xref: /dragonfly/sys/kern/usched_dfly.c (revision a3127495)
1 /*
2  * Copyright (c) 2012-2017 The DragonFly Project.  All rights reserved.
3  * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org>.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Matthew Dillon <dillon@backplane.com>,
7  * by Mihai Carabas <mihai.carabas@gmail.com>
8  * and many others.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in
18  *    the documentation and/or other materials provided with the
19  *    distribution.
20  * 3. Neither the name of The DragonFly Project nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific, prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
28  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  */
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/queue.h>
42 #include <sys/proc.h>
43 #include <sys/rtprio.h>
44 #include <sys/uio.h>
45 #include <sys/sysctl.h>
46 #include <sys/resourcevar.h>
47 #include <sys/spinlock.h>
48 #include <sys/cpu_topology.h>
49 #include <sys/thread2.h>
50 #include <sys/spinlock2.h>
51 
52 #include <sys/ktr.h>
53 
54 #include <machine/cpu.h>
55 #include <machine/smp.h>
56 
57 /*
58  * Priorities.  Note that with 32 run queues per scheduler each queue
59  * represents four priority levels.
60  */
61 
62 int dfly_rebalanced;
63 
64 #define MAXPRI			128
65 #define PRIMASK			(MAXPRI - 1)
66 #define PRIBASE_REALTIME	0
67 #define PRIBASE_NORMAL		MAXPRI
68 #define PRIBASE_IDLE		(MAXPRI * 2)
69 #define PRIBASE_THREAD		(MAXPRI * 3)
70 #define PRIBASE_NULL		(MAXPRI * 4)
71 
72 #define NQS	32			/* 32 run queues. */
73 #define PPQ	(MAXPRI / NQS)		/* priorities per queue */
74 #define PPQMASK	(PPQ - 1)
75 
76 /*
77  * NICE_QS	- maximum queues nice can shift the process
78  * EST_QS	- maximum queues estcpu can shift the process
79  *
80  * ESTCPUPPQ	- number of estcpu units per priority queue
81  * ESTCPUMAX	- number of estcpu units
82  *
83  * Remember that NICE runs over the whole -20 to +20 range.
84  */
85 #define NICE_QS		24	/* -20 to +20 shift in whole queues */
86 #define EST_QS		12	/* 0-MAX shift in whole queues */
87 #define ESTCPUPPQ	512
88 #define ESTCPUMAX	(ESTCPUPPQ * EST_QS)
89 #define PRIO_RANGE	(PRIO_MAX - PRIO_MIN + 1)
90 
91 #define ESTCPULIM(v)	min((v), ESTCPUMAX)
92 
93 TAILQ_HEAD(rq, lwp);
94 
95 #define lwp_priority	lwp_usdata.dfly.priority
96 #define lwp_forked	lwp_usdata.dfly.forked
97 #define lwp_rqindex	lwp_usdata.dfly.rqindex
98 #define lwp_estcpu	lwp_usdata.dfly.estcpu
99 #define lwp_estfast	lwp_usdata.dfly.estfast
100 #define lwp_uload	lwp_usdata.dfly.uload
101 #define lwp_rqtype	lwp_usdata.dfly.rqtype
102 #define lwp_qcpu	lwp_usdata.dfly.qcpu
103 #define lwp_rrcount	lwp_usdata.dfly.rrcount
104 
105 /*
106  * DFly scheduler pcpu structure.  Note that the pcpu uload field must
107  * be 64-bits to avoid overflowing in the situation where more than 32768
108  * processes are on a single cpu's queue.  Since high-end systems can
109  * easily run 900,000+ processes, we have to deal with it.
110  */
111 struct usched_dfly_pcpu {
112 	struct spinlock spin;
113 	struct thread	*helper_thread;
114 	u_short		scancpu;
115 	short		upri;
116 	long		uload;		/* 64-bits to avoid overflow (1) */
117 	int		ucount;
118 	int		unused01;
119 	struct lwp	*uschedcp;
120 	struct rq	queues[NQS];
121 	struct rq	rtqueues[NQS];
122 	struct rq	idqueues[NQS];
123 	u_int32_t	queuebits;
124 	u_int32_t	rtqueuebits;
125 	u_int32_t	idqueuebits;
126 	int		runqcount;
127 	int		cpuid;
128 	cpumask_t	cpumask;
129 	cpu_node_t	*cpunode;
130 };
131 
132 typedef struct usched_dfly_pcpu	*dfly_pcpu_t;
133 
134 static void dfly_acquire_curproc(struct lwp *lp);
135 static void dfly_release_curproc(struct lwp *lp);
136 static void dfly_select_curproc(globaldata_t gd);
137 static void dfly_setrunqueue(struct lwp *lp);
138 static void dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp);
139 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period,
140 				sysclock_t cpstamp);
141 static void dfly_recalculate_estcpu(struct lwp *lp);
142 static void dfly_resetpriority(struct lwp *lp);
143 static void dfly_forking(struct lwp *plp, struct lwp *lp);
144 static void dfly_exiting(struct lwp *lp, struct proc *);
145 static void dfly_uload_update(struct lwp *lp);
146 static void dfly_yield(struct lwp *lp);
147 static void dfly_changeqcpu_locked(struct lwp *lp,
148 				dfly_pcpu_t dd, dfly_pcpu_t rdd);
149 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
150 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
151 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
152 static void dfly_need_user_resched_remote(void *dummy);
153 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
154 					  struct lwp *chklp, int worst);
155 static void dfly_remrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
156 static void dfly_setrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
157 static void dfly_changedcpu(struct lwp *lp);
158 
159 struct usched usched_dfly = {
160 	{ NULL },
161 	"dfly", "Original DragonFly Scheduler",
162 	NULL,			/* default registration */
163 	NULL,			/* default deregistration */
164 	dfly_acquire_curproc,
165 	dfly_release_curproc,
166 	dfly_setrunqueue,
167 	dfly_schedulerclock,
168 	dfly_recalculate_estcpu,
169 	dfly_resetpriority,
170 	dfly_forking,
171 	dfly_exiting,
172 	dfly_uload_update,
173 	NULL,			/* setcpumask not supported */
174 	dfly_yield,
175 	dfly_changedcpu
176 };
177 
178 /*
179  * We have NQS (32) run queues per scheduling class.  For the normal
180  * class, there are 128 priorities scaled onto these 32 queues.  New
181  * processes are added to the last entry in each queue, and processes
182  * are selected for running by taking them from the head and maintaining
183  * a simple FIFO arrangement.  Realtime and Idle priority processes have
184  * and explicit 0-31 priority which maps directly onto their class queue
185  * index.  When a queue has something in it, the corresponding bit is
186  * set in the queuebits variable, allowing a single read to determine
187  * the state of all 32 queues and then a ffs() to find the first busy
188  * queue.
189  */
190 					/* currently running a user process */
191 static cpumask_t dfly_curprocmask = CPUMASK_INITIALIZER_ALLONES;
192 static cpumask_t dfly_rdyprocmask;	/* ready to accept a user process */
193 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
194 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
195 static struct sysctl_oid *usched_dfly_sysctl_tree;
196 
197 /* Debug info exposed through debug.* sysctl */
198 
199 static int usched_dfly_debug = -1;
200 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
201 	   &usched_dfly_debug, 0,
202 	   "Print debug information for this pid");
203 
204 static int usched_dfly_pid_debug = -1;
205 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
206 	   &usched_dfly_pid_debug, 0,
207 	   "Print KTR debug information for this pid");
208 
209 static int usched_dfly_chooser = 0;
210 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
211 	   &usched_dfly_chooser, 0,
212 	   "Print KTR debug information for this pid");
213 
214 /*
215  * WARNING!
216  *
217  * The fork bias can have a large effect on the system in the face of a
218  * make -j N or other high-forking applications.
219  *
220  * Larger values are much less invasive vs other things that
221  * might be running in the system, but can cause exec chains
222  * such as those typically generated by make to have higher
223  * latencies in the face of modest load.
224  *
225  * Lower values are more invasive but have reduced latencies
226  * for such exec chains.
227  *
228  *	make -j 10 buildkernel example, build times:
229  *
230  *	     +0	3:04
231  *	     +1 3:14	-5.2%	<-- default
232  *	     +2 3:22	-8.9%
233  *
234  * This issue occurs due to the way the scheduler affinity heuristics work.
235  * There is no way to really 'fix' the affinity heuristics because when it
236  * comes right down to it trying to instantly schedule a process on an
237  * available cpu (even if it will become unavailable a microsecond later)
238  * tends to cause processes to shift around between cpus and sockets too much
239  * and breaks the affinity.
