xref: /dflybsd-src/sys/kern/usched_bsd4.c (revision d4b8aec4bb44a374c3e91969c1a7df6569da7be3)
1 /*
2  * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $DragonFly: src/sys/kern/usched_bsd4.c,v 1.26 2008/11/01 23:31:19 dillon Exp $
27  */
28 
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
32 #include <sys/lock.h>
33 #include <sys/queue.h>
34 #include <sys/proc.h>
35 #include <sys/rtprio.h>
36 #include <sys/uio.h>
37 #include <sys/sysctl.h>
38 #include <sys/resourcevar.h>
39 #include <sys/spinlock.h>
40 #include <machine/cpu.h>
41 #include <machine/smp.h>
42 
43 #include <sys/thread2.h>
44 #include <sys/spinlock2.h>
45 #include <sys/mplock2.h>
46 
47 /*
48  * Priorities.  Note that with 32 run queues per scheduler each queue
49  * represents four priority levels.
50  */
51 
52 #define MAXPRI			128
53 #define PRIMASK			(MAXPRI - 1)
54 #define PRIBASE_REALTIME	0
55 #define PRIBASE_NORMAL		MAXPRI
56 #define PRIBASE_IDLE		(MAXPRI * 2)
57 #define PRIBASE_THREAD		(MAXPRI * 3)
58 #define PRIBASE_NULL		(MAXPRI * 4)
59 
60 #define NQS	32			/* 32 run queues. */
61 #define PPQ	(MAXPRI / NQS)		/* priorities per queue */
62 #define PPQMASK	(PPQ - 1)
63 
64 /*
65  * NICEPPQ	- number of nice units per priority queue
66  * ESTCPURAMP	- number of scheduler ticks for estcpu to switch queues
67  *
68  * ESTCPUPPQ	- number of estcpu units per priority queue
69  * ESTCPUMAX	- number of estcpu units
70  * ESTCPUINCR	- amount we have to increment p_estcpu per scheduling tick at
71  *		  100% cpu.
72  */
73 #define NICEPPQ		2
74 #define ESTCPURAMP	4
75 #define ESTCPUPPQ	512
76 #define ESTCPUMAX	(ESTCPUPPQ * NQS)
77 #define ESTCPUINCR	(ESTCPUPPQ / ESTCPURAMP)
78 #define PRIO_RANGE	(PRIO_MAX - PRIO_MIN + 1)
79 
80 #define ESTCPULIM(v)	min((v), ESTCPUMAX)
81 
82 TAILQ_HEAD(rq, lwp);
83 
84 #define lwp_priority	lwp_usdata.bsd4.priority
85 #define lwp_rqindex	lwp_usdata.bsd4.rqindex
86 #define lwp_origcpu	lwp_usdata.bsd4.origcpu
87 #define lwp_estcpu	lwp_usdata.bsd4.estcpu
88 #define lwp_rqtype	lwp_usdata.bsd4.rqtype
89 
90 static void bsd4_acquire_curproc(struct lwp *lp);
91 static void bsd4_release_curproc(struct lwp *lp);
92 static void bsd4_select_curproc(globaldata_t gd);
93 static void bsd4_setrunqueue(struct lwp *lp);
94 static void bsd4_schedulerclock(struct lwp *lp, sysclock_t period,
95 				sysclock_t cpstamp);
96 static void bsd4_recalculate_estcpu(struct lwp *lp);
97 static void bsd4_resetpriority(struct lwp *lp);
98 static void bsd4_forking(struct lwp *plp, struct lwp *lp);
99 static void bsd4_exiting(struct lwp *plp, struct lwp *lp);
100 static void bsd4_yield(struct lwp *lp);
101 
102 #ifdef SMP
103 static void need_user_resched_remote(void *dummy);
104 #endif
105 static struct lwp *chooseproc_locked(struct lwp *chklp);
106 static void bsd4_remrunqueue_locked(struct lwp *lp);
107 static void bsd4_setrunqueue_locked(struct lwp *lp);
108 
109 struct usched usched_bsd4 = {
110 	{ NULL },
111 	"bsd4", "Original DragonFly Scheduler",
112 	NULL,			/* default registration */
113 	NULL,			/* default deregistration */
114 	bsd4_acquire_curproc,
115 	bsd4_release_curproc,
116 	bsd4_setrunqueue,
117 	bsd4_schedulerclock,
118 	bsd4_recalculate_estcpu,
119 	bsd4_resetpriority,
120 	bsd4_forking,
121 	bsd4_exiting,
122 	NULL,			/* setcpumask not supported */
123 	bsd4_yield
124 };
125 
126 struct usched_bsd4_pcpu {
127 	struct thread helper_thread;
128 	short	rrcount;
129 	short	upri;
130 	struct lwp *uschedcp;
131 };
132 
133 typedef struct usched_bsd4_pcpu	*bsd4_pcpu_t;
134 
135 /*
136  * We have NQS (32) run queues per scheduling class.  For the normal
137  * class, there are 128 priorities scaled onto these 32 queues.  New
138  * processes are added to the last entry in each queue, and processes
139  * are selected for running by taking them from the head and maintaining
140  * a simple FIFO arrangement.  Realtime and Idle priority processes have
141  * and explicit 0-31 priority which maps directly onto their class queue
142  * index.  When a queue has something in it, the corresponding bit is
143  * set in the queuebits variable, allowing a single read to determine
144  * the state of all 32 queues and then a ffs() to find the first busy
145  * queue.
