vkernel - do not set SA_NODEFER for SIGIO and SIGUSR1
[dragonfly.git] / sys / kern / lwkt_thread.c
... / ...
CommitLineData
1/*
2 * Copyright (c) 2003-2010 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
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
16 * distribution.
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.
20 *
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
32 * SUCH DAMAGE.
33 */
34
35/*
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.
40 */
41
42#include <sys/param.h>
43#include <sys/systm.h>
44#include <sys/kernel.h>
45#include <sys/proc.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>
52#include <sys/lock.h>
53#include <sys/caps.h>
54#include <sys/spinlock.h>
55#include <sys/ktr.h>
56
57#include <sys/thread2.h>
58#include <sys/spinlock2.h>
59#include <sys/mplock2.h>
60
61#include <sys/dsched.h>
62
63#include <vm/vm.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>
71
72#include <machine/stdarg.h>
73#include <machine/smp.h>
74
75#if !defined(KTR_CTXSW)
76#define KTR_CTXSW KTR_ALL
77#endif
78KTR_INFO_MASTER(ctxsw);
79KTR_INFO(KTR_CTXSW, ctxsw, sw, 0, "#cpu[%d].td = %p",
80 sizeof(int) + sizeof(struct thread *));
81KTR_INFO(KTR_CTXSW, ctxsw, pre, 1, "#cpu[%d].td = %p",
82 sizeof(int) + sizeof(struct thread *));
83KTR_INFO(KTR_CTXSW, ctxsw, newtd, 2, "#threads[%p].name = %s",
84 sizeof (struct thread *) + sizeof(char *));
85KTR_INFO(KTR_CTXSW, ctxsw, deadtd, 3, "#threads[%p].name = <dead>", sizeof (struct thread *));
86
87static MALLOC_DEFINE(M_THREAD, "thread", "lwkt threads");
88
89#ifdef INVARIANTS
90static int panic_on_cscount = 0;
91#endif
92static __int64_t switch_count = 0;
93static __int64_t preempt_hit = 0;
94static __int64_t preempt_miss = 0;
95static __int64_t preempt_weird = 0;
96static __int64_t token_contention_count __debugvar = 0;
97static int lwkt_use_spin_port;
98static struct objcache *thread_cache;
99
100#ifdef SMP
101static void lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame);
102static void lwkt_setcpu_remote(void *arg);
103#endif
104static void lwkt_fairq_accumulate(globaldata_t gd, thread_t td);
105
106extern void cpu_heavy_restore(void);
107extern void cpu_lwkt_restore(void);
108extern void cpu_kthread_restore(void);
109extern void cpu_idle_restore(void);
110
111/*
112 * We can make all thread ports use the spin backend instead of the thread
113 * backend. This should only be set to debug the spin backend.
114 */
115TUNABLE_INT("lwkt.use_spin_port", &lwkt_use_spin_port);
116
117#ifdef INVARIANTS
118SYSCTL_INT(_lwkt, OID_AUTO, panic_on_cscount, CTLFLAG_RW, &panic_on_cscount, 0,
119 "Panic if attempting to switch lwkt's while mastering cpusync");
120#endif
121SYSCTL_QUAD(_lwkt, OID_AUTO, switch_count, CTLFLAG_RW, &switch_count, 0,
122 "Number of switched threads");
123SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_hit, CTLFLAG_RW, &preempt_hit, 0,
124 "Successful preemption events");
125SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_miss, CTLFLAG_RW, &preempt_miss, 0,
126 "Failed preemption events");
127SYSCTL_QUAD(_lwkt, OID_AUTO, preempt_weird, CTLFLAG_RW, &preempt_weird, 0,
128 "Number of preempted threads.");
129#ifdef INVARIANTS
130SYSCTL_QUAD(_lwkt, OID_AUTO, token_contention_count, CTLFLAG_RW,
131 &token_contention_count, 0, "spinning due to token contention");
132#endif
133static int fairq_enable = 1;
134SYSCTL_INT(_lwkt, OID_AUTO, fairq_enable, CTLFLAG_RW,
135 &fairq_enable, 0, "Turn on fairq priority accumulators");
136static int lwkt_spin_loops = 10;
137SYSCTL_INT(_lwkt, OID_AUTO, spin_loops, CTLFLAG_RW,
138 &lwkt_spin_loops, 0, "");
139static int lwkt_spin_delay = 1;
140SYSCTL_INT(_lwkt, OID_AUTO, spin_delay, CTLFLAG_RW,
141 &lwkt_spin_delay, 0, "Scheduler spin delay in microseconds 0=auto");
142static int lwkt_spin_method = 1;
143SYSCTL_INT(_lwkt, OID_AUTO, spin_method, CTLFLAG_RW,
144 &lwkt_spin_method, 0, "LWKT scheduler behavior when contended");
145static int lwkt_spin_fatal = 0; /* disabled */
146SYSCTL_INT(_lwkt, OID_AUTO, spin_fatal, CTLFLAG_RW,
147 &lwkt_spin_fatal, 0, "LWKT scheduler spin loops till fatal panic");
148static int preempt_enable = 1;
149SYSCTL_INT(_lwkt, OID_AUTO, preempt_enable, CTLFLAG_RW,
150 &preempt_enable, 0, "Enable preemption");
151static int lwkt_cache_threads = 32;
152SYSCTL_INT(_lwkt, OID_AUTO, cache_threads, CTLFLAG_RD,
153 &lwkt_cache_threads, 0, "thread+kstack cache");
154
155static __cachealign int lwkt_cseq_rindex;
156static __cachealign int lwkt_cseq_windex;
157
158/*
159 * These helper procedures handle the runq, they can only be called from
160 * within a critical section.
161 *
162 * WARNING! Prior to SMP being brought up it is possible to enqueue and
163 * dequeue threads belonging to other cpus, so be sure to use td->td_gd
164 * instead of 'mycpu' when referencing the globaldata structure. Once
165 * SMP live enqueuing and dequeueing only occurs on the current cpu.
166 */
167static __inline
168void
169_lwkt_dequeue(thread_t td)
170{
171 if (td->td_flags & TDF_RUNQ) {
172 struct globaldata *gd = td->td_gd;
173
174 td->td_flags &= ~TDF_RUNQ;
175 TAILQ_REMOVE(&gd->gd_tdrunq, td, td_threadq);
176 gd->gd_fairq_total_pri -= td->td_pri;
177 if (TAILQ_FIRST(&gd->gd_tdrunq) == NULL)
178 atomic_clear_int(&gd->gd_reqflags, RQF_RUNNING);
179 }
180}
181
182/*
183 * Priority enqueue.
184 *
185 * NOTE: There are a limited number of lwkt threads runnable since user
186 * processes only schedule one at a time per cpu.
187 */
188static __inline
189void
190_lwkt_enqueue(thread_t td)
191{
192 thread_t xtd;
193
194 if ((td->td_flags & (TDF_RUNQ|TDF_MIGRATING|TDF_BLOCKQ)) == 0) {
195 struct globaldata *gd = td->td_gd;
196
197 td->td_flags |= TDF_RUNQ;
198 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
199 if (xtd == NULL) {
200 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
201 atomic_set_int(&gd->gd_reqflags, RQF_RUNNING);
202 } else {
203 while (xtd && xtd->td_pri > td->td_pri)
204 xtd = TAILQ_NEXT(xtd, td_threadq);
205 if (xtd)
206 TAILQ_INSERT_BEFORE(xtd, td, td_threadq);
207 else
208 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, td, td_threadq);
209 }
210 gd->gd_fairq_total_pri += td->td_pri;
211 }
212}
213
214static __boolean_t
215_lwkt_thread_ctor(void *obj, void *privdata, int ocflags)
216{
217 struct thread *td = (struct thread *)obj;
218
219 td->td_kstack = NULL;
220 td->td_kstack_size = 0;
221 td->td_flags = TDF_ALLOCATED_THREAD;
222 return (1);
223}
224
225static void
226_lwkt_thread_dtor(void *obj, void *privdata)
227{
228 struct thread *td = (struct thread *)obj;
229
230 KASSERT(td->td_flags & TDF_ALLOCATED_THREAD,
231 ("_lwkt_thread_dtor: not allocated from objcache"));
232 KASSERT((td->td_flags & TDF_ALLOCATED_STACK) && td->td_kstack &&
233 td->td_kstack_size > 0,
234 ("_lwkt_thread_dtor: corrupted stack"));
235 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
236}
237
238/*
239 * Initialize the lwkt s/system.