240  *
241  * NOTE: Heavily concurrent builds typically have enough things on the pan
242  *	 that they remain time-efficient even with a higher bias.
243  */
244 static int usched_dfly_forkbias = 1;
245 SYSCTL_INT(_debug, OID_AUTO, dfly_forkbias, CTLFLAG_RW,
246 	   &usched_dfly_forkbias, 0,
247 	   "Fork bias for estcpu in whole queues");
248 
249 /*
250  * Tunning usched_dfly - configurable through kern.usched_dfly.
251  *
252  * weight1 - Tries to keep threads on their current cpu.  If you
253  *	     make this value too large the scheduler will not be
254  *	     able to load-balance large loads.
255  *
256  * weight2 - If non-zero, detects thread pairs undergoing synchronous
257  *	     communications and tries to move them closer together.
258  *	     Behavior is adjusted by bit 4 of features (0x10).
259  *
260  *	     WARNING!  Weight2 is a ridiculously sensitive parameter,
261  *	     a small value is recommended.
262  *
263  * weight3 - Weighting based on the number of recently runnable threads
264  *	     on the userland scheduling queue (ignoring their loads).
265  *	     A nominal value here prevents high-priority (low-load)
266  *	     threads from accumulating on one cpu core when other
267  *	     cores are available.
268  *
269  *	     This value should be left fairly small relative to weight1
270  *	     and weight4.
271  *
272  * weight4 - Weighting based on other cpu queues being available
273  *	     or running processes with higher lwp_priority's.
274  *
275  *	     This allows a thread to migrate to another nearby cpu if it
276  *	     is unable to run on the current cpu based on the other cpu
277  *	     being idle or running a lower priority (higher lwp_priority)
278  *	     thread.  This value should be large enough to override weight1
279  *
280  * features - These flags can be set or cleared to enable or disable various
281  *	      features.
282  *
283  *	      0x01	Enable idle-cpu pulling			(default)
284  *	      0x02	Enable proactive pushing		(default)
285  *	      0x04	Enable rebalancing rover		(default)
286  *	      0x08	Enable more proactive pushing		(default)
287  *	      0x10	(flip weight2 limit on same cpu)	(default)
288  *	      0x20	choose best cpu for forked process
289  *	      0x40	choose current cpu for forked process
290  *	      0x80	choose random cpu for forked process	(default)
291  */
292 static int usched_dfly_smt = 0;
293 static int usched_dfly_cache_coherent = 0;
294 static int usched_dfly_weight1 = 200;	/* keep thread on current cpu */
295 static int usched_dfly_weight2 = 180;	/* synchronous peer's current cpu */
296 static int usched_dfly_weight3 = 40;	/* number of threads on queue */
297 static int usched_dfly_weight4 = 160;	/* availability of idle cores */
298 static int usched_dfly_features = 0x8F;	/* allow pulls */
299 static int usched_dfly_fast_resched = 0;/* delta priority / resched */
300 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */
301 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
302 static int usched_dfly_decay = 8;
303 
304 /* KTR debug printings */
305 
306 KTR_INFO_MASTER(usched);
307 
308 #if !defined(KTR_USCHED_DFLY)
309 #define	KTR_USCHED_DFLY	KTR_ALL
310 #endif
311 
312 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
313     "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
314     pid_t pid, int old_cpuid, int curr);
315 
316 /*
317  * This function is called when the kernel intends to return to userland.
318  * It is responsible for making the thread the current designated userland
319  * thread for this cpu, blocking if necessary.
320  *
321  * The kernel will not depress our LWKT priority until after we return,
322  * in case we have to shove over to another cpu.
323  *
324  * We must determine our thread's disposition before we switch away.  This
325  * is very sensitive code.
326  *
327  * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
328  * TO ANOTHER CPU!  Because most of the kernel assumes that no migration will
329  * occur, this function is called only under very controlled circumstances.
330  */
331 static void
332 dfly_acquire_curproc(struct lwp *lp)
333 {
334 	globaldata_t gd;
335 	dfly_pcpu_t dd;
336 	dfly_pcpu_t rdd;
337 	thread_t td;
338 	int force_resched;
339 
340 	/*
341 	 * Make sure we aren't sitting on a tsleep queue.
342 	 */
343 	td = lp->lwp_thread;
344 	crit_enter_quick(td);
345 	if (td->td_flags & TDF_TSLEEPQ)
346 		tsleep_remove(td);
347 	dfly_recalculate_estcpu(lp);
348 
349 	gd = mycpu;
350 	dd = &dfly_pcpu[gd->gd_cpuid];
351 
352 	/*
353 	 * Process any pending interrupts/ipi's, then handle reschedule
354 	 * requests.  dfly_release_curproc() will try to assign a new
355 	 * uschedcp that isn't us and otherwise NULL it out.
356 	 */
357 	force_resched = 0;
358 	if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
359 	    lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
360 		force_resched = 1;
361 	}
362 
363 	if (user_resched_wanted()) {
364 		if (dd->uschedcp == lp)
365 			force_resched = 1;
366 		clear_user_resched();
367 		dfly_release_curproc(lp);
368 	}
369 
370 	/*
371 	 * Loop until we are the current user thread.
372 	 *
373 	 * NOTE: dd spinlock not held at top of loop.
374 	 */
375 	if (dd->uschedcp == lp)
376 		lwkt_yield_quick();
377 
378 	while (dd->uschedcp != lp) {
379 		lwkt_yield_quick();
380 
381 		spin_lock(&dd->spin);
382 
383 		/* This lwp is an outcast; force reschedule. */
384 		if (__predict_false(
385 		    CPUMASK_TESTBIT(lp->lwp_cpumask, gd->gd_cpuid) == 0) &&
386 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
387 			dfly_changeqcpu_locked(lp, dd, rdd);
388 			spin_unlock(&dd->spin);
389 			lwkt_deschedule(lp->lwp_thread);
390 			dfly_setrunqueue_dd(rdd, lp);
391 			lwkt_switch();
392 			gd = mycpu;
393 			dd = &dfly_pcpu[gd->gd_cpuid];
394 			continue;
395 		}
396 
397 		if (force_resched &&
398 		   (usched_dfly_features & 0x08) &&
399 		   (rdd = dfly_choose_best_queue(lp)) != dd) {
400 			/*
401 			 * We are not or are no longer the current lwp and a
402 			 * forced reschedule was requested.  Figure out the
403 			 * best cpu to run on (our current cpu will be given
404 			 * significant weight).
405 			 *
406 			 * (if a reschedule was not requested we want to
407 			 *  move this step after the uschedcp tests).
408 			 */
409 			dfly_changeqcpu_locked(lp, dd, rdd);
410 			spin_unlock(&dd->spin);
411 			lwkt_deschedule(lp->lwp_thread);
412 			dfly_setrunqueue_dd(rdd, lp);
413 			lwkt_switch();
414 			gd = mycpu;
415 			dd = &dfly_pcpu[gd->gd_cpuid];
416 			continue;
417 		}
418 
419 		/*
420 		 * Either no reschedule was requested or the best queue was
421 		 * dd, and no current process has been selected.  We can
422 		 * trivially become the current lwp on the current cpu.
423 		 */
424 		if (dd->uschedcp == NULL) {
425 			atomic_clear_int(&lp->lwp_thread->td_mpflags,
426 					 TDF_MP_DIDYIELD);
427 			ATOMIC_CPUMASK_ORBIT(dfly_curprocmask, gd->gd_cpuid);
428 			dd->uschedcp = lp;
429 			dd->upri = lp->lwp_priority;
430 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
431 			spin_unlock(&dd->spin);
432 			break;
433 		}
434 
435 		/*
436 		 * Put us back on the same run queue unconditionally.
437 		 *
438 		 * Set rrinterval to force placement at end of queue.
439 		 * Select the worst queue to ensure we round-robin,
440 		 * but do not change estcpu.