146  */
147 static struct rq bsd4_queues[NQS];
148 static struct rq bsd4_rtqueues[NQS];
149 static struct rq bsd4_idqueues[NQS];
150 static u_int32_t bsd4_queuebits;
151 static u_int32_t bsd4_rtqueuebits;
152 static u_int32_t bsd4_idqueuebits;
153 static cpumask_t bsd4_curprocmask = -1;	/* currently running a user process */
154 static cpumask_t bsd4_rdyprocmask;	/* ready to accept a user process */
155 static int	 bsd4_runqcount;
156 #ifdef SMP
157 static volatile int bsd4_scancpu;
158 #endif
159 static struct spinlock bsd4_spin;
160 static struct usched_bsd4_pcpu bsd4_pcpu[MAXCPU];
161 
162 SYSCTL_INT(_debug, OID_AUTO, bsd4_runqcount, CTLFLAG_RD, &bsd4_runqcount, 0, "");
163 #ifdef INVARIANTS
164 static int usched_nonoptimal;
165 SYSCTL_INT(_debug, OID_AUTO, usched_nonoptimal, CTLFLAG_RW,
166         &usched_nonoptimal, 0, "acquire_curproc() was not optimal");
167 static int usched_optimal;
168 SYSCTL_INT(_debug, OID_AUTO, usched_optimal, CTLFLAG_RW,
169         &usched_optimal, 0, "acquire_curproc() was optimal");
170 #endif
171 static int usched_debug = -1;
172 SYSCTL_INT(_debug, OID_AUTO, scdebug, CTLFLAG_RW, &usched_debug, 0, "");
173 #ifdef SMP
174 static int remote_resched_nonaffinity;
175 static int remote_resched_affinity;
176 static int choose_affinity;
177 SYSCTL_INT(_debug, OID_AUTO, remote_resched_nonaffinity, CTLFLAG_RD,
178         &remote_resched_nonaffinity, 0, "Number of remote rescheds");
179 SYSCTL_INT(_debug, OID_AUTO, remote_resched_affinity, CTLFLAG_RD,
180         &remote_resched_affinity, 0, "Number of remote rescheds");
181 SYSCTL_INT(_debug, OID_AUTO, choose_affinity, CTLFLAG_RD,
182         &choose_affinity, 0, "chooseproc() was smart");
183 #endif
184 
185 static int usched_bsd4_rrinterval = (ESTCPUFREQ + 9) / 10;
186 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_rrinterval, CTLFLAG_RW,
187         &usched_bsd4_rrinterval, 0, "");
188 static int usched_bsd4_decay = ESTCPUINCR / 2;
189 SYSCTL_INT(_kern, OID_AUTO, usched_bsd4_decay, CTLFLAG_RW,
190         &usched_bsd4_decay, 0, "");
191 
192 /*
193  * Initialize the run queues at boot time.
194  */
195 static void
196 rqinit(void *dummy)
197 {
198 	int i;
199 
200 	spin_init(&bsd4_spin);
201 	for (i = 0; i < NQS; i++) {
202 		TAILQ_INIT(&bsd4_queues[i]);
203 		TAILQ_INIT(&bsd4_rtqueues[i]);
204 		TAILQ_INIT(&bsd4_idqueues[i]);
205 	}
206 	atomic_clear_int(&bsd4_curprocmask, 1);
207 }
208 SYSINIT(runqueue, SI_BOOT2_USCHED, SI_ORDER_FIRST, rqinit, NULL)
209 
210 /*
211  * BSD4_ACQUIRE_CURPROC
212  *
213  * This function is called when the kernel intends to return to userland.
214  * It is responsible for making the thread the current designated userland
215  * thread for this cpu, blocking if necessary.
216  *
217  * The kernel has already depressed our LWKT priority so we must not switch
218  * until we have either assigned or disposed of the thread.
219  *
220  * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
221  * TO ANOTHER CPU!  Because most of the kernel assumes that no migration will
222  * occur, this function is called only under very controlled circumstances.
223  *
224  * MPSAFE
225  */
226 static void
227 bsd4_acquire_curproc(struct lwp *lp)
228 {
229 	globaldata_t gd;
230 	bsd4_pcpu_t dd;
231 	struct lwp *olp;
232 
233 	crit_enter();
234 	bsd4_recalculate_estcpu(lp);
235 
236 	/*
237 	 * If a reschedule was requested give another thread the
238 	 * driver's seat.
239 	 */
240 	if (user_resched_wanted()) {
241 		clear_user_resched();
242 		bsd4_release_curproc(lp);
243 	}
244 
245 	/*
246 	 * Loop until we are the current user thread
247 	 */
248 	do {
249 		/*
250 		 * Reload after a switch or setrunqueue/switch possibly
251 		 * moved us to another cpu.
252 		 */
253 		clear_lwkt_resched();
254 		gd = mycpu;
255 		dd = &bsd4_pcpu[gd->gd_cpuid];
256 
257 		/*
258 		 * Become the currently scheduled user thread for this cpu
259 		 * if we can do so trivially.
260 		 *
261 		 * We can steal another thread's current thread designation
262 		 * on this cpu since if we are running that other thread
263 		 * must not be, so we can safely deschedule it.
264 		 */
265 		if (dd->uschedcp == lp) {
266 			dd->upri = lp->lwp_priority;
267 		} else if (dd->uschedcp == NULL) {
268 			atomic_set_int(&bsd4_curprocmask, gd->gd_cpumask);
269 			dd->uschedcp = lp;
270 			dd->upri = lp->lwp_priority;
271 		} else if (dd->upri > lp->lwp_priority) {
272 			olp = dd->uschedcp;
273 			dd->uschedcp = lp;
274 			dd->upri = lp->lwp_priority;
275 			lwkt_deschedule(olp->lwp_thread);
276 			bsd4_setrunqueue(olp);
277 		} else {
278 			lwkt_deschedule(lp->lwp_thread);
279 			bsd4_setrunqueue(lp);
280 			lwkt_switch();
281 		}
282 
283 		/*
284 		 * Other threads at our current user priority have already
285 		 * put in their bids, but we must run any kernel threads
286 		 * at higher priorities, and we could lose our bid to
287 		 * another thread trying to return to user mode in the
288 		 * process.