240 *
241 * Nominally cache up to 32 thread + kstack structures.
242 */
243void
244lwkt_init(void)
245{
246 TUNABLE_INT("lwkt.cache_threads", &lwkt_cache_threads);
247 thread_cache = objcache_create_mbacked(
248 M_THREAD, sizeof(struct thread),
249 NULL, lwkt_cache_threads,
250 _lwkt_thread_ctor, _lwkt_thread_dtor, NULL);
251}
252
253/*
254 * Schedule a thread to run. As the current thread we can always safely
255 * schedule ourselves, and a shortcut procedure is provided for that
256 * function.
257 *
258 * (non-blocking, self contained on a per cpu basis)
259 */
260void
261lwkt_schedule_self(thread_t td)
262{
263 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
264 crit_enter_quick(td);
265 KASSERT(td != &td->td_gd->gd_idlethread,
266 ("lwkt_schedule_self(): scheduling gd_idlethread is illegal!"));
267 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
268 _lwkt_enqueue(td);
269 crit_exit_quick(td);
270}
271
272/*
273 * Deschedule a thread.
274 *
275 * (non-blocking, self contained on a per cpu basis)
276 */
277void
278lwkt_deschedule_self(thread_t td)
279{
280 crit_enter_quick(td);
281 _lwkt_dequeue(td);
282 crit_exit_quick(td);
283}
284
285/*
286 * LWKTs operate on a per-cpu basis
287 *
288 * WARNING! Called from early boot, 'mycpu' may not work yet.
289 */
290void
291lwkt_gdinit(struct globaldata *gd)
292{
293 TAILQ_INIT(&gd->gd_tdrunq);
294 TAILQ_INIT(&gd->gd_tdallq);
295}
296
297/*
298 * Create a new thread. The thread must be associated with a process context
299 * or LWKT start address before it can be scheduled. If the target cpu is
300 * -1 the thread will be created on the current cpu.
301 *
302 * If you intend to create a thread without a process context this function
303 * does everything except load the startup and switcher function.
304 */
305thread_t
306lwkt_alloc_thread(struct thread *td, int stksize, int cpu, int flags)
307{
308 globaldata_t gd = mycpu;
309 void *stack;
310
311 /*
312 * If static thread storage is not supplied allocate a thread. Reuse
313 * a cached free thread if possible. gd_freetd is used to keep an exiting
314 * thread intact through the exit.
315 */
316 if (td == NULL) {
317 crit_enter_gd(gd);
318 if ((td = gd->gd_freetd) != NULL) {
319 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
320 TDF_RUNQ)) == 0);
321 gd->gd_freetd = NULL;
322 } else {
323 td = objcache_get(thread_cache, M_WAITOK);
324 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|
325 TDF_RUNQ)) == 0);
326 }
327 crit_exit_gd(gd);
328 KASSERT((td->td_flags &
329 (TDF_ALLOCATED_THREAD|TDF_RUNNING)) == TDF_ALLOCATED_THREAD,
330 ("lwkt_alloc_thread: corrupted td flags 0x%X", td->td_flags));
331 flags |= td->td_flags & (TDF_ALLOCATED_THREAD|TDF_ALLOCATED_STACK);
332 }
333
334 /*
335 * Try to reuse cached stack.
336 */
337 if ((stack = td->td_kstack) != NULL && td->td_kstack_size != stksize) {
338 if (flags & TDF_ALLOCATED_STACK) {
339 kmem_free(&kernel_map, (vm_offset_t)stack, td->td_kstack_size);
340 stack = NULL;
341 }
342 }
343 if (stack == NULL) {
344 stack = (void *)kmem_alloc_stack(&kernel_map, stksize);
345 flags |= TDF_ALLOCATED_STACK;
346 }
347 if (cpu < 0)
348 lwkt_init_thread(td, stack, stksize, flags, gd);
349 else
350 lwkt_init_thread(td, stack, stksize, flags, globaldata_find(cpu));
351 return(td);
352}
353
354/*
355 * Initialize a preexisting thread structure. This function is used by
356 * lwkt_alloc_thread() and also used to initialize the per-cpu idlethread.
357 *
358 * All threads start out in a critical section at a priority of
359 * TDPRI_KERN_DAEMON. Higher level code will modify the priority as
360 * appropriate. This function may send an IPI message when the
361 * requested cpu is not the current cpu and consequently gd_tdallq may
362 * not be initialized synchronously from the point of view of the originating
363 * cpu.
364 *
365 * NOTE! we have to be careful in regards to creating threads for other cpus
366 * if SMP has not yet been activated.
367 */
368#ifdef SMP
369
370static void
371lwkt_init_thread_remote(void *arg)
372{
373 thread_t td = arg;
374
375 /*
376 * Protected by critical section held by IPI dispatch
377 */
378 TAILQ_INSERT_TAIL(&td->td_gd->gd_tdallq, td, td_allq);
379}
380
381#endif
382
383/*
384 * lwkt core thread structural initialization.
385 *
386 * NOTE: All threads are initialized as mpsafe threads.
387 */
388void
389lwkt_init_thread(thread_t td, void *stack, int stksize, int flags,
390 struct globaldata *gd)
391{
392 globaldata_t mygd = mycpu;
393
394 bzero(td, sizeof(struct thread));
395 td->td_kstack = stack;
396 td->td_kstack_size = stksize;
397 td->td_flags = flags;
398 td->td_gd = gd;
399 td->td_pri = TDPRI_KERN_DAEMON;
400 td->td_critcount = 1;
401 td->td_toks_stop = &td->td_toks_base;
402 if (lwkt_use_spin_port)
403 lwkt_initport_spin(&td->td_msgport);
404 else
405 lwkt_initport_thread(&td->td_msgport, td);
406 pmap_init_thread(td);
407#ifdef SMP
408 /*
409 * Normally initializing a thread for a remote cpu requires sending an
410 * IPI. However, the idlethread is setup before the other cpus are
411 * activated so we have to treat it as a special case. XXX manipulation
412 * of gd_tdallq requires the BGL.
413 */
414 if (gd == mygd || td == &gd->gd_idlethread) {
415 crit_enter_gd(mygd);
416 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
417 crit_exit_gd(mygd);
418 } else {
419 lwkt_send_ipiq(gd, lwkt_init_thread_remote, td);
420 }
421#else
422 crit_enter_gd(mygd);
423 TAILQ_INSERT_TAIL(&gd->gd_tdallq, td, td_allq);
424 crit_exit_gd(mygd);
425#endif
426
427 dsched_new_thread(td);
428}
429
430void
431lwkt_set_comm(thread_t td, const char *ctl, ...)
432{
433 __va_list va;
434
435 __va_start(va, ctl);
436 kvsnprintf(td->td_comm, sizeof(td->td_comm), ctl, va);
437 __va_end(va);
438 KTR_LOG(ctxsw_newtd, td, &td->td_comm[0]);
439}
440
441void
442lwkt_hold(thread_t td)
443{
444 atomic_add_int(&td->td_refs, 1);
445}
446
447void
448lwkt_rele(thread_t td)
449{
450 KKASSERT(td->td_refs > 0);
451 atomic_add_int(&td->td_refs, -1);
452}
453
454void
455lwkt_wait_free(thread_t td)
456{
457 while (td->td_refs)
458 tsleep(td, 0, "tdreap", hz);
459}
460
461void
462lwkt_free_thread(thread_t td)
463{
464 KKASSERT(td->td_refs == 0);
465 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK|TDF_RUNQ)) == 0);
466 if (td->td_flags & TDF_ALLOCATED_THREAD) {
467 objcache_put(thread_cache, td);
468 } else if (td->td_flags & TDF_ALLOCATED_STACK) {
469 /* client-allocated struct with internally allocated stack */
470 KASSERT(td->td_kstack && td->td_kstack_size > 0,
471 ("lwkt_free_thread: corrupted stack"));
472 kmem_free(&kernel_map, (vm_offset_t)td->td_kstack, td->td_kstack_size);
473 td->td_kstack = NULL;
474 td->td_kstack_size = 0;
475 }
476 KTR_LOG(ctxsw_deadtd, td);
477}
478
479
480/*
481 * Switch to the next runnable lwkt. If no LWKTs are runnable then
482 * switch to the idlethread. Switching must occur within a critical
483 * section to avoid races with the scheduling queue.