441 		 */
442 		if (lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) {
443 			u_int32_t tsqbits;
444 
445 			switch(lp->lwp_rqtype) {
446 			case RTP_PRIO_NORMAL:
447 				tsqbits = dd->queuebits;
448 				spin_unlock(&dd->spin);
449 
450 				lp->lwp_rrcount = usched_dfly_rrinterval;
451 				if (tsqbits)
452 					lp->lwp_rqindex = bsrl(tsqbits);
453 				break;
454 			default:
455 				spin_unlock(&dd->spin);
456 				break;
457 			}
458 			lwkt_deschedule(lp->lwp_thread);
459 			dfly_setrunqueue_dd(dd, lp);
460 			atomic_clear_int(&lp->lwp_thread->td_mpflags,
461 					 TDF_MP_DIDYIELD);
462 			lwkt_switch();
463 			gd = mycpu;
464 			dd = &dfly_pcpu[gd->gd_cpuid];
465 			continue;
466 		}
467 
468 		/*
469 		 * Can we steal the current designated user thread?
470 		 *
471 		 * If we do the other thread will stall when it tries to
472 		 * return to userland, possibly rescheduling elsewhere.
473 		 *
474 		 * It is important to do a masked test to avoid the edge
475 		 * case where two near-equal-priority threads are constantly
476 		 * interrupting each other.
477 		 *
478 		 * In the exact match case another thread has already gained
479 		 * uschedcp and lowered its priority, if we steal it the
480 		 * other thread will stay stuck on the LWKT runq and not
481 		 * push to another cpu.  So don't steal on equal-priority even
482 		 * though it might appear to be more beneficial due to not
483 		 * having to switch back to the other thread's context.
484 		 *
485 		 * usched_dfly_fast_resched requires that two threads be
486 		 * significantly far apart in priority in order to interrupt.
487 		 *
488 		 * If better but not sufficiently far apart, the current
489 		 * uschedcp will be interrupted at the next scheduler clock.
490 		 */
491 		if (dd->uschedcp &&
492 		   (dd->upri & ~PPQMASK) >
493 		   (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
494 			dd->uschedcp = lp;
495 			dd->upri = lp->lwp_priority;
496 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
497 			spin_unlock(&dd->spin);
498 			break;
499 		}
500 		/*
501 		 * We are not the current lwp, figure out the best cpu
502 		 * to run on (our current cpu will be given significant
503 		 * weight).  Loop on cpu change.
504 		 */
505 		if ((usched_dfly_features & 0x02) &&
506 		    force_resched == 0 &&
507 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
508 			dfly_changeqcpu_locked(lp, dd, rdd);
509 			spin_unlock(&dd->spin);
510 			lwkt_deschedule(lp->lwp_thread);
511 			dfly_setrunqueue_dd(rdd, lp);
512 			lwkt_switch();
513 			gd = mycpu;
514 			dd = &dfly_pcpu[gd->gd_cpuid];
515 			continue;
516 		}
517 
518 		/*
519 		 * We cannot become the current lwp, place the lp on the
520 		 * run-queue of this or another cpu and deschedule ourselves.
521 		 *
522 		 * When we are reactivated we will have another chance.
523 		 *
524 		 * Reload after a switch or setrunqueue/switch possibly
525 		 * moved us to another cpu.
526 		 */
527 		spin_unlock(&dd->spin);
528 		lwkt_deschedule(lp->lwp_thread);
529 		dfly_setrunqueue_dd(dd, lp);
530 		lwkt_switch();
531 		gd = mycpu;
532 		dd = &dfly_pcpu[gd->gd_cpuid];
533 	}
534 
535 	/*
536 	 * Make sure upri is synchronized, then yield to LWKT threads as
537 	 * needed before returning.  This could result in another reschedule.
538 	 * XXX
539 	 */
540 	crit_exit_quick(td);
541 
542 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
543 }
544 
545 /*
546  * DFLY_RELEASE_CURPROC
547  *
548  * This routine detaches the current thread from the userland scheduler,
549  * usually because the thread needs to run or block in the kernel (at
550  * kernel priority) for a while.
551  *
552  * This routine is also responsible for selecting a new thread to
553  * make the current thread.
554  *
555  * NOTE: This implementation differs from the dummy example in that
556  * dfly_select_curproc() is able to select the current process, whereas
557  * dummy_select_curproc() is not able to select the current process.
558  * This means we have to NULL out uschedcp.
559  *
560  * Additionally, note that we may already be on a run queue if releasing
561  * via the lwkt_switch() in dfly_setrunqueue().
562  */
563 static void
564 dfly_release_curproc(struct lwp *lp)
565 {
566 	globaldata_t gd = mycpu;
567 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
568 
569 	/*
570 	 * Make sure td_wakefromcpu is defaulted.  This will be overwritten
571 	 * by wakeup().
572 	 */
573 	if (dd->uschedcp == lp) {
574 		KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
575 		spin_lock(&dd->spin);
576 		if (dd->uschedcp == lp) {
577 			dd->uschedcp = NULL;	/* don't let lp be selected */
578 			dd->upri = PRIBASE_NULL;
579 			ATOMIC_CPUMASK_NANDBIT(dfly_curprocmask, gd->gd_cpuid);
580 			spin_unlock(&dd->spin);
581 			dfly_select_curproc(gd);
582 		} else {
583 			spin_unlock(&dd->spin);
584 		}
585 	}
586 }
587 
588 /*
589  * DFLY_SELECT_CURPROC
590  *
591  * Select a new current process for this cpu and clear any pending user
592  * reschedule request.  The cpu currently has no current process.
593  *
594  * This routine is also responsible for equal-priority round-robining,
595  * typically triggered from dfly_schedulerclock().  In our dummy example
596  * all the 'user' threads are LWKT scheduled all at once and we just
597  * call lwkt_switch().
598  *
599  * The calling process is not on the queue and cannot be selected.
600  */
601 static
602 void
603 dfly_select_curproc(globaldata_t gd)
604 {
605 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
606 	struct lwp *nlp;
607 	int cpuid = gd->gd_cpuid;
608 
609 	crit_enter_gd(gd);
610 
611 	spin_lock(&dd->spin);
612 	nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
613 
614 	if (nlp) {
615 		ATOMIC_CPUMASK_ORBIT(dfly_curprocmask, cpuid);
616 		dd->upri = nlp->lwp_priority;
617 		dd->uschedcp = nlp;
618 #if 0
619 		dd->rrcount = 0;		/* reset round robin */
620 #endif
621 		spin_unlock(&dd->spin);
622 		lwkt_acquire(nlp->lwp_thread);
623 		lwkt_schedule(nlp->lwp_thread);
624 	} else {
625 		spin_unlock(&dd->spin);
626 	}
627 	crit_exit_gd(gd);
628 }
629 
630 /*
631  * Place the specified lwp on the user scheduler's run queue.  This routine
632  * must be called with the thread descheduled.  The lwp must be runnable.
633  * It must not be possible for anyone else to explicitly schedule this thread.
634  *
635  * The thread may be the current thread as a special case.
636  */
637 static void
638 dfly_setrunqueue(struct lwp *lp)
639 {
640 	dfly_pcpu_t dd;
641 	dfly_pcpu_t rdd;
642 
643 	/*
644 	 * First validate the process LWKT state.
645 	 */
646 	KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
647 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
648 	    ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
649 	     lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
650 	KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
651 
652 	/*
653 	 * NOTE: dd/rdd do not necessarily represent the current cpu.
654 	 *	 Instead they may represent the cpu the thread was last
655 	 *	 scheduled on or inherited by its parent.
656 	 */
657 	dd = &dfly_pcpu[lp->lwp_qcpu];
658 	rdd = dd;
659 
660 	/*
661 	 * This process is not supposed to be scheduled anywhere or assigned
662 	 * as the current process anywhere.  Assert the condition.
663 	 */
664 	KKASSERT(rdd->uschedcp != lp);
665 
666 	/*
667 	 * Ok, we have to setrunqueue some target cpu and request a reschedule
668 	 * if necessary.
669 	 *
670 	 * We have to choose the best target cpu.  It might not be the current
671 	 * target even if the current cpu has no running user thread (for
672 	 * example, because the current cpu might be a hyperthread and its
673 	 * sibling has a thread assigned).
674 	 *
675 	 * If we just forked it is most optimal to run the child on the same
676 	 * cpu just in case the parent decides to wait for it (thus getting
677 	 * off that cpu).  As long as there is nothing else runnable on the
678 	 * cpu, that is.  If we did this unconditionally a parent forking
679 	 * multiple children before waiting (e.g. make -j N) leaves other
680 	 * cpus idle that could be working.