289 		 *
290 		 * If we lose our bid we will be descheduled and put on
291 		 * the run queue.  When we are reactivated we will have
292 		 * another chance.
293 		 */
294 		lwkt_yield();
295 	} while (dd->uschedcp != lp);
296 
297 	crit_exit();
298 	KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
299 }
300 
301 /*
302  * BSD4_RELEASE_CURPROC
303  *
304  * This routine detaches the current thread from the userland scheduler,
305  * usually because the thread needs to run or block in the kernel (at
306  * kernel priority) for a while.
307  *
308  * This routine is also responsible for selecting a new thread to
309  * make the current thread.
310  *
311  * NOTE: This implementation differs from the dummy example in that
312  * bsd4_select_curproc() is able to select the current process, whereas
313  * dummy_select_curproc() is not able to select the current process.
314  * This means we have to NULL out uschedcp.
315  *
316  * Additionally, note that we may already be on a run queue if releasing
317  * via the lwkt_switch() in bsd4_setrunqueue().
318  *
319  * WARNING!  The MP lock may be in an unsynchronized state due to the
320  * way get_mplock() works and the fact that this function may be called
321  * from a passive release during a lwkt_switch().   try_mplock() will deal
322  * with this for us but you should be aware that td_mpcount may not be
323  * useable.
324  *
325  * MPSAFE
326  */
327 static void
328 bsd4_release_curproc(struct lwp *lp)
329 {
330 	globaldata_t gd = mycpu;
331 	bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
332 
333 	if (dd->uschedcp == lp) {
334 		crit_enter();
335 		KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
336 		dd->uschedcp = NULL;	/* don't let lp be selected */
337 		dd->upri = PRIBASE_NULL;
338 		atomic_clear_int(&bsd4_curprocmask, gd->gd_cpumask);
339 		bsd4_select_curproc(gd);
340 		crit_exit();
341 	}
342 }
343 
344 /*
345  * BSD4_SELECT_CURPROC
346  *
347  * Select a new current process for this cpu and clear any pending user
348  * reschedule request.  The cpu currently has no current process.
349  *
350  * This routine is also responsible for equal-priority round-robining,
351  * typically triggered from bsd4_schedulerclock().  In our dummy example
352  * all the 'user' threads are LWKT scheduled all at once and we just
353  * call lwkt_switch().
354  *
355  * The calling process is not on the queue and cannot be selected.
356  *
357  * MPSAFE
358  */
359 static
360 void
361 bsd4_select_curproc(globaldata_t gd)
362 {
363 	bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
364 	struct lwp *nlp;
365 	int cpuid = gd->gd_cpuid;
366 
367 	crit_enter_gd(gd);
368 
369 	spin_lock_wr(&bsd4_spin);
370 	if ((nlp = chooseproc_locked(dd->uschedcp)) != NULL) {
371 		atomic_set_int(&bsd4_curprocmask, 1 << cpuid);
372 		dd->upri = nlp->lwp_priority;
373 		dd->uschedcp = nlp;
374 		spin_unlock_wr(&bsd4_spin);
375 #ifdef SMP
376 		lwkt_acquire(nlp->lwp_thread);
377 #endif
378 		lwkt_schedule(nlp->lwp_thread);
379 	} else if (bsd4_runqcount && (bsd4_rdyprocmask & (1 << cpuid))) {
380 		atomic_clear_int(&bsd4_rdyprocmask, 1 << cpuid);
381 		spin_unlock_wr(&bsd4_spin);
382 		lwkt_schedule(&dd->helper_thread);
383 	} else {
384 		spin_unlock_wr(&bsd4_spin);
385 	}
386 	crit_exit_gd(gd);
387 }
388 
389 /*
390  * BSD4_SETRUNQUEUE
391  *
392  * Place the specified lwp on the user scheduler's run queue.  This routine
393  * must be called with the thread descheduled.  The lwp must be runnable.
394  *
395  * The thread may be the current thread as a special case.
396  *
397  * MPSAFE
398  */
399 static void
400 bsd4_setrunqueue(struct lwp *lp)
401 {
402 	globaldata_t gd;
403 	bsd4_pcpu_t dd;
404 #ifdef SMP
405 	int cpuid;
406 	cpumask_t mask;
407 	cpumask_t tmpmask;
408 #endif
409 
410 	/*
411 	 * First validate the process state relative to the current cpu.
412 	 * We don't need the spinlock for this, just a critical section.
413 	 * We are in control of the process.
414 	 */
415 	crit_enter();
416 	KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
417 	KASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0,
418 	    ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
419 	     lp->lwp_tid, lp->lwp_proc->p_flag, lp->lwp_flag));
420 	KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
421 
422 	/*
423 	 * Note: gd and dd are relative to the target thread's last cpu,
424 	 * NOT our current cpu.
425 	 */
426 	gd = lp->lwp_thread->td_gd;
427 	dd = &bsd4_pcpu[gd->gd_cpuid];
428 
429 	/*
430 	 * This process is not supposed to be scheduled anywhere or assigned
431 	 * as the current process anywhere.  Assert the condition.
432 	 */
433 	KKASSERT(dd->uschedcp != lp);
434 
435 #ifndef SMP
436 	/*
437 	 * If we are not SMP we do not have a scheduler helper to kick
438 	 * and must directly activate the process if none are scheduled.
439 	 *
440 	 * This is really only an issue when bootstrapping init since
441 	 * the caller in all other cases will be a user process, and
442 	 * even if released (dd->uschedcp == NULL), that process will
443 	 * kickstart the scheduler when it returns to user mode from
444 	 * the kernel.