484 *
485 * We always have full control over our cpu's run queue. Other cpus
486 * that wish to manipulate our queue must use the cpu_*msg() calls to
487 * talk to our cpu, so a critical section is all that is needed and
488 * the result is very, very fast thread switching.
489 *
490 * The LWKT scheduler uses a fixed priority model and round-robins at
491 * each priority level. User process scheduling is a totally
492 * different beast and LWKT priorities should not be confused with
493 * user process priorities.
494 *
495 * PREEMPTION NOTE: Preemption occurs via lwkt_preempt(). lwkt_switch()
496 * is not called by the current thread in the preemption case, only when
497 * the preempting thread blocks (in order to return to the original thread).
498 *
499 * SPECIAL NOTE ON SWITCH ATOMICY: Certain operations such as thread
500 * migration and tsleep deschedule the current lwkt thread and call
501 * lwkt_switch(). In particular, the target cpu of the migration fully
502 * expects the thread to become non-runnable and can deadlock against
503 * cpusync operations if we run any IPIs prior to switching the thread out.
504 *
505 * WE MUST BE VERY CAREFUL NOT TO RUN SPLZ DIRECTLY OR INDIRECTLY IF
506 * THE CURRENT THREAD HAS BEEN DESCHEDULED!
507 */
508void
509lwkt_switch(void)
510{
511 globaldata_t gd = mycpu;
512 thread_t td = gd->gd_curthread;
513 thread_t ntd;
514 thread_t xtd;
515 int spinning = lwkt_spin_loops; /* loops before HLTing */
516 int reqflags;
517 int cseq;
518 int oseq;
519 int fatal_count;
520
521 /*
522 * Switching from within a 'fast' (non thread switched) interrupt or IPI
523 * is illegal. However, we may have to do it anyway if we hit a fatal
524 * kernel trap or we have paniced.
525 *
526 * If this case occurs save and restore the interrupt nesting level.
527 */
528 if (gd->gd_intr_nesting_level) {
529 int savegdnest;
530 int savegdtrap;
531
532 if (gd->gd_trap_nesting_level == 0 && panic_cpu_gd != mycpu) {
533 panic("lwkt_switch: Attempt to switch from a "
534 "a fast interrupt, ipi, or hard code section, "
535 "td %p\n",
536 td);
537 } else {
538 savegdnest = gd->gd_intr_nesting_level;
539 savegdtrap = gd->gd_trap_nesting_level;
540 gd->gd_intr_nesting_level = 0;
541 gd->gd_trap_nesting_level = 0;
542 if ((td->td_flags & TDF_PANICWARN) == 0) {
543 td->td_flags |= TDF_PANICWARN;
544 kprintf("Warning: thread switch from interrupt, IPI, "
545 "or hard code section.\n"
546 "thread %p (%s)\n", td, td->td_comm);
547 print_backtrace(-1);
548 }
549 lwkt_switch();
550 gd->gd_intr_nesting_level = savegdnest;
551 gd->gd_trap_nesting_level = savegdtrap;
552 return;
553 }
554 }
555
556 /*
557 * Passive release (used to transition from user to kernel mode
558 * when we block or switch rather then when we enter the kernel).
559 * This function is NOT called if we are switching into a preemption
560 * or returning from a preemption. Typically this causes us to lose
561 * our current process designation (if we have one) and become a true
562 * LWKT thread, and may also hand the current process designation to
563 * another process and schedule thread.
564 */
565 if (td->td_release)
566 td->td_release(td);
567
568 crit_enter_gd(gd);
569 if (TD_TOKS_HELD(td))
570 lwkt_relalltokens(td);
571
572 /*
573 * We had better not be holding any spin locks, but don't get into an
574 * endless panic loop.
575 */
576 KASSERT(gd->gd_spinlocks_wr == 0 || panicstr != NULL,
577 ("lwkt_switch: still holding %d exclusive spinlocks!",
578 gd->gd_spinlocks_wr));
579
580
581#ifdef SMP
582#ifdef INVARIANTS
583 if (td->td_cscount) {
584 kprintf("Diagnostic: attempt to switch while mastering cpusync: %p\n",
585 td);
586 if (panic_on_cscount)
587 panic("switching while mastering cpusync");
588 }
589#endif
590#endif
591
592 /*
593 * If we had preempted another thread on this cpu, resume the preempted
594 * thread. This occurs transparently, whether the preempted thread
595 * was scheduled or not (it may have been preempted after descheduling
596 * itself).
597 *
598 * We have to setup the MP lock for the original thread after backing
599 * out the adjustment that was made to curthread when the original
600 * was preempted.
601 */
602 if ((ntd = td->td_preempted) != NULL) {
603 KKASSERT(ntd->td_flags & TDF_PREEMPT_LOCK);
604 ntd->td_flags |= TDF_PREEMPT_DONE;
605
606 /*
607 * The interrupt may have woken a thread up, we need to properly
608 * set the reschedule flag if the originally interrupted thread is
609 * at a lower priority.
610 */
611 if (TAILQ_FIRST(&gd->gd_tdrunq) &&
612 TAILQ_FIRST(&gd->gd_tdrunq)->td_pri > ntd->td_pri) {
613 need_lwkt_resched();
614 }
615 /* YYY release mp lock on switchback if original doesn't need it */
616 goto havethread_preempted;
617 }
618
619 /*
620 * Implement round-robin fairq with priority insertion. The priority
621 * insertion is handled by _lwkt_enqueue()
622 *
623 * If we cannot obtain ownership of the tokens we cannot immediately
624 * schedule the target thread.
625 *
626 * Reminder: Again, we cannot afford to run any IPIs in this path if
627 * the current thread has been descheduled.
628 */
629 for (;;) {
630 /*
631 * Clear RQF_AST_LWKT_RESCHED (we handle the reschedule request)
632 * and set RQF_WAKEUP (prevent unnecessary IPIs from being
633 * received).
634 */
635 for (;;) {
636 reqflags = gd->gd_reqflags;
637 if (atomic_cmpset_int(&gd->gd_reqflags, reqflags,
638 (reqflags & ~RQF_AST_LWKT_RESCHED) |
639 RQF_WAKEUP)) {
640 break;
641 }
642 }
643
644 /*
645 * Hotpath - pull the head of the run queue and attempt to schedule
646 * it. Fairq exhaustion moves the task to the end of the list. If
647 * no threads are runnable we switch to the idle thread.
648 */
649 for (;;) {
650 ntd = TAILQ_FIRST(&gd->gd_tdrunq);
651
652 if (ntd == NULL) {
653 /*
654 * Runq is empty, switch to idle and clear RQF_WAKEUP
655 * to allow it to halt.
656 */
657 ntd = &gd->gd_idlethread;
658#ifdef SMP
659 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
660 ASSERT_NO_TOKENS_HELD(ntd);
661#endif
662 cpu_time.cp_msg[0] = 0;
663 cpu_time.cp_stallpc = 0;
664 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
665 goto haveidle;
666 }
667
668 if (ntd->td_fairq_accum >= 0)
669 break;
670
671 /*splz_check(); cannot do this here, see above */
672 lwkt_fairq_accumulate(gd, ntd);
673 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
674 TAILQ_INSERT_TAIL(&gd->gd_tdrunq, ntd, td_threadq);
675 }
676
677 /*
678 * Hotpath - schedule ntd. Leaves RQF_WAKEUP set to prevent
679 * unwanted decontention IPIs.
680 *
681 * NOTE: For UP there is no mplock and lwkt_getalltokens()
682 * always succeeds.