681 	 */
682 	if (lp->lwp_forked) {
683 		lp->lwp_forked = 0;
684 		if (usched_dfly_features & 0x20)
685 			rdd = dfly_choose_best_queue(lp);
686 		else if (usched_dfly_features & 0x40)
687 			rdd = &dfly_pcpu[lp->lwp_qcpu];
688 		else if (usched_dfly_features & 0x80)
689 			rdd = dfly_choose_queue_simple(rdd, lp);
690 		else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
691 			rdd = dfly_choose_best_queue(lp);
692 		else
693 			rdd = &dfly_pcpu[lp->lwp_qcpu];
694 	} else {
695 		rdd = dfly_choose_best_queue(lp);
696 		/* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
697 	}
698 	if (lp->lwp_qcpu != rdd->cpuid) {
699 		spin_lock(&dd->spin);
700 		dfly_changeqcpu_locked(lp, dd, rdd);
701 		spin_unlock(&dd->spin);
702 	}
703 	dfly_setrunqueue_dd(rdd, lp);
704 }
705 
706 /*
707  * Change qcpu to rdd->cpuid.  The dd the lp is CURRENTLY on must be
708  * spin-locked on-call.  rdd does not have to be.
709  */
710 static void
711 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
712 {
713 	if (lp->lwp_qcpu != rdd->cpuid) {
714 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
715 			atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
716 			atomic_add_long(&dd->uload, -lp->lwp_uload);
717 			atomic_add_int(&dd->ucount, -1);
718 		}
719 		lp->lwp_qcpu = rdd->cpuid;
720 	}
721 }
722 
723 /*
724  * Place lp on rdd's runqueue.  Nothing is locked on call.  This function
725  * also performs all necessary ancillary notification actions.
726  */
727 static void
728 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
729 {
730 	globaldata_t rgd;
731 
732 	/*
733 	 * We might be moving the lp to another cpu's run queue, and once
734 	 * on the runqueue (even if it is our cpu's), another cpu can rip
735 	 * it away from us.
736 	 *
737 	 * TDF_MIGRATING might already be set if this is part of a
738 	 * remrunqueue+setrunqueue sequence.
739 	 */
740 	if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
741 		lwkt_giveaway(lp->lwp_thread);
742 
743 	rgd = globaldata_find(rdd->cpuid);
744 
745 	/*
746 	 * We lose control of the lp the moment we release the spinlock
747 	 * after having placed it on the queue.  i.e. another cpu could pick
748 	 * it up, or it could exit, or its priority could be further
749 	 * adjusted, or something like that.
750 	 *
751 	 * WARNING! rdd can point to a foreign cpu!
752 	 */
753 	spin_lock(&rdd->spin);
754 	dfly_setrunqueue_locked(rdd, lp);
755 
756 	/*
757 	 * Potentially interrupt the currently-running thread
758 	 */
759 	if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
760 		/*
761 		 * Currently running thread is better or same, do not
762 		 * interrupt.
763 		 */
764 		spin_unlock(&rdd->spin);
765 	} else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
766 		   usched_dfly_fast_resched) {
767 		/*
768 		 * Currently running thread is not better, but not so bad
769 		 * that we need to interrupt it.  Let it run for one more
770 		 * scheduler tick.
771 		 */
772 		if (rdd->uschedcp &&
773 		    rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
774 			rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
775 		}
776 		spin_unlock(&rdd->spin);
777 	} else if (rgd == mycpu) {
778 		/*
779 		 * We should interrupt the currently running thread, which
780 		 * is on the current cpu.  However, if DIDYIELD is set we
781 		 * round-robin unconditionally and do not interrupt it.
782 		 */
783 		spin_unlock(&rdd->spin);
784 		if (rdd->uschedcp == NULL)
785 			wakeup_mycpu(rdd->helper_thread); /* XXX */
786 		if ((lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) == 0)
787 			need_user_resched();
788 	} else {
789 		/*
790 		 * We should interrupt the currently running thread, which
791 		 * is on a different cpu.
792 		 */
793 		spin_unlock(&rdd->spin);
794 		lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
795 	}
796 }
797 
798 /*
799  * This routine is called from a systimer IPI.  It MUST be MP-safe and
800  * the BGL IS NOT HELD ON ENTRY.  This routine is called at ESTCPUFREQ on
801  * each cpu.
802  */
803 static
804 void
805 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
806 {
807 	globaldata_t gd = mycpu;
808 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
809 
810 	/*
811 	 * Spinlocks also hold a critical section so there should not be
812 	 * any active.
813 	 */
814 	KKASSERT(gd->gd_spinlocks == 0 || dumping);
815 
816 	/*
817 	 * If lp is NULL we might be contended and lwkt_switch() may have
818 	 * cycled into the idle thread.  Apply the tick to the current
819 	 * process on this cpu if it is contended.
820 	 */
821 	if (gd->gd_curthread == &gd->gd_idlethread) {
822 		lp = dd->uschedcp;
823 		if (lp && (lp->lwp_thread == NULL ||
824 			   lp->lwp_thread->td_contended == 0)) {
825 			lp = NULL;
826 		}
827 	}
828 
829 	/*
830 	 * Dock thread for tick
831 	 */
832 	if (lp) {
833 		/*
834 		 * Do we need to round-robin?  We round-robin 10 times a
835 		 * second.  This should only occur for cpu-bound batch
836 		 * processes.
837 		 */
838 		if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
839 			lp->lwp_thread->td_wakefromcpu = -1;
840 			need_user_resched();
841 		}
842 
843 		/*
844 		 * Adjust estcpu upward using a real time equivalent
845 		 * calculation, and recalculate lp's priority.  Estcpu
846 		 * is increased such that it will cap-out over a period
847 		 * of one second.
848 		 */
849 		lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu +
850 					   ESTCPUMAX / ESTCPUFREQ + 1);
851 		dfly_resetpriority(lp);
852 	}
853 
854 	/*
855 	 * Rebalance two cpus every 8 ticks, pulling the worst thread
856 	 * from the worst cpu's queue into a rotating cpu number.
857 	 *
858 	 * This mechanic is needed because the push algorithms can
859 	 * steady-state in an non-optimal configuration.  We need to mix it
860 	 * up a little, even if it means breaking up a paired thread, so
861 	 * the push algorithms can rebalance the degenerate conditions.
862 	 * This portion of the algorithm exists to ensure stability at the
863 	 * selected weightings.
864 	 *
865 	 * Because we might be breaking up optimal conditions we do not want
866 	 * to execute this too quickly, hence we only rebalance approximately
867 	 * ~7-8 times per second.  The push's, on the otherhand, are capable
868 	 * moving threads to other cpus at a much higher rate.
869 	 *
870 	 * We choose the most heavily loaded thread from the worst queue
871 	 * in order to ensure that multiple heavy-weight threads on the same
872 	 * queue get broken up, and also because these threads are the most
873 	 * likely to be able to remain in place.  Hopefully then any pairings,
874 	 * if applicable, migrate to where these threads are.
875 	 */
876 	if ((usched_dfly_features & 0x04) &&
877 	    ((u_int)sched_ticks & 7) == 0 &&
878 	    (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
879 		/*
880 		 * Our cpu is up.
881 		 */
882 		struct lwp *nlp;
883 		dfly_pcpu_t rdd;
884 
885 		rdd = dfly_choose_worst_queue(dd);
886 		if (rdd) {
887 			spin_lock(&dd->spin);
888 			if (spin_trylock(&rdd->spin)) {
889 				nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
890 				spin_unlock(&rdd->spin);
891 				if (nlp == NULL)
892 					spin_unlock(&dd->spin);
893 			} else {
894 				spin_unlock(&dd->spin);
895 				nlp = NULL;
896 			}
897 		} else {
898 			nlp = NULL;
899 		}
900 		/* dd->spin held if nlp != NULL */
901 
902 		/*
903 		 * Either schedule it or add it to our queue.
904 		 */
905 		if (nlp &&
906 		    (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
907 			ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, dd->cpumask);
908 			dd->upri = nlp->lwp_priority;
909 			dd->uschedcp = nlp;
910 #if 0
911 			dd->rrcount = 0;	/* reset round robin */
912 #endif
913 			spin_unlock(&dd->spin);
914 			lwkt_acquire(nlp->lwp_thread);
915 			lwkt_schedule(nlp->lwp_thread);
916 		} else if (nlp) {
917 			dfly_setrunqueue_locked(dd, nlp);
918 			spin_unlock(&dd->spin);
919 		}
920 	}
921 }
922 
923 /*
924  * Called from acquire and from kern_synch's one-second timer (one of the
925  * callout helper threads) with a critical section held.
926  *
927  * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
928  * overall system load.
929  *
930  * Note that no recalculation occurs for a process which sleeps and wakes
931  * up in the same tick.  That is, a system doing thousands of context
932  * switches per second will still only do serious estcpu calculations
933  * ESTCPUFREQ times per second.