445 	 */
446 	if (dd->uschedcp == NULL) {
447 		atomic_set_int(&bsd4_curprocmask, gd->gd_cpumask);
448 		dd->uschedcp = lp;
449 		dd->upri = lp->lwp_priority;
450 		lwkt_schedule(lp->lwp_thread);
451 		crit_exit();
452 		return;
453 	}
454 #endif
455 
456 #ifdef SMP
457 	/*
458 	 * XXX fixme.  Could be part of a remrunqueue/setrunqueue
459 	 * operation when the priority is recalculated, so TDF_MIGRATING
460 	 * may already be set.
461 	 */
462 	if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
463 		lwkt_giveaway(lp->lwp_thread);
464 #endif
465 
466 	/*
467 	 * We lose control of lp the moment we release the spinlock after
468 	 * having placed lp on the queue.  i.e. another cpu could pick it
469 	 * up and it could exit, or its priority could be further adjusted,
470 	 * or something like that.
471 	 */
472 	spin_lock_wr(&bsd4_spin);
473 	bsd4_setrunqueue_locked(lp);
474 
475 #ifdef SMP
476 	/*
477 	 * Kick the scheduler helper on one of the other cpu's
478 	 * and request a reschedule if appropriate.
479 	 */
480 	cpuid = (bsd4_scancpu & 0xFFFF) % ncpus;
481 	++bsd4_scancpu;
482 	mask = ~bsd4_curprocmask & bsd4_rdyprocmask &
483 		lp->lwp_cpumask & smp_active_mask;
484 	spin_unlock_wr(&bsd4_spin);
485 
486 	while (mask) {
487 		tmpmask = ~((1 << cpuid) - 1);
488 		if (mask & tmpmask)
489 			cpuid = bsfl(mask & tmpmask);
490 		else
491 			cpuid = bsfl(mask);
492 		gd = globaldata_find(cpuid);
493 		dd = &bsd4_pcpu[cpuid];
494 
495 		if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
496 			if (gd == mycpu)
497 				need_user_resched_remote(NULL);
498 			else
499 				lwkt_send_ipiq(gd, need_user_resched_remote, NULL);
500 			break;
501 		}
502 		mask &= ~(1 << cpuid);
503 	}
504 #else
505 	/*
506 	 * Request a reschedule if appropriate.
507 	 */
508 	spin_unlock_wr(&bsd4_spin);
509 	if ((dd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
510 		need_user_resched();
511 	}
512 #endif
513 	crit_exit();
514 }
515 
516 /*
517  * This routine is called from a systimer IPI.  It MUST be MP-safe and
518  * the BGL IS NOT HELD ON ENTRY.  This routine is called at ESTCPUFREQ on
519  * each cpu.
520  *
521  * Because this is effectively a 'fast' interrupt, we cannot safely
522  * use spinlocks unless gd_spinlock_rd is NULL and gd_spinlocks_wr is 0,
523  * even if the spinlocks are 'non conflicting'.  This is due to the way
524  * spinlock conflicts against cached read locks are handled.
525  *
526  * MPSAFE
527  */
528 static
529 void
530 bsd4_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
531 {
532 	globaldata_t gd = mycpu;
533 	bsd4_pcpu_t dd = &bsd4_pcpu[gd->gd_cpuid];
534 
535 	/*
536 	 * Do we need to round-robin?  We round-robin 10 times a second.
537 	 * This should only occur for cpu-bound batch processes.
538 	 */
539 	if (++dd->rrcount >= usched_bsd4_rrinterval) {
540 		dd->rrcount = 0;
541 		need_user_resched();
542 	}
543 
544 	/*
545 	 * As the process accumulates cpu time p_estcpu is bumped and may
546 	 * push the process into another scheduling queue.  It typically
547 	 * takes 4 ticks to bump the queue.
548 	 */
549 	lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
550 
551 	/*
552 	 * Reducing p_origcpu over time causes more of our estcpu to be
553 	 * returned to the parent when we exit.  This is a small tweak
554 	 * for the batch detection heuristic.
555 	 */
556 	if (lp->lwp_origcpu)
557 		--lp->lwp_origcpu;
558 
559 	/*
560 	 * We can only safely call bsd4_resetpriority(), which uses spinlocks,
561 	 * if we aren't interrupting a thread that is using spinlocks.
562 	 * Otherwise we can deadlock with another cpu waiting for our read
563 	 * spinlocks to clear.
564 	 */
565 	if (gd->gd_spinlock_rd == NULL && gd->gd_spinlocks_wr == 0)
566 		bsd4_resetpriority(lp);
567 	else
568 		need_user_resched();
569 }
570 
571 /*
572  * Called from acquire and from kern_synch's one-second timer (one of the
573  * callout helper threads) with a critical section held.
574  *
575  * Decay p_estcpu based on the number of ticks we haven't been running
576  * and our p_nice.  As the load increases each process observes a larger
577  * number of idle ticks (because other processes are running in them).
578  * This observation leads to a larger correction which tends to make the
579  * system more 'batchy'.
580  *
581  * Note that no recalculation occurs for a process which sleeps and wakes
582  * up in the same tick.  That is, a system doing thousands of context
583  * switches per second will still only do serious estcpu calculations
584  * ESTCPUFREQ times per second.
585  *
586  * MPSAFE
587  */
588 static
589 void
590 bsd4_recalculate_estcpu(struct lwp *lp)
591 {
592 	globaldata_t gd = mycpu;
593 	sysclock_t cpbase;
594 	int loadfac;
595 	int ndecay;
596 	int nticks;
597 	int nleft;
598 
599 	/*
600 	 * We have to subtract periodic to get the last schedclock
601 	 * timeout time, otherwise we would get the upcoming timeout.
