The LWKT scheduler will not
1343 * reschedule the thread until it can obtain the giant lock for it.
1346 lwkt_mp_lock_contested(void)