934  */
935 static
936 void
937 dfly_recalculate_estcpu(struct lwp *lp)
938 {
939 	globaldata_t gd = mycpu;
940 	sysclock_t cpbase;
941 	sysclock_t ttlticks;
942 	int estcpu;
943 	int decay_factor;
944 	int ucount;
945 
946 	/*
947 	 * We have to subtract periodic to get the last schedclock
948 	 * timeout time, otherwise we would get the upcoming timeout.
949 	 * Keep in mind that a process can migrate between cpus and
950 	 * while the scheduler clock should be very close, boundary
951 	 * conditions could lead to a small negative delta.
952 	 */
953 	cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
954 
955 	if (lp->lwp_slptime > 1) {
956 		/*
957 		 * Too much time has passed, do a coarse correction.
958 		 */
959 		lp->lwp_estcpu = lp->lwp_estcpu >> 1;
960 		dfly_resetpriority(lp);
961 		lp->lwp_cpbase = cpbase;
962 		lp->lwp_cpticks = 0;
963 		lp->lwp_estfast = 0;
964 	} else if (lp->lwp_cpbase != cpbase) {
965 		/*
966 		 * Adjust estcpu if we are in a different tick.  Don't waste
967 		 * time if we are in the same tick.
968 		 *
969 		 * First calculate the number of ticks in the measurement
970 		 * interval.  The ttlticks calculation can wind up 0 due to
971 		 * a bug in the handling of lwp_slptime  (as yet not found),
972 		 * so make sure we do not get a divide by 0 panic.
973 		 */
974 		ttlticks = (cpbase - lp->lwp_cpbase) /
975 			   gd->gd_schedclock.periodic;
976 		if ((ssysclock_t)ttlticks < 0) {
977 			ttlticks = 0;
978 			lp->lwp_cpbase = cpbase;
979 		}
980 		if (ttlticks < 4)
981 			return;
982 		updatepcpu(lp, lp->lwp_cpticks, ttlticks);
983 
984 		/*
985 		 * Calculate instant estcpu based percentage of (one) cpu
986 		 * used and exponentially average it into the current
987 		 * lwp_estcpu.
988 		 */
989 		ucount = dfly_pcpu[lp->lwp_qcpu].ucount;
990 		estcpu = lp->lwp_cpticks * ESTCPUMAX / ttlticks;
991 
992 		/*
993 		 * The higher ttlticks gets, the more meaning the calculation
994 		 * has and the smaller our decay_factor in the exponential
995 		 * average.
996 		 *
997 		 * The uload calculation has been removed because it actually
998 		 * makes things worse, causing processes which use less cpu
999 		 * (such as a browser) to be pumped up and treated the same
1000 		 * as a cpu-bound process (such as a make).  The same effect
1001 		 * can occur with sufficient load without the uload
1002 		 * calculation, but occurs less quickly and takes more load.
1003 		 * In addition, the less cpu a process uses the smaller the
1004 		 * effect of the overload.
1005 		 */
1006 		if (ttlticks >= hz)
1007 			decay_factor = 1;
1008 		else
1009 			decay_factor = hz - ttlticks;
1010 
1011 		lp->lwp_estcpu = ESTCPULIM(
1012 				(lp->lwp_estcpu * ttlticks + estcpu) /
1013 				(ttlticks + 1));
1014 		if (usched_dfly_debug == lp->lwp_proc->p_pid)
1015 			kprintf(" finalestcpu %d %d\n", estcpu, lp->lwp_estcpu);
1016 
1017 #if 0
1018 		/*
1019 		 * Calculate the percentage of one cpu being used then
1020 		 * compensate for any system load in excess of ncpus.
1021 		 *
1022 		 * For example, if we have 8 cores and 16 running cpu-bound
1023 		 * processes then all things being equal each process will
1024 		 * get 50% of one cpu.  We need to pump this value back
1025 		 * up to 100% so the estcpu calculation properly adjusts
1026 		 * the process's dynamic priority.
1027 		 *
1028 		 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
1029 		 */
1030 
1031 		estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
1032 		ucount = dfly_ucount;
1033 		if (ucount > ncpus) {
1034 			estcpu += estcpu * (ucount - ncpus) / ncpus;
1035 		}
1036 
1037 		if (usched_dfly_debug == lp->lwp_proc->p_pid) {
1038 			kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
1039 				lp->lwp_proc->p_pid, lp,
1040 				estcpu, lp->lwp_estcpu,
1041 				lp->lwp_cpticks, ttlticks);
1042 		}
1043 
1044 		/*
1045 		 * Adjust lp->lwp_esetcpu.  The decay factor determines how
1046 		 * quickly lwp_estcpu collapses to its realtime calculation.
1047 		 * A slower collapse gives us a more accurate number over
1048 		 * the long term but can create problems with bursty threads
1049 		 * or threads which become cpu hogs.
1050 		 *
1051 		 * To solve this problem, newly started lwps and lwps which
1052 		 * are restarting after having been asleep for a while are
1053 		 * given a much, much faster decay in order to quickly
1054 		 * detect whether they become cpu-bound.
1055 		 *
1056 		 * NOTE: p_nice is accounted for in dfly_resetpriority(),
1057 		 *	 and not here, but we must still ensure that a
1058 		 *	 cpu-bound nice -20 process does not completely
1059 		 *	 override a cpu-bound nice +20 process.
1060 		 *
1061 		 * NOTE: We must use ESTCPULIM() here to deal with any
1062 		 *	 overshoot.
1063 		 */
1064 		decay_factor = usched_dfly_decay;
1065 		if (decay_factor < 1)
1066 			decay_factor = 1;
1067 		if (decay_factor > 1024)
1068 			decay_factor = 1024;
1069 
1070 		if (lp->lwp_estfast < usched_dfly_decay) {
1071 			++lp->lwp_estfast;
1072 			lp->lwp_estcpu = ESTCPULIM(
1073 				(lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
1074 				(lp->lwp_estfast + 1));
1075 		} else {
1076 			lp->lwp_estcpu = ESTCPULIM(
1077 				(lp->lwp_estcpu * decay_factor + estcpu) /
1078 				(decay_factor + 1));
1079 		}
1080 
1081 		if (usched_dfly_debug == lp->lwp_proc->p_pid)
1082 			kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
1083 #endif
1084 		dfly_resetpriority(lp);
1085 		lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
1086 		lp->lwp_cpticks = 0;
1087 	}
1088 }
1089 
1090 /*
1091  * Compute the priority of a process when running in user mode.
1092  * Arrange to reschedule if the resulting priority is better
1093  * than that of the current process.
1094  *
1095  * This routine may be called with any process.
1096  *
1097  * This routine is called by fork1() for initial setup with the process of
1098  * the run queue, and also may be called normally with the process on or
1099  * off the run queue.
1100  */
1101 static void
1102 dfly_resetpriority(struct lwp *lp)
1103 {
1104 	dfly_pcpu_t rdd;
1105 	int newpriority;
1106 	u_short newrqtype;
1107 	int rcpu;
1108 	int checkpri;
1109 	int estcpu;
1110 	int delta_uload;
1111 
1112 	crit_enter();
1113 
1114 	/*
1115 	 * Lock the scheduler (lp) belongs to.  This can be on a different
1116 	 * cpu.  Handle races.  This loop breaks out with the appropriate
1117 	 * rdd locked.
1118 	 */
1119 	for (;;) {
1120 		rcpu = lp->lwp_qcpu;
1121 		cpu_ccfence();
1122 		rdd = &dfly_pcpu[rcpu];
1123 		spin_lock(&rdd->spin);
1124 		if (rcpu == lp->lwp_qcpu)
1125 			break;
1126 		spin_unlock(&rdd->spin);
1127 	}
1128 
1129 	/*
1130 	 * Calculate the new priority and queue type
1131 	 */
1132 	newrqtype = lp->lwp_rtprio.type;
1133 
1134 	switch(newrqtype) {
1135 	case RTP_PRIO_REALTIME:
1136 	case RTP_PRIO_FIFO:
1137 		newpriority = PRIBASE_REALTIME +
1138 			     (lp->lwp_rtprio.prio & PRIMASK);
1139 		break;
1140 	case RTP_PRIO_NORMAL:
1141 		/*
1142 		 * Calculate the new priority.