602 	 * Keep in mind that a process can migrate between cpus and
603 	 * while the scheduler clock should be very close, boundary
604 	 * conditions could lead to a small negative delta.
605 	 */
606 	cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
607 
608 	if (lp->lwp_slptime > 1) {
609 		/*
610 		 * Too much time has passed, do a coarse correction.
611 		 */
612 		lp->lwp_estcpu = lp->lwp_estcpu >> 1;
613 		bsd4_resetpriority(lp);
614 		lp->lwp_cpbase = cpbase;
615 		lp->lwp_cpticks = 0;
616 	} else if (lp->lwp_cpbase != cpbase) {
617 		/*
618 		 * Adjust estcpu if we are in a different tick.  Don't waste
619 		 * time if we are in the same tick.
620 		 *
621 		 * First calculate the number of ticks in the measurement
622 		 * interval.  The nticks calculation can wind up 0 due to
623 		 * a bug in the handling of lwp_slptime  (as yet not found),
624 		 * so make sure we do not get a divide by 0 panic.
625 		 */
626 		nticks = (cpbase - lp->lwp_cpbase) / gd->gd_schedclock.periodic;
627 		if (nticks <= 0)
628 			nticks = 1;
629 		updatepcpu(lp, lp->lwp_cpticks, nticks);
630 
631 		if ((nleft = nticks - lp->lwp_cpticks) < 0)
632 			nleft = 0;
633 		if (usched_debug == lp->lwp_proc->p_pid) {
634 			kprintf("pid %d tid %d estcpu %d cpticks %d nticks %d nleft %d",
635 				lp->lwp_proc->p_pid, lp->lwp_tid, lp->lwp_estcpu,
636 				lp->lwp_cpticks, nticks, nleft);
637 		}
638 
639 		/*
640 		 * Calculate a decay value based on ticks remaining scaled
641 		 * down by the instantanious load and p_nice.
642 		 */
643 		if ((loadfac = bsd4_runqcount) < 2)
644 			loadfac = 2;
645 		ndecay = nleft * usched_bsd4_decay * 2 *
646 			(PRIO_MAX * 2 - lp->lwp_proc->p_nice) / (loadfac * PRIO_MAX * 2);
647 
648 		/*
649 		 * Adjust p_estcpu.  Handle a border case where batch jobs
650 		 * can get stalled long enough to decay to zero when they
651 		 * shouldn't.
652 		 */
653 		if (lp->lwp_estcpu > ndecay * 2)
654 			lp->lwp_estcpu -= ndecay;
655 		else
656 			lp->lwp_estcpu >>= 1;
657 
658 		if (usched_debug == lp->lwp_proc->p_pid)
659 			kprintf(" ndecay %d estcpu %d\n", ndecay, lp->lwp_estcpu);
660 		bsd4_resetpriority(lp);
661 		lp->lwp_cpbase = cpbase;
662 		lp->lwp_cpticks = 0;
663 	}
664 }
665 
666 /*
667  * Compute the priority of a process when running in user mode.
668  * Arrange to reschedule if the resulting priority is better
669  * than that of the current process.
670  *
671  * This routine may be called with any process.
672  *
673  * This routine is called by fork1() for initial setup with the process
674  * of the run queue, and also may be called normally with the process on or
675  * off the run queue.
676  *
677  * MPSAFE
678  */
679 static void
680 bsd4_resetpriority(struct lwp *lp)
681 {
682 	bsd4_pcpu_t dd;
683 	int newpriority;
684 	u_short newrqtype;
685 	int reschedcpu;
686 
687 	/*
688 	 * Calculate the new priority and queue type
689 	 */
690 	crit_enter();
691 	spin_lock_wr(&bsd4_spin);
692 
693 	newrqtype = lp->lwp_rtprio.type;
694 
695 	switch(newrqtype) {
696 	case RTP_PRIO_REALTIME:
697 	case RTP_PRIO_FIFO:
698 		newpriority = PRIBASE_REALTIME +
699 			     (lp->lwp_rtprio.prio & PRIMASK);
700 		break;
701 	case RTP_PRIO_NORMAL:
702 		newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
703 		newpriority += lp->lwp_estcpu * PPQ / ESTCPUPPQ;
704 		newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
705 			      NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
706 		newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
707 		break;
708 	case RTP_PRIO_IDLE:
709 		newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
710 		break;
711 	case RTP_PRIO_THREAD:
712 		newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
713 		break;
714 	default:
715 		panic("Bad RTP_PRIO %d", newrqtype);
716 		/* NOT REACHED */
717 	}
718 
719 	/*
720 	 * The newpriority incorporates the queue type so do a simple masked
721 	 * check to determine if the process has moved to another queue.  If
722 	 * it has, and it is currently on a run queue, then move it.
723 	 */
724 	if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
725 		lp->lwp_priority = newpriority;
726 		if (lp->lwp_flag & LWP_ONRUNQ) {
727 			bsd4_remrunqueue_locked(lp);
728 			lp->lwp_rqtype = newrqtype;
729 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
730 			bsd4_setrunqueue_locked(lp);
731 			reschedcpu = lp->lwp_thread->td_gd->gd_cpuid;
732 		} else {
733 			lp->lwp_rqtype = newrqtype;
734 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
735 			reschedcpu = -1;
736 		}
737 	} else {
738 		lp->lwp_priority = newpriority;
739 		reschedcpu = -1;
740 	}
741 	spin_unlock_wr(&bsd4_spin);
742 
743 	/*
744 	 * Determine if we need to reschedule the target cpu.  This only
745 	 * occurs if the LWP is already on a scheduler queue, which means
746 	 * that idle cpu notification has already occured.  At most we
747 	 * need only issue a need_user_resched() on the appropriate cpu.