683 */
684 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
685 goto havethread;
686
687 /*
688 * Coldpath (SMP only since tokens always succeed on UP)
689 *
690 * We had some contention on the thread we wanted to schedule.
691 * What we do now is try to find a thread that we can schedule
692 * in its stead until decontention reschedules on our cpu.
693 *
694 * The coldpath scan does NOT rearrange threads in the run list
695 * and it also ignores the accumulator.
696 *
697 * We do not immediately schedule a user priority thread, instead
698 * we record it in xtd and continue looking for kernel threads.
699 * A cpu can only have one user priority thread (normally) so just
700 * record the first one.
701 *
702 * NOTE: This scan will also include threads whos fairq's were
703 * accumulated in the first loop.
704 */
705 ++token_contention_count;
706 xtd = NULL;
707 while ((ntd = TAILQ_NEXT(ntd, td_threadq)) != NULL) {
708 /*
709 * Try to switch to this thread. If the thread is running at
710 * user priority we clear WAKEUP to allow decontention IPIs
711 * (since this thread is simply running until the one we wanted
712 * decontends), and we make sure that LWKT_RESCHED is not set.
713 *
714 * Otherwise for kernel threads we leave WAKEUP set to avoid
715 * unnecessary decontention IPIs.
716 */
717 if (ntd->td_pri < TDPRI_KERN_LPSCHED) {
718 if (xtd == NULL)
719 xtd = ntd;
720 continue;
721 }
722
723 /*
724 * Do not let the fairq get too negative. Even though we are
725 * ignoring it atm once the scheduler decontends a very negative
726 * thread will get moved to the end of the queue.
727 */
728 if (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd)) {
729 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
730 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
731 goto havethread;
732 }
733
734 /*
735 * Well fubar, this thread is contended as well, loop
736 */
737 /* */
738 }
739
740 /*
741 * We exhausted the run list but we may have recorded a user
742 * thread to try. We have three choices based on
743 * lwkt.decontention_method.
744 *
745 * (0) Atomically clear RQF_WAKEUP in order to receive decontention
746 * IPIs (to interrupt the user process) and test
747 * RQF_AST_LWKT_RESCHED at the same time.
748 *
749 * This results in significant decontention IPI traffic but may
750 * be more responsive.
751 *
752 * (1) Leave RQF_WAKEUP set so we do not receive a decontention IPI.
753 * An automatic LWKT reschedule will occur on the next hardclock
754 * (typically 100hz).
755 *
756 * This results in no decontention IPI traffic but may be less
757 * responsive. This is the default.
758 *
759 * (2) Refuse to schedule the user process at this time.
760 *
761 * This is highly experimental and should not be used under
762 * normal circumstances. This can cause a user process to
763 * get starved out in situations where kernel threads are
764 * fighting each other for tokens.
765 */
766 if (xtd) {
767 ntd = xtd;
768
769 switch(lwkt_spin_method) {
770 case 0:
771 for (;;) {
772 reqflags = gd->gd_reqflags;
773 if (atomic_cmpset_int(&gd->gd_reqflags,
774 reqflags,
775 reqflags & ~RQF_WAKEUP)) {
776 break;
777 }
778 }
779 break;
780 case 1:
781 reqflags = gd->gd_reqflags;
782 break;
783 default:
784 goto skip;
785 break;
786 }
787 if ((reqflags & RQF_AST_LWKT_RESCHED) == 0 &&
788 (TD_TOKS_NOT_HELD(ntd) || lwkt_getalltokens(ntd))
789 ) {
790 if (ntd->td_fairq_accum < -TDFAIRQ_MAX(gd))
791 ntd->td_fairq_accum = -TDFAIRQ_MAX(gd);
792 goto havethread;
793 }
794
795skip:
796 /*
797 * Make sure RQF_WAKEUP is set if we failed to schedule the
798 * user thread to prevent the idle thread from halting.
799 */
800 atomic_set_int(&gd->gd_reqflags, RQF_WAKEUP);
801 }
802
803 /*
804 * We exhausted the run list, meaning that all runnable threads
805 * are contended.
806 */
807 cpu_pause();
808 ntd = &gd->gd_idlethread;
809#ifdef SMP
810 if (gd->gd_trap_nesting_level == 0 && panicstr == NULL)
811 ASSERT_NO_TOKENS_HELD(ntd);
812 /* contention case, do not clear contention mask */
813#endif
814
815 /*
816 * Ok, we might want to spin a few times as some tokens are held for
817 * very short periods of time and IPI overhead is 1uS or worse
818 * (meaning it is usually better to spin). Regardless we have to
819 * call splz_check() to be sure to service any interrupts blocked
820 * by our critical section, otherwise we could livelock e.g. IPIs.
821 *
822 * The IPI mechanic is really a last resort. In nearly all other
823 * cases RQF_WAKEUP is left set to prevent decontention IPIs.
824 *
825 * When we decide not to spin we clear RQF_WAKEUP and switch to
826 * the idle thread. Clearing RQF_WEAKEUP allows the idle thread
827 * to halt and decontended tokens will issue an IPI to us. The
828 * idle thread will check for pending reschedules already set
829 * (RQF_AST_LWKT_RESCHED) before actually halting so we don't have
830 * to here.
831 *
832 * Also, if TDF_RUNQ is not set the current thread is trying to
833 * deschedule, possibly in an atomic fashion. We cannot afford to
834 * stay here.
835 */
836 if (spinning <= 0 || (td->td_flags & TDF_RUNQ) == 0) {
837 atomic_clear_int(&gd->gd_reqflags, RQF_WAKEUP);
838 goto haveidle;
839 }
840 --spinning;
841
842 /*
843 * When spinning a delay is required both to avoid livelocks from
844 * token order reversals (a thread may be trying to acquire multiple
845 * tokens), and also to reduce cpu cache management traffic.
846 *
847 * In order to scale to a large number of CPUs we use a time slot
848 * resequencer to force contending cpus into non-contending
849 * time-slots. The scheduler may still contend with the lock holder
850 * but will not (generally) contend with all the other cpus trying
851 * trying to get the same token.
852 *
853 * The resequencer uses a FIFO counter mechanic. The owner of the
854 * rindex at the head of the FIFO is allowed to pull itself off
855 * the FIFO and fetchadd is used to enter into the FIFO. This bit
856 * of code is VERY cache friendly and forces all spinning schedulers
857 * into their own time slots.
858 *
859 * This code has been tested to 48-cpus and caps the cache
860 * contention load at ~1uS intervals regardless of the number of
861 * cpus. Scaling beyond 64 cpus might require additional smarts
862 * (such as separate FIFOs for specific token cases).
863 *
864 * WARNING! We can't call splz_check() or anything else here as
865 * it could cause a deadlock.
866 */
867#if defined(INVARIANTS) && defined(__amd64__)
868 if ((read_rflags() & PSL_I) == 0) {
869 cpu_enable_intr();
870 panic("lwkt_switch() called with interrupts disabled");
871 }
872#endif
873 cseq = atomic_fetchadd_int(&lwkt_cseq_windex, 1);
874 fatal_count = lwkt_spin_fatal;
875 while ((oseq = lwkt_cseq_rindex) != cseq) {
876 cpu_ccfence();
877#if !defined(_KERNEL_VIRTUAL)
878 if (cpu_mi_feature & CPU_MI_MONITOR) {
879 cpu_mmw_pause_int(&lwkt_cseq_rindex, oseq);
880 } else
881#endif
882 {
883 DELAY(1);
884 cpu_lfence();
885 }
886 if (fatal_count && --fatal_count == 0)
887 panic("lwkt_switch: fatal spin wait");
888 }
889 cseq = lwkt_spin_delay; /* don't trust the system operator */
890 cpu_ccfence();
891 if (cseq < 1)
892 cseq = 1;
893 if (cseq > 1000)
894 cseq = 1000;
895 DELAY(cseq);
896 atomic_add_int(&lwkt_cseq_rindex, 1);
897 splz_check(); /* ok, we already checked that td is still scheduled */
898 /* highest level for(;;) loop */
899 }
900
901havethread:
902 /*
903 * We must always decrement td_fairq_accum on non-idle threads just
904 * in case a thread never gets a tick due to being in a continuous
905 * critical section. The page-zeroing code does this, for example.