1143 		 *
1144 		 * nice contributes up to NICE_QS queues (typ 32 - full range)
1145 		 * estcpu contributes up to EST_QS queues (typ 16)
1146 		 *
1147 		 * A nice +20 process receives 1/10 cpu vs nice+0.  Niced
1148 		 * process more than 20 apart may receive no cpu, so cpu
1149 		 * bound nice -20 can prevent a nice +5 from getting any
1150 		 * cpu.  A nice+0, being in the middle, always gets some cpu
1151 		 * no matter what.
1152 		 */
1153 		estcpu = lp->lwp_estcpu;
1154 		newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) *
1155 			      (NICE_QS * PPQ) / PRIO_RANGE;
1156 		newpriority += estcpu * PPQ / ESTCPUPPQ;
1157 		if (newpriority < 0)
1158 			newpriority = 0;
1159 		if (newpriority >= MAXPRI)
1160 			newpriority = MAXPRI - 1;
1161 		newpriority += PRIBASE_NORMAL;
1162 		break;
1163 	case RTP_PRIO_IDLE:
1164 		newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1165 		break;
1166 	case RTP_PRIO_THREAD:
1167 		newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1168 		break;
1169 	default:
1170 		panic("Bad RTP_PRIO %d", newrqtype);
1171 		/* NOT REACHED */
1172 	}
1173 
1174 	/*
1175 	 * The LWKT scheduler doesn't dive usched structures, give it a hint
1176 	 * on the relative priority of user threads running in the kernel.
1177 	 * The LWKT scheduler will always ensure that a user thread running
1178 	 * in the kernel will get cpu some time, regardless of its upri,
1179 	 * but can decide not to instantly switch from one kernel or user
1180 	 * mode user thread to a kernel-mode user thread when it has a less
1181 	 * desireable user priority.
1182 	 *
1183 	 * td_upri has normal sense (higher values are more desireable), so
1184 	 * negate it (this is a different field lp->lwp_priority)
1185 	 */
1186 	lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1187 
1188 	/*
1189 	 * The newpriority incorporates the queue type so do a simple masked
1190 	 * check to determine if the process has moved to another queue.  If
1191 	 * it has, and it is currently on a run queue, then move it.
1192 	 *
1193 	 * Since uload is ~PPQMASK masked, no modifications are necessary if
1194 	 * we end up in the same run queue.
1195 	 *
1196 	 * Reset rrcount if moving to a higher-priority queue, otherwise
1197 	 * retain rrcount.
1198 	 */
1199 	if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1200 		if (lp->lwp_priority < newpriority)
1201 			lp->lwp_rrcount = 0;
1202 		if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1203 			dfly_remrunqueue_locked(rdd, lp);
1204 			lp->lwp_priority = newpriority;
1205 			lp->lwp_rqtype = newrqtype;
1206 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1207 			dfly_setrunqueue_locked(rdd, lp);
1208 			checkpri = 1;
1209 		} else {
1210 			lp->lwp_priority = newpriority;
1211 			lp->lwp_rqtype = newrqtype;
1212 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1213 			checkpri = 0;
1214 		}
1215 	} else {
1216 		/*
1217 		 * In the same PPQ, uload cannot change.
1218 		 */
1219 		lp->lwp_priority = newpriority;
1220 		checkpri = 1;
1221 		rcpu = -1;
1222 	}
1223 
1224 	/*
1225 	 * Adjust effective load.
1226 	 *
1227 	 * Calculate load then scale up or down geometrically based on p_nice.
1228 	 * Processes niced up (positive) are less important, and processes
1229 	 * niced downard (negative) are more important.  The higher the uload,
1230 	 * the more important the thread.
1231 	 */
1232 	/* 0-511, 0-100% cpu */
1233 	delta_uload = lp->lwp_estcpu / NQS;
1234 	delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1235 	delta_uload -= lp->lwp_uload;
1236 	if (lp->lwp_uload + delta_uload < -32767) {
1237 		delta_uload = -32768 - lp->lwp_uload;
1238 	} else if (lp->lwp_uload + delta_uload > 32767) {
1239 		delta_uload = 32767 - lp->lwp_uload;
1240 	}
1241 	lp->lwp_uload += delta_uload;
1242 	if (lp->lwp_mpflags & LWP_MP_ULOAD)
1243 		atomic_add_long(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1244 
1245 	/*
1246 	 * Determine if we need to reschedule the target cpu.  This only
1247 	 * occurs if the LWP is already on a scheduler queue, which means
1248 	 * that idle cpu notification has already occured.  At most we
1249 	 * need only issue a need_user_resched() on the appropriate cpu.
1250 	 *
1251 	 * The LWP may be owned by a CPU different from the current one,
1252 	 * in which case dd->uschedcp may be modified without an MP lock
1253 	 * or a spinlock held.  The worst that happens is that the code
1254 	 * below causes a spurious need_user_resched() on the target CPU
1255 	 * and dd->pri to be wrong for a short period of time, both of
1256 	 * which are harmless.
1257 	 *
1258 	 * If checkpri is 0 we are adjusting the priority of the current
1259 	 * process, possibly higher (less desireable), so ignore the upri
1260 	 * check which will fail in that case.
1261 	 */
1262 	if (rcpu >= 0) {
1263 		if (CPUMASK_TESTBIT(dfly_rdyprocmask, rcpu) &&
1264 		    (checkpri == 0 ||
1265 		     (rdd->upri & ~PRIMASK) >
1266 		     (lp->lwp_priority & ~PRIMASK))) {
1267 			if (rcpu == mycpu->gd_cpuid) {
1268 				spin_unlock(&rdd->spin);
1269 				need_user_resched();
1270 			} else {
1271 				spin_unlock(&rdd->spin);
1272 				lwkt_send_ipiq(globaldata_find(rcpu),
1273 					       dfly_need_user_resched_remote,
1274 					       NULL);
1275 			}
1276 		} else {
1277 			spin_unlock(&rdd->spin);
1278 		}
1279 	} else {
1280 		spin_unlock(&rdd->spin);
1281 	}
1282 	crit_exit();
1283 }
1284 
1285 static
1286 void
1287 dfly_yield(struct lwp *lp)
1288 {
1289 	if (lp->lwp_qcpu != mycpu->gd_cpuid)
1290 		return;
1291 	KKASSERT(lp == curthread->td_lwp);
1292 
1293 	/*
1294 	 * Don't set need_user_resched() or mess with rrcount or anything.
1295 	 * the TDF flag will override everything as long as we release.
1296 	 */
1297 	atomic_set_int(&lp->lwp_thread->td_mpflags, TDF_MP_DIDYIELD);
1298 	dfly_release_curproc(lp);
1299 }
1300 
1301 /*
1302  * Thread was forcefully migrated to another cpu.  Normally forced migrations
1303  * are used for iterations and the kernel returns to the original cpu before
1304  * returning and this is not needed.  However, if the kernel migrates a
1305  * thread to another cpu and wants to leave it there, it has to call this
1306  * scheduler helper.
1307  *
1308  * Note that the lwkt_migratecpu() function also released the thread, so
1309  * we don't have to worry about that.
1310  */
1311 static
1312 void
1313 dfly_changedcpu(struct lwp *lp)
1314 {
1315 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1316 	dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid];
1317 
1318 	if (dd != rdd) {
1319 		spin_lock(&dd->spin);
1320 		dfly_changeqcpu_locked(lp, dd, rdd);
1321 		spin_unlock(&dd->spin);
1322 	}
1323 }
1324 
1325 /*
1326  * Called from fork1() when a new child process is being created.
1327  *
1328  * Give the child process an initial estcpu that is more batch then
1329  * its parent and dock the parent for the fork (but do not
1330  * reschedule the parent).
1331  *
1332  * fast
1333  *
1334  * XXX lwp should be "spawning" instead of "forking"
1335  */
1336 static void
1337 dfly_forking(struct lwp *plp, struct lwp *lp)
1338 {
1339 	int estcpu;
1340 
1341 	/*
1342 	 * Put the child 4 queue slots (out of 32) higher than the parent
1343 	 * (less desireable than the parent).
1344 	 */
1345 	lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu +
1346 				   ESTCPUPPQ * usched_dfly_forkbias);
1347 	lp->lwp_forked = 1;
1348 	lp->lwp_estfast = 0;
1349 
1350 	/*
1351 	 * Even though the lp will be scheduled specially the first time
1352 	 * due to lp->lwp_forked, it is important to initialize lwp_qcpu
1353 	 * to avoid favoring a fixed cpu.