748 	 *
749 	 * The LWP may be owned by a CPU different from the current one,
750 	 * in which case dd->uschedcp may be modified without an MP lock
751 	 * or a spinlock held.  The worst that happens is that the code
752 	 * below causes a spurious need_user_resched() on the target CPU
753 	 * and dd->pri to be wrong for a short period of time, both of
754 	 * which are harmless.
755 	 */
756 	if (reschedcpu >= 0) {
757 		dd = &bsd4_pcpu[reschedcpu];
758 		if ((dd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK)) {
759 			dd->upri = lp->lwp_priority;
760 #ifdef SMP
761 			if (reschedcpu == mycpu->gd_cpuid) {
762 				need_user_resched();
763 			} else {
764 				lwkt_send_ipiq(lp->lwp_thread->td_gd,
765 					       need_user_resched_remote, NULL);
766 			}
767 #else
768 			need_user_resched();
769 #endif
770 		}
771 	}
772 	crit_exit();
773 }
774 
775 /*
776  * MPSAFE
777  */
778 static
779 void
780 bsd4_yield(struct lwp *lp)
781 {
782 #if 0
783 	/* FUTURE (or something similar) */
784 	switch(lp->lwp_rqtype) {
785 	case RTP_PRIO_NORMAL:
786 		lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
787 		break;
788 	default:
789 		break;
790 	}
791 #endif
792         need_user_resched();
793 }
794 
795 /*
796  * Called from fork1() when a new child process is being created.
797  *
798  * Give the child process an initial estcpu that is more batch then
799  * its parent and dock the parent for the fork (but do not
800  * reschedule the parent).   This comprises the main part of our batch
801  * detection heuristic for both parallel forking and sequential execs.
802  *
803  * Interactive processes will decay the boosted estcpu quickly while batch
804  * processes will tend to compound it.
805  * XXX lwp should be "spawning" instead of "forking"
806  *
807  * MPSAFE
808  */
809 static void
810 bsd4_forking(struct lwp *plp, struct lwp *lp)
811 {
812 	lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ);
813 	lp->lwp_origcpu = lp->lwp_estcpu;
814 	plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ);
815 }
816 
817 /*
818  * Called when the parent reaps a child.   Propogate cpu use by the child
819  * back to the parent.
820  *
821  * MPSAFE
822  */
823 static void
824 bsd4_exiting(struct lwp *plp, struct lwp *lp)
825 {
826 	int delta;
827 
828 	if (plp->lwp_proc->p_pid != 1) {
829 		delta = lp->lwp_estcpu - lp->lwp_origcpu;
830 		if (delta > 0)
831 			plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + delta);
832 	}
833 }
834 
835 
836 /*
837  * chooseproc() is called when a cpu needs a user process to LWKT schedule,
838  * it selects a user process and returns it.  If chklp is non-NULL and chklp
839  * has a better or equal priority then the process that would otherwise be
840  * chosen, NULL is returned.
841  *
842  * Until we fix the RUNQ code the chklp test has to be strict or we may
843  * bounce between processes trying to acquire the current process designation.
844  *
845  * MPSAFE - must be called with bsd4_spin exclusive held.  The spinlock is
846  *	    left intact through the entire routine.
847  */
848 static
849 struct lwp *
850 chooseproc_locked(struct lwp *chklp)
851 {
852 	struct lwp *lp;
853 	struct rq *q;
854 	u_int32_t *which, *which2;
855 	u_int32_t pri;
856 	u_int32_t rtqbits;
857 	u_int32_t tsqbits;
858 	u_int32_t idqbits;
859 	cpumask_t cpumask;
860 
861 	rtqbits = bsd4_rtqueuebits;
862 	tsqbits = bsd4_queuebits;
863 	idqbits = bsd4_idqueuebits;
864 	cpumask = mycpu->gd_cpumask;
865 
866 #ifdef SMP
867 again:
868 #endif
869 	if (rtqbits) {
870 		pri = bsfl(rtqbits);
871 		q = &bsd4_rtqueues[pri];
872 		which = &bsd4_rtqueuebits;
873 		which2 = &rtqbits;
874 	} else if (tsqbits) {
875 		pri = bsfl(tsqbits);
876 		q = &bsd4_queues[pri];
877 		which = &bsd4_queuebits;
878 		which2 = &tsqbits;
879 	} else if (idqbits) {
880 		pri = bsfl(idqbits);
881 		q = &bsd4_idqueues[pri];
882 		which = &bsd4_idqueuebits;
883 		which2 = &idqbits;
884 	} else {
885 		return NULL;
886 	}
887 	lp = TAILQ_FIRST(q);
888 	KASSERT(lp, ("chooseproc: no lwp on busy queue"));
889 
890 #ifdef SMP
891 	while ((lp->lwp_cpumask & cpumask) == 0) {
892 		lp = TAILQ_NEXT(lp, lwp_procq);
893 		if (lp == NULL) {
894 			*which2 &= ~(1 << pri);
895 			goto again;
896 		}
897 	}
898 #endif
899 
900 	/*
901 	 * If the passed lwp <chklp> is reasonably close to the selected
902 	 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
903 	 *
904 	 * Note that we must error on the side of <chklp> to avoid bouncing
905 	 * between threads in the acquire code.
906 	 */
907 	if (chklp) {
908 		if (chklp->lwp_priority < lp->lwp_priority + PPQ)
909 			return(NULL);
910 	}
911 
912 #ifdef SMP
913 	/*
914 	 * If the chosen lwp does not reside on this cpu spend a few
915 	 * cycles looking for a better candidate at the same priority level.
916 	 * This is a fallback check, setrunqueue() tries to wakeup the
917 	 * correct cpu and is our front-line affinity.