906 *
907 * If the thread we came up with is a higher or equal priority verses
908 * the thread at the head of the queue we move our thread to the
909 * front. This way we can always check the front of the queue.
910 *
911 * Clear gd_idle_repeat when doing a normal switch to a non-idle
912 * thread.
913 */
914 ++gd->gd_cnt.v_swtch;
915 --ntd->td_fairq_accum;
916 ntd->td_wmesg = NULL;
917 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
918 if (ntd != xtd && ntd->td_pri >= xtd->td_pri) {
919 TAILQ_REMOVE(&gd->gd_tdrunq, ntd, td_threadq);
920 TAILQ_INSERT_HEAD(&gd->gd_tdrunq, ntd, td_threadq);
921 }
922 gd->gd_idle_repeat = 0;
923
924havethread_preempted:
925 /*
926 * If the new target does not need the MP lock and we are holding it,
927 * release the MP lock. If the new target requires the MP lock we have
928 * already acquired it for the target.
929 */
930 ;
931haveidle:
932 KASSERT(ntd->td_critcount,
933 ("priority problem in lwkt_switch %d %d",
934 td->td_critcount, ntd->td_critcount));
935
936 if (td != ntd) {
937 /*
938 * Execute the actual thread switch operation. This function
939 * returns to the current thread and returns the previous thread
940 * (which may be different from the thread we switched to).
941 *
942 * We are responsible for marking ntd as TDF_RUNNING.
943 */
944 ++switch_count;
945 KTR_LOG(ctxsw_sw, gd->gd_cpuid, ntd);
946 ntd->td_flags |= TDF_RUNNING;
947 lwkt_switch_return(td->td_switch(ntd));
948 /* ntd invalid, td_switch() can return a different thread_t */
949 }
950 /* NOTE: current cpu may have changed after switch */
951 crit_exit_quick(td);
952}
953
954/*
955 * Called by assembly in the td_switch (thread restore path) for thread
956 * bootstrap cases which do not 'return' to lwkt_switch().
957 */
958void
959lwkt_switch_return(thread_t otd)
960{
961#ifdef SMP
962 globaldata_t rgd;
963
964 /*
965 * Check if otd was migrating. Now that we are on ntd we can finish
966 * up the migration. This is a bit messy but it is the only place
967 * where td is known to be fully descheduled.
968 *
969 * We can only activate the migration if otd was migrating but not
970 * held on the cpu due to a preemption chain. We still have to
971 * clear TDF_RUNNING on the old thread either way.
972 *
973 * We are responsible for clearing the previously running thread's
974 * TDF_RUNNING.
975 */
976 if ((rgd = otd->td_migrate_gd) != NULL &&
977 (otd->td_flags & TDF_PREEMPT_LOCK) == 0) {
978 KKASSERT((otd->td_flags & (TDF_MIGRATING | TDF_RUNNING)) ==
979 (TDF_MIGRATING | TDF_RUNNING));
980 otd->td_migrate_gd = NULL;
981 otd->td_flags &= ~TDF_RUNNING;
982 lwkt_send_ipiq(rgd, lwkt_setcpu_remote, otd);
983 } else {
984 otd->td_flags &= ~TDF_RUNNING;
985 }
986#else
987 otd->td_flags &= ~TDF_RUNNING;
988#endif
989}
990
991/*
992 * Request that the target thread preempt the current thread. Preemption
993 * only works under a specific set of conditions:
994 *
995 * - We are not preempting ourselves
996 * - The target thread is owned by the current cpu
997 * - We are not currently being preempted
998 * - The target is not currently being preempted
999 * - We are not holding any spin locks
1000 * - The target thread is not holding any tokens
1001 * - We are able to satisfy the target's MP lock requirements (if any).
1002 *
1003 * THE CALLER OF LWKT_PREEMPT() MUST BE IN A CRITICAL SECTION. Typically
1004 * this is called via lwkt_schedule() through the td_preemptable callback.
1005 * critcount is the managed critical priority that we should ignore in order
1006 * to determine whether preemption is possible (aka usually just the crit
1007 * priority of lwkt_schedule() itself).
1008 *
1009 * XXX at the moment we run the target thread in a critical section during
1010 * the preemption in order to prevent the target from taking interrupts
1011 * that *WE* can't. Preemption is strictly limited to interrupt threads
1012 * and interrupt-like threads, outside of a critical section, and the
1013 * preempted source thread will be resumed the instant the target blocks
1014 * whether or not the source is scheduled (i.e. preemption is supposed to
1015 * be as transparent as possible).
1016 */
1017void
1018lwkt_preempt(thread_t ntd, int critcount)
1019{
1020 struct globaldata *gd = mycpu;
1021 thread_t xtd;
1022 thread_t td;
1023 int save_gd_intr_nesting_level;
1024
1025 /*
1026 * The caller has put us in a critical section. We can only preempt
1027 * if the caller of the caller was not in a critical section (basically
1028 * a local interrupt), as determined by the 'critcount' parameter. We
1029 * also can't preempt if the caller is holding any spinlocks (even if
1030 * he isn't in a critical section). This also handles the tokens test.
1031 *
1032 * YYY The target thread must be in a critical section (else it must
1033 * inherit our critical section? I dunno yet).
1034 *
1035 * Set need_lwkt_resched() unconditionally for now YYY.
1036 */
1037 KASSERT(ntd->td_critcount, ("BADCRIT0 %d", ntd->td_pri));
1038
1039 if (preempt_enable == 0) {
1040 ++preempt_miss;
1041 return;
1042 }
1043
1044 td = gd->gd_curthread;
1045 if (ntd->td_pri <= td->td_pri) {
1046 ++preempt_miss;
1047 return;
1048 }
1049 if (td->td_critcount > critcount) {
1050 ++preempt_miss;
1051 need_lwkt_resched();
1052 return;
1053 }
1054#ifdef SMP
1055 if (ntd->td_gd != gd) {
1056 ++preempt_miss;
1057 need_lwkt_resched();
1058 return;
1059 }
1060#endif
1061 /*
1062 * We don't have to check spinlocks here as they will also bump
1063 * td_critcount.
1064 *
1065 * Do not try to preempt if the target thread is holding any tokens.
1066 * We could try to acquire the tokens but this case is so rare there
1067 * is no need to support it.
1068 */
1069 KKASSERT(gd->gd_spinlocks_wr == 0);
1070
1071 if (TD_TOKS_HELD(ntd)) {
1072 ++preempt_miss;
1073 need_lwkt_resched();
1074 return;
1075 }
1076 if (td == ntd || ((td->td_flags | ntd->td_flags) & TDF_PREEMPT_LOCK)) {
1077 ++preempt_weird;
1078 need_lwkt_resched();
1079 return;
1080 }
1081 if (ntd->td_preempted) {
1082 ++preempt_hit;
1083 need_lwkt_resched();
1084 return;
1085 }
1086
1087 /*
1088 * Since we are able to preempt the current thread, there is no need to
1089 * call need_lwkt_resched().
1090 *
1091 * We must temporarily clear gd_intr_nesting_level around the switch
1092 * since switchouts from the target thread are allowed (they will just
1093 * return to our thread), and since the target thread has its own stack.
1094 *
1095 * A preemption must switch back to the original thread, assert the
1096 * case.
1097 */
1098 ++preempt_hit;
1099 ntd->td_preempted = td;
1100 td->td_flags |= TDF_PREEMPT_LOCK;
1101 KTR_LOG(ctxsw_pre, gd->gd_cpuid, ntd);
1102 save_gd_intr_nesting_level = gd->gd_intr_nesting_level;
1103 gd->gd_intr_nesting_level = 0;
1104 ntd->td_flags |= TDF_RUNNING;
1105 xtd = td->td_switch(ntd);
1106 KKASSERT(xtd == ntd);
1107 lwkt_switch_return(xtd);
1108 gd->gd_intr_nesting_level = save_gd_intr_nesting_level;
1109
1110 KKASSERT(ntd->td_preempted && (td->td_flags & TDF_PREEMPT_DONE));
1111 ntd->td_preempted = NULL;
1112 td->td_flags &= ~(TDF_PREEMPT_LOCK|TDF_PREEMPT_DONE);
1113}
1114
1115/*
1116 * Conditionally call splz() if gd_reqflags indicates work is pending.