1354 	 */
1355 #if 0
1356 	static uint16_t save_cpu;
1357 	lp->lwp_qcpu = ++save_cpu % ncpus;
1358 #else
1359 	lp->lwp_qcpu = plp->lwp_qcpu;
1360 	if (CPUMASK_TESTBIT(lp->lwp_cpumask, lp->lwp_qcpu) == 0)
1361 		lp->lwp_qcpu = BSFCPUMASK(lp->lwp_cpumask);
1362 #endif
1363 
1364 	/*
1365 	 * Dock the parent a cost for the fork, protecting us from fork
1366 	 * bombs.  If the parent is forking quickly this makes both the
1367 	 * parent and child more batchy.
1368 	 */
1369 	estcpu = plp->lwp_estcpu + ESTCPUPPQ / 16;
1370 	plp->lwp_estcpu = ESTCPULIM(estcpu);
1371 }
1372 
1373 /*
1374  * Called when a lwp is being removed from this scheduler, typically
1375  * during lwp_exit().  We have to clean out any ULOAD accounting before
1376  * we can let the lp go.  The dd->spin lock is not needed for uload
1377  * updates.
1378  *
1379  * Scheduler dequeueing has already occurred, no further action in that
1380  * regard is needed.
1381  */
1382 static void
1383 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1384 {
1385 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1386 
1387 	if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1388 		atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1389 		atomic_add_long(&dd->uload, -lp->lwp_uload);
1390 		atomic_add_int(&dd->ucount, -1);
1391 	}
1392 }
1393 
1394 /*
1395  * This function cannot block in any way, but spinlocks are ok.
1396  *
1397  * Update the uload based on the state of the thread (whether it is going
1398  * to sleep or running again).  The uload is meant to be a longer-term
1399  * load and not an instantanious load.
1400  */
1401 static void
1402 dfly_uload_update(struct lwp *lp)
1403 {
1404 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1405 
1406 	if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1407 		if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1408 			spin_lock(&dd->spin);
1409 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1410 				atomic_set_int(&lp->lwp_mpflags,
1411 					       LWP_MP_ULOAD);
1412 				atomic_add_long(&dd->uload, lp->lwp_uload);
1413 				atomic_add_int(&dd->ucount, 1);
1414 			}
1415 			spin_unlock(&dd->spin);
1416 		}
1417 	} else if (lp->lwp_slptime > 0) {
1418 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1419 			spin_lock(&dd->spin);
1420 			if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1421 				atomic_clear_int(&lp->lwp_mpflags,
1422 						 LWP_MP_ULOAD);
1423 				atomic_add_long(&dd->uload, -lp->lwp_uload);
1424 				atomic_add_int(&dd->ucount, -1);
1425 			}
1426 			spin_unlock(&dd->spin);
1427 		}
1428 	}
1429 }
1430 
1431 /*
1432  * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1433  * it selects a user process and returns it.  If chklp is non-NULL and chklp
1434  * has a better or equal priority then the process that would otherwise be
1435  * chosen, NULL is returned.
1436  *
1437  * Until we fix the RUNQ code the chklp test has to be strict or we may
1438  * bounce between processes trying to acquire the current process designation.
1439  *
1440  * Must be called with rdd->spin locked.  The spinlock is left intact through
1441  * the entire routine.  dd->spin does not have to be locked.
1442  *
1443  * If worst is non-zero this function finds the worst thread instead of the
1444  * best thread (used by the schedulerclock-based rover).
1445  */
1446 static
1447 struct lwp *
1448 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1449 		       struct lwp *chklp, int worst)
1450 {
1451 	struct lwp *lp;
1452 	struct rq *q;
1453 	u_int32_t *which;
1454 	u_int32_t pri;
1455 	u_int32_t rtqbits;
1456 	u_int32_t tsqbits;
1457 	u_int32_t idqbits;
1458 
1459 	rtqbits = rdd->rtqueuebits;
1460 	tsqbits = rdd->queuebits;
1461 	idqbits = rdd->idqueuebits;
1462 
1463 	if (worst) {
1464 		if (idqbits) {
1465 			pri = bsrl(idqbits);
1466 			q = &rdd->idqueues[pri];
1467 			which = &rdd->idqueuebits;
1468 		} else if (tsqbits) {
1469 			pri = bsrl(tsqbits);
1470 			q = &rdd->queues[pri];
1471 			which = &rdd->queuebits;
1472 		} else if (rtqbits) {
1473 			pri = bsrl(rtqbits);
1474 			q = &rdd->rtqueues[pri];
1475 			which = &rdd->rtqueuebits;
1476 		} else {
1477 			return (NULL);
1478 		}
1479 		lp = TAILQ_LAST(q, rq);
1480 	} else {
1481 		if (rtqbits) {
1482 			pri = bsfl(rtqbits);
1483 			q = &rdd->rtqueues[pri];
1484 			which = &rdd->rtqueuebits;
1485 		} else if (tsqbits) {
1486 			pri = bsfl(tsqbits);
1487 			q = &rdd->queues[pri];
1488 			which = &rdd->queuebits;
1489 		} else if (idqbits) {
1490 			pri = bsfl(idqbits);
1491 			q = &rdd->idqueues[pri];
1492 			which = &rdd->idqueuebits;
1493 		} else {
1494 			return (NULL);
1495 		}
1496 		lp = TAILQ_FIRST(q);
1497 	}
1498 	KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1499 
1500 	/*
1501 	 * If the passed lwp <chklp> is reasonably close to the selected
1502 	 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1503 	 *
1504 	 * Note that we must error on the side of <chklp> to avoid bouncing
1505 	 * between threads in the acquire code.
1506 	 */
1507 	if (chklp) {
1508 		if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1509 			return(NULL);
1510 	}
1511 
1512 	KTR_COND_LOG(usched_chooseproc,
1513 	    lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1514 	    lp->lwp_proc->p_pid,
1515 	    lp->lwp_thread->td_gd->gd_cpuid,
1516 	    mycpu->gd_cpuid);
1517 
1518 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1519 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1520 	TAILQ_REMOVE(q, lp, lwp_procq);
1521 	--rdd->runqcount;
1522 	if (TAILQ_EMPTY(q))
1523 		*which &= ~(1 << pri);
1524 
1525 	/*
1526 	 * If we are choosing a process from rdd with the intent to
1527 	 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1528 	 * is still held.
1529 	 */
1530 	if (rdd != dd) {
1531 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1532 			atomic_add_long(&rdd->uload, -lp->lwp_uload);
1533 			atomic_add_int(&rdd->ucount, -1);
1534 		}
1535 		lp->lwp_qcpu = dd->cpuid;
1536 		atomic_add_long(&dd->uload, lp->lwp_uload);
1537 		atomic_add_int(&dd->ucount, 1);
1538 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1539 	}
1540 	return lp;
1541 }
1542 
1543 /*
1544  * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1545  *
1546  * Choose a cpu node to schedule lp on, hopefully nearby its current
1547  * node.
1548  *
1549  * We give the current node a modest advantage for obvious reasons.
1550  *
1551  * We also give the node the thread was woken up FROM a slight advantage
1552  * in order to try to schedule paired threads which synchronize/block waiting
1553  * for each other fairly close to each other.  Similarly in a network setting
1554  * this feature will also attempt to place a user process near the kernel
1555  * protocol thread that is feeding it data.  THIS IS A CRITICAL PART of the
1556  * algorithm as it heuristically groups synchronizing processes for locality
1557  * of reference in multi-socket systems.
1558  *
1559  * We check against running processes and give a big advantage if there
1560  * are none running.
1561  *
1562  * The caller will normally dfly_setrunqueue() lp on the returned queue.
1563  *
1564  * When the topology is known choose a cpu whos group has, in aggregate,
1565  * has the lowest weighted load.
1566  */
1567 static
1568 dfly_pcpu_t
1569 dfly_choose_best_queue(struct lwp *lp)
1570 {
1571 	cpumask_t wakemask;
1572 	cpumask_t mask;
1573 	cpu_node_t *cpup;
1574 	cpu_node_t *cpun;
1575 	cpu_node_t *cpub;
1576 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1577 	dfly_pcpu_t rdd;
1578 	int wakecpu;
1579 	int cpuid;
1580 	int n;
1581 	int count;
1582 	long load;
1583 	long lowest_load;
1584 
1585 	/*
1586 	 * When the topology is unknown choose a random cpu that is hopefully
1587 	 * idle.