918 	 */
919 	if (lp->lwp_thread->td_gd != mycpu &&
920 	    (chklp = TAILQ_NEXT(lp, lwp_procq)) != NULL
921 	) {
922 		if (chklp->lwp_thread->td_gd == mycpu) {
923 			++choose_affinity;
924 			lp = chklp;
925 		}
926 	}
927 #endif
928 
929 	TAILQ_REMOVE(q, lp, lwp_procq);
930 	--bsd4_runqcount;
931 	if (TAILQ_EMPTY(q))
932 		*which &= ~(1 << pri);
933 	KASSERT((lp->lwp_flag & LWP_ONRUNQ) != 0, ("not on runq6!"));
934 	lp->lwp_flag &= ~LWP_ONRUNQ;
935 	return lp;
936 }
937 
938 #ifdef SMP
939 
940 /*
941  * Called via an ipi message to reschedule on another cpu.  If no
942  * user thread is active on the target cpu we wake the scheduler
943  * helper thread up to help schedule one.
944  *
945  * MPSAFE
946  */
947 static
948 void
949 need_user_resched_remote(void *dummy)
950 {
951 	globaldata_t gd = mycpu;
952 	bsd4_pcpu_t  dd = &bsd4_pcpu[gd->gd_cpuid];
953 
954 	if (dd->uschedcp == NULL && (bsd4_rdyprocmask & gd->gd_cpumask)) {
955 		atomic_clear_int(&bsd4_rdyprocmask, gd->gd_cpumask);
956 		lwkt_schedule(&dd->helper_thread);
957 	} else {
958 		need_user_resched();
959 	}
960 }
961 
962 #endif
963 
964 /*
965  * bsd4_remrunqueue_locked() removes a given process from the run queue
966  * that it is on, clearing the queue busy bit if it becomes empty.
967  *
968  * Note that user process scheduler is different from the LWKT schedule.
969  * The user process scheduler only manages user processes but it uses LWKT
970  * underneath, and a user process operating in the kernel will often be
971  * 'released' from our management.
972  *
973  * MPSAFE - bsd4_spin must be held exclusively on call
974  */
975 static void
976 bsd4_remrunqueue_locked(struct lwp *lp)
977 {
978 	struct rq *q;
979 	u_int32_t *which;
980 	u_int8_t pri;
981 
982 	KKASSERT(lp->lwp_flag & LWP_ONRUNQ);
983 	lp->lwp_flag &= ~LWP_ONRUNQ;
984 	--bsd4_runqcount;
985 	KKASSERT(bsd4_runqcount >= 0);
986 
987 	pri = lp->lwp_rqindex;
988 	switch(lp->lwp_rqtype) {
989 	case RTP_PRIO_NORMAL:
990 		q = &bsd4_queues[pri];
991 		which = &bsd4_queuebits;
992 		break;
993 	case RTP_PRIO_REALTIME:
994 	case RTP_PRIO_FIFO:
995 		q = &bsd4_rtqueues[pri];
996 		which = &bsd4_rtqueuebits;
997 		break;
998 	case RTP_PRIO_IDLE:
999 		q = &bsd4_idqueues[pri];
1000 		which = &bsd4_idqueuebits;
1001 		break;
1002 	default:
1003 		panic("remrunqueue: invalid rtprio type");
1004 		/* NOT REACHED */
1005 	}
1006 	TAILQ_REMOVE(q, lp, lwp_procq);
1007 	if (TAILQ_EMPTY(q)) {
1008 		KASSERT((*which & (1 << pri)) != 0,
1009 			("remrunqueue: remove from empty queue"));
1010 		*which &= ~(1 << pri);
1011 	}
1012 }
1013 
1014 /*
1015  * bsd4_setrunqueue_locked()
1016  *
1017  * Add a process whos rqtype and rqindex had previously been calculated
1018  * onto the appropriate run queue.   Determine if the addition requires
1019  * a reschedule on a cpu and return the cpuid or -1.
1020  *
1021  * NOTE: Lower priorities are better priorities.
1022  *
1023  * MPSAFE - bsd4_spin must be held exclusively on call
1024  */
1025 static void
1026 bsd4_setrunqueue_locked(struct lwp *lp)
1027 {
1028 	struct rq *q;
1029 	u_int32_t *which;
1030 	int pri;
1031 
1032 	KKASSERT((lp->lwp_flag & LWP_ONRUNQ) == 0);
1033 	lp->lwp_flag |= LWP_ONRUNQ;
1034 	++bsd4_runqcount;
1035 
1036 	pri = lp->lwp_rqindex;
1037 
1038 	switch(lp->lwp_rqtype) {
1039 	case RTP_PRIO_NORMAL:
1040 		q = &bsd4_queues[pri];
1041 		which = &bsd4_queuebits;
1042 		break;
1043 	case RTP_PRIO_REALTIME:
1044 	case RTP_PRIO_FIFO:
1045 		q = &bsd4_rtqueues[pri];
1046 		which = &bsd4_rtqueuebits;
1047 		break;
1048 	case RTP_PRIO_IDLE:
1049 		q = &bsd4_idqueues[pri];
1050 		which = &bsd4_idqueuebits;
1051 		break;
1052 	default:
1053 		panic("remrunqueue: invalid rtprio type");
1054 		/* NOT REACHED */
1055 	}
1056 
1057 	/*
1058 	 * Add to the correct queue and set the appropriate bit.  If no
1059 	 * lower priority (i.e. better) processes are in the queue then
1060 	 * we want a reschedule, calculate the best cpu for the job.
1061 	 *
1062 	 * Always run reschedules on the LWPs original cpu.