1117 * This will work inside a critical section but not inside a hard code
1118 * section.
1119 *
1120 * (self contained on a per cpu basis)
1121 */
1122void
1123splz_check(void)
1124{
1125 globaldata_t gd = mycpu;
1126 thread_t td = gd->gd_curthread;
1127
1128 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) &&
1129 gd->gd_intr_nesting_level == 0 &&
1130 td->td_nest_count < 2)
1131 {
1132 splz();
1133 }
1134}
1135
1136/*
1137 * This version is integrated into crit_exit, reqflags has already
1138 * been tested but td_critcount has not.
1139 *
1140 * We only want to execute the splz() on the 1->0 transition of
1141 * critcount and not in a hard code section or if too deeply nested.
1142 */
1143void
1144lwkt_maybe_splz(thread_t td)
1145{
1146 globaldata_t gd = td->td_gd;
1147
1148 if (td->td_critcount == 0 &&
1149 gd->gd_intr_nesting_level == 0 &&
1150 td->td_nest_count < 2)
1151 {
1152 splz();
1153 }
1154}
1155
1156/*
1157 * This function is used to negotiate a passive release of the current
1158 * process/lwp designation with the user scheduler, allowing the user
1159 * scheduler to schedule another user thread. The related kernel thread
1160 * (curthread) continues running in the released state.
1161 */
1162void
1163lwkt_passive_release(struct thread *td)
1164{
1165 struct lwp *lp = td->td_lwp;
1166
1167 td->td_release = NULL;
1168 lwkt_setpri_self(TDPRI_KERN_USER);
1169 lp->lwp_proc->p_usched->release_curproc(lp);
1170}
1171
1172
1173/*
1174 * This implements a normal yield. This routine is virtually a nop if
1175 * there is nothing to yield to but it will always run any pending interrupts
1176 * if called from a critical section.
1177 *
1178 * This yield is designed for kernel threads without a user context.
1179 *
1180 * (self contained on a per cpu basis)
1181 */
1182void
1183lwkt_yield(void)
1184{
1185 globaldata_t gd = mycpu;
1186 thread_t td = gd->gd_curthread;
1187 thread_t xtd;
1188
1189 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1190 splz();
1191 if (td->td_fairq_accum < 0) {
1192 lwkt_schedule_self(curthread);
1193 lwkt_switch();
1194 } else {
1195 xtd = TAILQ_FIRST(&gd->gd_tdrunq);
1196 if (xtd && xtd->td_pri > td->td_pri) {
1197 lwkt_schedule_self(curthread);
1198 lwkt_switch();
1199 }
1200 }
1201}
1202
1203/*
1204 * This yield is designed for kernel threads with a user context.
1205 *
1206 * The kernel acting on behalf of the user is potentially cpu-bound,
1207 * this function will efficiently allow other threads to run and also
1208 * switch to other processes by releasing.
1209 *
1210 * The lwkt_user_yield() function is designed to have very low overhead
1211 * if no yield is determined to be needed.
1212 */
1213void
1214lwkt_user_yield(void)
1215{
1216 globaldata_t gd = mycpu;
1217 thread_t td = gd->gd_curthread;
1218
1219 /*
1220 * Always run any pending interrupts in case we are in a critical
1221 * section.
1222 */
1223 if ((gd->gd_reqflags & RQF_IDLECHECK_MASK) && td->td_nest_count < 2)
1224 splz();
1225
1226 /*
1227 * Switch (which forces a release) if another kernel thread needs
1228 * the cpu, if userland wants us to resched, or if our kernel
1229 * quantum has run out.
1230 */
1231 if (lwkt_resched_wanted() ||
1232 user_resched_wanted() ||
1233 td->td_fairq_accum < 0)
1234 {
1235 lwkt_switch();
1236 }
1237
1238#if 0
1239 /*
1240 * Reacquire the current process if we are released.
1241 *
1242 * XXX not implemented atm. The kernel may be holding locks and such,
1243 * so we want the thread to continue to receive cpu.
1244 */
1245 if (td->td_release == NULL && lp) {
1246 lp->lwp_proc->p_usched->acquire_curproc(lp);
1247 td->td_release = lwkt_passive_release;
1248 lwkt_setpri_self(TDPRI_USER_NORM);
1249 }
1250#endif
1251}
1252
1253/*
1254 * Generic schedule. Possibly schedule threads belonging to other cpus and
1255 * deal with threads that might be blocked on a wait queue.
1256 *
1257 * We have a little helper inline function which does additional work after
1258 * the thread has been enqueued, including dealing with preemption and
1259 * setting need_lwkt_resched() (which prevents the kernel from returning
1260 * to userland until it has processed higher priority threads).
1261 *
1262 * It is possible for this routine to be called after a failed _enqueue
1263 * (due to the target thread migrating, sleeping, or otherwise blocked).
1264 * We have to check that the thread is actually on the run queue!
1265 *
1266 * reschedok is an optimized constant propagated from lwkt_schedule() or
1267 * lwkt_schedule_noresched(). By default it is non-zero, causing a
1268 * reschedule to be requested if the target thread has a higher priority.
1269 * The port messaging code will set MSG_NORESCHED and cause reschedok to
1270 * be 0, prevented undesired reschedules.
1271 */
1272static __inline
1273void
1274_lwkt_schedule_post(globaldata_t gd, thread_t ntd, int ccount, int reschedok)
1275{
1276 thread_t otd;
1277
1278 if (ntd->td_flags & TDF_RUNQ) {
1279 if (ntd->td_preemptable && reschedok) {
1280 ntd->td_preemptable(ntd, ccount); /* YYY +token */
1281 } else if (reschedok) {
1282 otd = curthread;
1283 if (ntd->td_pri > otd->td_pri)
1284 need_lwkt_resched();
1285 }
1286
1287 /*
1288 * Give the thread a little fair share scheduler bump if it
1289 * has been asleep for a while. This is primarily to avoid
1290 * a degenerate case for interrupt threads where accumulator
1291 * crosses into negative territory unnecessarily.
1292 */
1293 if (ntd->td_fairq_lticks != ticks) {
1294 ntd->td_fairq_lticks = ticks;
1295 ntd->td_fairq_accum += gd->gd_fairq_total_pri;
1296 if (ntd->td_fairq_accum > TDFAIRQ_MAX(gd))
1297 ntd->td_fairq_accum = TDFAIRQ_MAX(gd);
1298 }
1299 }
1300}
1301
1302static __inline
1303void
1304_lwkt_schedule(thread_t td, int reschedok)
1305{
1306 globaldata_t mygd = mycpu;
1307
1308 KASSERT(td != &td->td_gd->gd_idlethread,
1309 ("lwkt_schedule(): scheduling gd_idlethread is illegal!"));
1310 KKASSERT((td->td_flags & TDF_MIGRATING) == 0);
1311 crit_enter_gd(mygd);
1312 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1313 if (td == mygd->gd_curthread) {
1314 _lwkt_enqueue(td);
1315 } else {
1316 /*
1317 * If we own the thread, there is no race (since we are in a
1318 * critical section). If we do not own the thread there might
1319 * be a race but the target cpu will deal with it.
1320 */
1321#ifdef SMP
1322 if (td->td_gd == mygd) {
1323 _lwkt_enqueue(td);
1324 _lwkt_schedule_post(mygd, td, 1, reschedok);
1325 } else {
1326 lwkt_send_ipiq3(td->td_gd, lwkt_schedule_remote, td, 0);
1327 }
1328#else
1329 _lwkt_enqueue(td);
1330 _lwkt_schedule_post(mygd, td, 1, reschedok);
1331#endif
1332 }
1333 crit_exit_gd(mygd);
1334}
1335
1336void
1337lwkt_schedule(thread_t td)
1338{
1339 _lwkt_schedule(td, 1);
1340}
1341
1342void
1343lwkt_schedule_noresched(thread_t td)
1344{
1345 _lwkt_schedule(td, 0);
1346}
1347
1348#ifdef SMP
1349
1350/*
1351 * When scheduled remotely if frame != NULL the IPIQ is being
1352 * run via doreti or an interrupt then preemption can be allowed.
1353 *
1354 * To allow preemption we have to drop the critical section so only
1355 * one is present in _lwkt_schedule_post.
1356 */
1357static void
1358lwkt_schedule_remote(void *arg, int arg2, struct intrframe *frame)
1359{
1360 thread_t td = curthread;
1361 thread_t ntd = arg;
1362
1363 if (frame && ntd->td_preemptable) {
1364 crit_exit_noyield(td);
1365 _lwkt_schedule(ntd, 1);
1366 crit_enter_quick(td);
1367 } else {
1368 _lwkt_schedule(ntd, 1);
1369 }
1370}
1371
1372/*
1373 * Thread migration using a 'Pull' method. The thread may or may not be
1374 * the current thread. It MUST be descheduled and in a stable state.
1375 * lwkt_giveaway() must be called on the cpu owning the thread.
1376 *
1377 * At any point after lwkt_giveaway() is called, the target cpu may
1378 * 'pull' the thread by calling lwkt_acquire().
1379 *
1380 * We have to make sure the thread is not sitting on a per-cpu tsleep
1381 * queue or it will blow up when it moves to another cpu.
1382 *
1383 * MPSAFE - must be called under very specific conditions.
1384 */
1385void
1386lwkt_giveaway(thread_t td)
1387{
1388 globaldata_t gd = mycpu;
1389
1390 crit_enter_gd(gd);
1391 if (td->td_flags & TDF_TSLEEPQ)
1392 tsleep_remove(td);
1393 KKASSERT(td->td_gd == gd);
1394 TAILQ_REMOVE(&gd->gd_tdallq, td, td_allq);
1395 td->td_flags |= TDF_MIGRATING;
1396 crit_exit_gd(gd);
1397}
1398
1399void
1400lwkt_acquire(thread_t td)
1401{
1402 globaldata_t gd;
1403 globaldata_t mygd;
1404 int retry = 10000000;
1405
1406 KKASSERT(td->td_flags & TDF_MIGRATING);
1407 gd = td->td_gd;
1408 mygd = mycpu;
1409 if (gd != mycpu) {
1410 cpu_lfence();
1411 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1412 crit_enter_gd(mygd);
1413 DEBUG_PUSH_INFO("lwkt_acquire");
1414 while (td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) {
1415#ifdef SMP
1416 lwkt_process_ipiq();
1417#endif
1418 cpu_lfence();
1419 if (--retry == 0) {
1420 kprintf("lwkt_acquire: stuck: td %p td->td_flags %08x\n",
1421 td, td->td_flags);
1422 retry = 10000000;
1423 }
1424 }
1425 DEBUG_POP_INFO();
1426 cpu_mfence();
1427 td->td_gd = mygd;
1428 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1429 td->td_flags &= ~TDF_MIGRATING;
1430 crit_exit_gd(mygd);
1431 } else {
1432 crit_enter_gd(mygd);
1433 TAILQ_INSERT_TAIL(&mygd->gd_tdallq, td, td_allq);
1434 td->td_flags &= ~TDF_MIGRATING;
1435 crit_exit_gd(mygd);
1436 }
1437}
1438
1439#endif
1440
1441/*
1442 * Generic deschedule. Descheduling threads other then your own should be
1443 * done only in carefully controlled circumstances. Descheduling is
1444 * asynchronous.
1445 *
1446 * This function may block if the cpu has run out of messages.
1447 */
1448void
1449lwkt_deschedule(thread_t td)
1450{
1451 crit_enter();
1452#ifdef SMP
1453 if (td == curthread) {
1454 _lwkt_dequeue(td);
1455 } else {
1456 if (td->td_gd == mycpu) {
1457 _lwkt_dequeue(td);
1458 } else {
1459 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_deschedule, td);
1460 }
1461 }
1462#else
1463 _lwkt_dequeue(td);
1464#endif
1465 crit_exit();
1466}
1467
1468/*
1469 * Set the target thread's priority. This routine does not automatically
1470 * switch to a higher priority thread, LWKT threads are not designed for
1471 * continuous priority changes. Yield if you want to switch.
1472 */
1473void
1474lwkt_setpri(thread_t td, int pri)
1475{
1476 KKASSERT(td->td_gd == mycpu);
1477 if (td->td_pri != pri) {
1478 KKASSERT(pri >= 0);
1479 crit_enter();
1480 if (td->td_flags & TDF_RUNQ) {
1481 _lwkt_dequeue(td);
1482 td->td_pri = pri;
1483 _lwkt_enqueue(td);
1484 } else {
1485 td->td_pri = pri;
1486 }
1487 crit_exit();
1488 }
1489}
1490
1491/*
1492 * Set the initial priority for a thread prior to it being scheduled for
1493 * the first time. The thread MUST NOT be scheduled before or during
1494 * this call. The thread may be assigned to a cpu other then the current
1495 * cpu.
1496 *
1497 * Typically used after a thread has been created with TDF_STOPPREQ,
1498 * and before the thread is initially scheduled.
1499 */
1500void
1501lwkt_setpri_initial(thread_t td, int pri)
1502{
1503 KKASSERT(pri >= 0);
1504 KKASSERT((td->td_flags & TDF_RUNQ) == 0);
1505 td->td_pri = pri;
1506}
1507
1508void
1509lwkt_setpri_self(int pri)
1510{
1511 thread_t td = curthread;
1512
1513 KKASSERT(pri >= 0 && pri <= TDPRI_MAX);
1514 crit_enter();
1515 if (td->td_flags & TDF_RUNQ) {
1516 _lwkt_dequeue(td);
1517 td->td_pri = pri;
1518 _lwkt_enqueue(td);
1519 } else {
1520 td->td_pri = pri;
1521 }
1522 crit_exit();
1523}
1524
1525/*
1526 * 1/hz tick (typically 10ms) x TDFAIRQ_SCALE (typ 8) = 80ms full cycle.
1527 *
1528 * Example: two competing threads, same priority N. decrement by (2*N)
1529 * increment by N*8, each thread will get 4 ticks.
1530 */
1531void
1532lwkt_fairq_schedulerclock(thread_t td)
1533{
1534 globaldata_t gd;
1535
1536 if (fairq_enable) {
1537 while (td) {
1538 gd = td->td_gd;
1539 if (td != &gd->gd_idlethread) {
1540 td->td_fairq_accum -= gd->gd_fairq_total_pri;
1541 if (td->td_fairq_accum < -TDFAIRQ_MAX(gd))
1542 td->td_fairq_accum = -TDFAIRQ_MAX(gd);
1543 if (td->td_fairq_accum < 0)
1544 need_lwkt_resched();
1545 td->td_fairq_lticks = ticks;
1546 }
1547 td = td->td_preempted;
1548 }
1549 }
1550}
1551
1552static void
1553lwkt_fairq_accumulate(globaldata_t gd, thread_t td)
1554{
1555 td->td_fairq_accum += td->td_pri * TDFAIRQ_SCALE;
1556 if (td->td_fairq_accum > TDFAIRQ_MAX(td->td_gd))
1557 td->td_fairq_accum = TDFAIRQ_MAX(td->td_gd);
1558}
1559
1560/*
1561 * Migrate the current thread to the specified cpu.
1562 *
1563 * This is accomplished by descheduling ourselves from the current cpu
1564 * and setting td_migrate_gd. The lwkt_switch() code will detect that the
1565 * 'old' thread wants to migrate after it has been completely switched out
1566 * and will complete the migration.
1567 *
1568 * TDF_MIGRATING prevents scheduling races while the thread is being migrated.
1569 *
1570 * We must be sure to release our current process designation (if a user
1571 * process) before clearing out any tsleepq we are on because the release
1572 * code may re-add us.
1573 *
1574 * We must be sure to remove ourselves from the current cpu's tsleepq
1575 * before potentially moving to another queue. The thread can be on
1576 * a tsleepq due to a left-over tsleep_interlock().