1063 	 */
1064 	TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1065 	*which |= 1 << pri;
1066 }
1067 
1068 #ifdef SMP
1069 
1070 /*
1071  * For SMP systems a user scheduler helper thread is created for each
1072  * cpu and is used to allow one cpu to wakeup another for the purposes of
1073  * scheduling userland threads from setrunqueue().  UP systems do not
1074  * need the helper since there is only one cpu.  We can't use the idle
1075  * thread for this because we need to hold the MP lock.  Additionally,
1076  * doing things this way allows us to HLT idle cpus on MP systems.
1077  *
1078  * MPSAFE
1079  */
1080 static void
1081 sched_thread(void *dummy)
1082 {
1083     globaldata_t gd;
1084     bsd4_pcpu_t  dd;
1085     struct lwp *nlp;
1086     cpumask_t cpumask;
1087     int cpuid;
1088 #if 0
1089     cpumask_t tmpmask;
1090     int tmpid;
1091 #endif
1092 
1093     gd = mycpu;
1094     cpuid = gd->gd_cpuid;	/* doesn't change */
1095     cpumask = gd->gd_cpumask;	/* doesn't change */
1096     dd = &bsd4_pcpu[cpuid];
1097 
1098     /*
1099      * The scheduler thread does not need to hold the MP lock.  Since we
1100      * are woken up only when no user processes are scheduled on a cpu, we
1101      * can run at an ultra low priority.
1102      */
1103     rel_mplock();
1104     lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1105 
1106     for (;;) {
1107 	/*
1108 	 * We use the LWKT deschedule-interlock trick to avoid racing
1109 	 * bsd4_rdyprocmask.  This means we cannot block through to the
1110 	 * manual lwkt_switch() call we make below.
1111 	 */
1112 	crit_enter_gd(gd);
1113 	lwkt_deschedule_self(gd->gd_curthread);
1114 	spin_lock_wr(&bsd4_spin);
1115 	atomic_set_int(&bsd4_rdyprocmask, cpumask);
1116 
1117 	clear_user_resched();	/* This satisfied the reschedule request */
1118 	dd->rrcount = 0;	/* Reset the round-robin counter */
1119 
1120 	if ((bsd4_curprocmask & cpumask) == 0) {
1121 		/*
1122 		 * No thread is currently scheduled.
1123 		 */
1124 		KKASSERT(dd->uschedcp == NULL);
1125 		if ((nlp = chooseproc_locked(NULL)) != NULL) {
1126 			atomic_set_int(&bsd4_curprocmask, cpumask);
1127 			dd->upri = nlp->lwp_priority;
1128 			dd->uschedcp = nlp;
1129 			spin_unlock_wr(&bsd4_spin);
1130 			lwkt_acquire(nlp->lwp_thread);
1131 			lwkt_schedule(nlp->lwp_thread);
1132 		} else {
1133 			spin_unlock_wr(&bsd4_spin);
1134 		}
1135 #if 0
1136 	/*
1137 	 * Disabled for now, this can create an infinite loop.
1138 	 */
1139 	} else if (bsd4_runqcount) {
1140 		/*
1141 		 * Someone scheduled us but raced.  In order to not lose
1142 		 * track of the fact that there may be a LWP ready to go,
1143 		 * forward the request to another cpu if available.
1144 		 *
1145 		 * Rotate through cpus starting with cpuid + 1.  Since cpuid
1146 		 * is already masked out by gd_other_cpus, just use ~cpumask.
1147 		 */
1148 		tmpmask = bsd4_rdyprocmask & mycpu->gd_other_cpus &
1149 			  ~bsd4_curprocmask;
1150 		if (tmpmask) {
1151 			if (tmpmask & ~(cpumask - 1))
1152 				tmpid = bsfl(tmpmask & ~(cpumask - 1));
1153 			else
1154 				tmpid = bsfl(tmpmask);
1155 			bsd4_scancpu = tmpid;
1156 			atomic_clear_int(&bsd4_rdyprocmask, 1 << tmpid);
1157 			spin_unlock_wr(&bsd4_spin);
1158 			lwkt_schedule(&bsd4_pcpu[tmpid].helper_thread);
1159 		} else {
1160 			spin_unlock_wr(&bsd4_spin);
1161 		}
1162 #endif
1163 	} else {
1164 		/*
1165 		 * The runq is empty.
1166 		 */
1167 		spin_unlock_wr(&bsd4_spin);
1168 	}
1169 	crit_exit_gd(gd);
1170 	lwkt_switch();
1171     }
1172 }
1173 
1174 /*
1175  * Setup our scheduler helpers.  Note that curprocmask bit 0 has already
1176  * been cleared by rqinit() and we should not mess with it further.
1177  */
1178 static void
1179 sched_thread_cpu_init(void)
1180 {
1181     int i;
1182 
1183     if (bootverbose)
1184 	kprintf("start scheduler helpers on cpus:");
1185 
1186     for (i = 0; i < ncpus; ++i) {
1187 	bsd4_pcpu_t dd = &bsd4_pcpu[i];
1188 	cpumask_t mask = 1 << i;
1189 
1190 	if ((mask & smp_active_mask) == 0)
1191 	    continue;
1192 
1193 	if (bootverbose)
1194 	    kprintf(" %d", i);
1195 
1196 	lwkt_create(sched_thread, NULL, NULL, &dd->helper_thread,
1197 		    TDF_STOPREQ, i, "usched %d", i);
1198 
1199 	/*
1200 	 * Allow user scheduling on the target cpu.  cpu #0 has already
1201 	 * been enabled in rqinit().
1202 	 */
1203 	if (i)
1204 	    atomic_clear_int(&bsd4_curprocmask, mask);
1205 	atomic_set_int(&bsd4_rdyprocmask, mask);
1206 	dd->upri = PRIBASE_NULL;
1207     }
1208     if (bootverbose)
1209 	kprintf("\n");
1210 }
1211 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1212 	sched_thread_cpu_init, NULL)
1213 
1214 #endif
1215 
1216