1577 */
1578
1579void
1580lwkt_setcpu_self(globaldata_t rgd)
1581{
1582#ifdef SMP
1583 thread_t td = curthread;
1584
1585 if (td->td_gd != rgd) {
1586 crit_enter_quick(td);
1587
1588 if (td->td_release)
1589 td->td_release(td);
1590 if (td->td_flags & TDF_TSLEEPQ)
1591 tsleep_remove(td);
1592
1593 /*
1594 * Set TDF_MIGRATING to prevent a spurious reschedule while we are
1595 * trying to deschedule ourselves and switch away, then deschedule
1596 * ourself, remove us from tdallq, and set td_migrate_gd. Finally,
1597 * call lwkt_switch() to complete the operation.
1598 */
1599 td->td_flags |= TDF_MIGRATING;
1600 lwkt_deschedule_self(td);
1601 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1602 td->td_migrate_gd = rgd;
1603 lwkt_switch();
1604
1605 /*
1606 * We are now on the target cpu
1607 */
1608 KKASSERT(rgd == mycpu);
1609 TAILQ_INSERT_TAIL(&rgd->gd_tdallq, td, td_allq);
1610 crit_exit_quick(td);
1611 }
1612#endif
1613}
1614
1615void
1616lwkt_migratecpu(int cpuid)
1617{
1618#ifdef SMP
1619 globaldata_t rgd;
1620
1621 rgd = globaldata_find(cpuid);
1622 lwkt_setcpu_self(rgd);
1623#endif
1624}
1625
1626#ifdef SMP
1627/*
1628 * Remote IPI for cpu migration (called while in a critical section so we
1629 * do not have to enter another one).
1630 *
1631 * The thread (td) has already been completely descheduled from the
1632 * originating cpu and we can simply assert the case. The thread is
1633 * assigned to the new cpu and enqueued.
1634 *
1635 * The thread will re-add itself to tdallq when it resumes execution.
1636 */
1637static void
1638lwkt_setcpu_remote(void *arg)
1639{
1640 thread_t td = arg;
1641 globaldata_t gd = mycpu;
1642
1643 KKASSERT((td->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1644 td->td_gd = gd;
1645 cpu_mfence();
1646 td->td_flags &= ~TDF_MIGRATING;
1647 KKASSERT(td->td_migrate_gd == NULL);
1648 KKASSERT(td->td_lwp == NULL || (td->td_lwp->lwp_flag & LWP_ONRUNQ) == 0);
1649 _lwkt_enqueue(td);
1650}
1651#endif
1652
1653struct lwp *
1654lwkt_preempted_proc(void)
1655{
1656 thread_t td = curthread;
1657 while (td->td_preempted)
1658 td = td->td_preempted;
1659 return(td->td_lwp);
1660}
1661
1662/*
1663 * Create a kernel process/thread/whatever. It shares it's address space
1664 * with proc0 - ie: kernel only.
1665 *
1666 * NOTE! By default new threads are created with the MP lock held. A
1667 * thread which does not require the MP lock should release it by calling
1668 * rel_mplock() at the start of the new thread.
1669 */
1670int
1671lwkt_create(void (*func)(void *), void *arg, struct thread **tdp,
1672 thread_t template, int tdflags, int cpu, const char *fmt, ...)
1673{
1674 thread_t td;
1675 __va_list ap;
1676
1677 td = lwkt_alloc_thread(template, LWKT_THREAD_STACK, cpu,
1678 tdflags);
1679 if (tdp)
1680 *tdp = td;
1681 cpu_set_thread_handler(td, lwkt_exit, func, arg);
1682
1683 /*
1684 * Set up arg0 for 'ps' etc
1685 */
1686 __va_start(ap, fmt);
1687 kvsnprintf(td->td_comm, sizeof(td->td_comm), fmt, ap);
1688 __va_end(ap);
1689
1690 /*
1691 * Schedule the thread to run
1692 */
1693 if ((td->td_flags & TDF_STOPREQ) == 0)
1694 lwkt_schedule(td);
1695 else
1696 td->td_flags &= ~TDF_STOPREQ;
1697 return 0;
1698}
1699
1700/*
1701 * Destroy an LWKT thread. Warning! This function is not called when
1702 * a process exits, cpu_proc_exit() directly calls cpu_thread_exit() and
1703 * uses a different reaping mechanism.
1704 */
1705void
1706lwkt_exit(void)
1707{
1708 thread_t td = curthread;
1709 thread_t std;
1710 globaldata_t gd;
1711
1712 /*
1713 * Do any cleanup that might block here
1714 */
1715 if (td->td_flags & TDF_VERBOSE)
1716 kprintf("kthread %p %s has exited\n", td, td->td_comm);
1717 caps_exit(td);
1718 biosched_done(td);
1719 dsched_exit_thread(td);
1720
1721 /*
1722 * Get us into a critical section to interlock gd_freetd and loop
1723 * until we can get it freed.
1724 *
1725 * We have to cache the current td in gd_freetd because objcache_put()ing
1726 * it would rip it out from under us while our thread is still active.
1727 */
1728 gd = mycpu;
1729 crit_enter_quick(td);
1730 while ((std = gd->gd_freetd) != NULL) {
1731 KKASSERT((std->td_flags & (TDF_RUNNING|TDF_PREEMPT_LOCK)) == 0);
1732 gd->gd_freetd = NULL;
1733 objcache_put(thread_cache, std);
1734 }
1735
1736 /*
1737 * Remove thread resources from kernel lists and deschedule us for
1738 * the last time. We cannot block after this point or we may end
1739 * up with a stale td on the tsleepq.
1740 */
1741 if (td->td_flags & TDF_TSLEEPQ)
1742 tsleep_remove(td);
1743 lwkt_deschedule_self(td);
1744 lwkt_remove_tdallq(td);
1745 KKASSERT(td->td_refs == 0);
1746
1747 /*
1748 * Final cleanup
1749 */
1750 KKASSERT(gd->gd_freetd == NULL);
1751 if (td->td_flags & TDF_ALLOCATED_THREAD)
1752 gd->gd_freetd = td;
1753 cpu_thread_exit();
1754}
1755
1756void
1757lwkt_remove_tdallq(thread_t td)
1758{
1759 KKASSERT(td->td_gd == mycpu);
1760 TAILQ_REMOVE(&td->td_gd->gd_tdallq, td, td_allq);
1761}
1762
1763/*
1764 * Code reduction and branch prediction improvements. Call/return
1765 * overhead on modern cpus often degenerates into 0 cycles due to
1766 * the cpu's branch prediction hardware and return pc cache. We
1767 * can take advantage of this by not inlining medium-complexity
1768 * functions and we can also reduce the branch prediction impact
1769 * by collapsing perfectly predictable branches into a single
1770 * procedure instead of duplicating it.
1771 *
1772 * Is any of this noticeable? Probably not, so I'll take the
1773 * smaller code size.
1774 */
1775void
1776crit_exit_wrapper(__DEBUG_CRIT_ARG__)
1777{
1778 _crit_exit(mycpu __DEBUG_CRIT_PASS_ARG__);
1779}
1780
1781void
1782crit_panic(void)
1783{
1784 thread_t td = curthread;
1785 int lcrit = td->td_critcount;
1786
1787 td->td_critcount = 0;
1788 panic("td_critcount is/would-go negative! %p %d", td, lcrit);
1789 /* NOT REACHED */
1790}
1791
1792#ifdef SMP
1793
1794/*
1795 * Called from debugger/panic on cpus which have been stopped. We must still
1796 * process the IPIQ while stopped, even if we were stopped while in a critical
1797 * section (XXX).
1798 *
1799 * If we are dumping also try to process any pending interrupts. This may
1800 * or may not work depending on the state of the cpu at the point it was
1801 * stopped.
1802 */
1803void
1804lwkt_smp_stopped(void)
1805{
1806 globaldata_t gd = mycpu;
1807
1808 crit_enter_gd(gd);
1809 if (dumping) {
1810 lwkt_process_ipiq();
1811 splz();
1812 } else {
1813 lwkt_process_ipiq();
1814 }
1815 crit_exit_gd(gd);
1816}
1817
1818#endif