1588 	 */
1589 	if (dd->cpunode == NULL)
1590 		return (dfly_choose_queue_simple(dd, lp));
1591 
1592 	/*
1593 	 * Pairing mask
1594 	 */
1595 	if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1596 		wakemask = dfly_pcpu[wakecpu].cpumask;
1597 	else
1598 		CPUMASK_ASSZERO(wakemask);
1599 
1600 	/*
1601 	 * When the topology is known choose a cpu whos group has, in
1602 	 * aggregate, has the lowest weighted load.
1603 	 */
1604 	cpup = root_cpu_node;
1605 	rdd = dd;
1606 
1607 	while (cpup) {
1608 		/*
1609 		 * Degenerate case super-root
1610 		 */
1611 		if (cpup->child_no == 1) {
1612 			cpup = cpup->child_node[0];
1613 			continue;
1614 		}
1615 
1616 		/*
1617 		 * Terminal cpunode
1618 		 */
1619 		if (cpup->child_no == 0) {
1620 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1621 			break;
1622 		}
1623 
1624 		cpub = NULL;
1625 		lowest_load = 0x7FFFFFFFFFFFFFFFLLU;
1626 
1627 		for (n = 0; n < cpup->child_no; ++n) {
1628 			/*
1629 			 * Accumulate load information for all cpus
1630 			 * which are members of this node.
1631 			 */
1632 			cpun = cpup->child_node[n];
1633 			mask = cpun->members;
1634 			CPUMASK_ANDMASK(mask, usched_global_cpumask);
1635 			CPUMASK_ANDMASK(mask, smp_active_mask);
1636 			CPUMASK_ANDMASK(mask, lp->lwp_cpumask);
1637 			if (CPUMASK_TESTZERO(mask))
1638 				continue;
1639 
1640 			count = 0;
1641 			load = 0;
1642 
1643 			while (CPUMASK_TESTNZERO(mask)) {
1644 				cpuid = BSFCPUMASK(mask);
1645 				rdd = &dfly_pcpu[cpuid];
1646 				load += rdd->uload;
1647 				load += rdd->ucount * usched_dfly_weight3;
1648 
1649 				if (rdd->uschedcp == NULL &&
1650 				    rdd->runqcount == 0 &&
1651 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1652 				) {
1653 					load -= usched_dfly_weight4;
1654 				}
1655 #if 0
1656 				else if (rdd->upri > lp->lwp_priority + PPQ) {
1657 					load -= usched_dfly_weight4 / 2;
1658 				}
1659 #endif
1660 				CPUMASK_NANDBIT(mask, cpuid);
1661 				++count;
1662 			}
1663 
1664 			/*
1665 			 * Compensate if the lp is already accounted for in
1666 			 * the aggregate uload for this mask set.  We want
1667 			 * to calculate the loads as if lp were not present,
1668 			 * otherwise the calculation is bogus.
1669 			 */
1670 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1671 			    CPUMASK_TESTMASK(dd->cpumask, cpun->members)) {
1672 				load -= lp->lwp_uload;
1673 				load -= usched_dfly_weight3;
1674 			}
1675 
1676 			load /= count;
1677 
1678 			/*
1679 			 * Advantage the cpu group (lp) is already on.
1680 			 */
1681 			if (CPUMASK_TESTMASK(cpun->members, dd->cpumask))
1682 				load -= usched_dfly_weight1;
1683 
1684 			/*
1685 			 * Advantage the cpu group we want to pair (lp) to,
1686 			 * but don't let it go to the exact same cpu as
1687 			 * the wakecpu target.
1688 			 *
1689 			 * We do this by checking whether cpun is a
1690 			 * terminal node or not.  All cpun's at the same
1691 			 * level will either all be terminal or all not
1692 			 * terminal.
1693 			 *
1694 			 * If it is and we match we disadvantage the load.
1695 			 * If it is and we don't match we advantage the load.
1696 			 *
1697 			 * Also note that we are effectively disadvantaging
1698 			 * all-but-one by the same amount, so it won't effect
1699 			 * the weight1 factor for the all-but-one nodes.
1700 			 */
1701 			if (CPUMASK_TESTMASK(cpun->members, wakemask)) {
1702 				if (cpun->child_no != 0) {
1703 					/* advantage */
1704 					load -= usched_dfly_weight2;
1705 				} else {
1706 					if (usched_dfly_features & 0x10)
1707 						load += usched_dfly_weight2;
1708 					else
1709 						load -= usched_dfly_weight2;
1710 				}
1711 			}
1712 
1713 			/*
1714 			 * Calculate the best load
1715 			 */
1716 			if (cpub == NULL || lowest_load > load ||
1717 			    (lowest_load == load &&
1718 			     CPUMASK_TESTMASK(cpun->members, dd->cpumask))
1719 			) {
1720 				lowest_load = load;
1721 				cpub = cpun;
1722 			}
1723 		}
1724 		cpup = cpub;
1725 	}
1726 	/* Dispatch this outcast to a proper CPU. */
1727 	if (__predict_false(CPUMASK_TESTBIT(lp->lwp_cpumask, rdd->cpuid) == 0))
1728 		rdd = &dfly_pcpu[BSFCPUMASK(lp->lwp_cpumask)];
1729 	if (usched_dfly_chooser > 0) {
1730 		--usched_dfly_chooser;		/* only N lines */
1731 		kprintf("lp %02d->%02d %s\n",
1732 			lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1733 	}
1734 	return (rdd);
1735 }
1736 
1737 /*
1738  * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1739  *
1740  * Choose the worst queue close to dd's cpu node with a non-empty runq
1741  * that is NOT dd.  Also require that the moving of the highest-load thread
1742  * from rdd to dd does not cause the uload's to cross each other.
1743  *
1744  * This is used by the thread chooser when the current cpu's queues are
1745  * empty to steal a thread from another cpu's queue.  We want to offload
1746  * the most heavily-loaded queue.
1747  */
1748 static
1749 dfly_pcpu_t
1750 dfly_choose_worst_queue(dfly_pcpu_t dd)
1751 {
1752 	cpumask_t mask;
1753 	cpu_node_t *cpup;
1754 	cpu_node_t *cpun;
1755 	cpu_node_t *cpub;
1756 	dfly_pcpu_t rdd;
1757 	int cpuid;
1758 	int n;
1759 	int count;
1760 	long load;
1761 	long highest_load;
1762 #if 0
1763 	int pri;
1764 	int hpri;
1765 #endif
1766 
1767 	/*
1768 	 * When the topology is unknown choose a random cpu that is hopefully
1769 	 * idle.
1770 	 */
1771 	if (dd->cpunode == NULL) {
1772 		return (NULL);
1773 	}
1774 
1775 	/*
1776 	 * When the topology is known choose a cpu whos group has, in
1777 	 * aggregate, has the highest weighted load.
1778 	 */
1779 	cpup = root_cpu_node;
1780 	rdd = dd;
1781 	while (cpup) {
1782 		/*
1783 		 * Degenerate case super-root
1784 		 */
1785 		if (cpup->child_no == 1) {
1786 			cpup = cpup->child_node[0];
1787 			continue;
1788 		}
1789 
1790 		/*
1791 		 * Terminal cpunode
1792 		 */
1793 		if (cpup->child_no == 0) {
1794 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1795 			break;
1796 		}
1797 
1798 		cpub = NULL;
1799 		highest_load = 0;
1800 
1801 		for (n = 0; n < cpup->child_no; ++n) {
1802 			/*
1803 			 * Accumulate load information for all cpus
1804 			 * which are members of this node.
1805 			 */
1806 			cpun = cpup->child_node[n];
1807 			mask = cpun->members;
1808 			CPUMASK_ANDMASK(mask, usched_global_cpumask);
1809 			CPUMASK_ANDMASK(mask, smp_active_mask);
1810 			if (CPUMASK_TESTZERO(mask))
1811 				continue;
1812 
1813 			count = 0;
1814 			load = 0;
1815 
1816 			while (CPUMASK_TESTNZERO(mask)) {
1817 				cpuid = BSFCPUMASK(mask);
1818 				rdd = &dfly_pcpu[cpuid];
1819 				load += rdd->uload;
1820 				load += (long)rdd->ucount * usched_dfly_weight3;
1821 
1822 				if (rdd->uschedcp == NULL &&
1823 				    rdd->runqcount == 0 &&
1824 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1825 				) {
1826 					load -= usched_dfly_weight4;
1827 				}
1828 #if 0
1829 				else if (rdd->upri > dd->upri + PPQ) {
1830 					load -= usched_dfly_weight4 / 2;
1831 				}
1832 #endif
1833 				CPUMASK_NANDBIT(mask, cpuid);
1834 				++count;
1835 			}
1836 			load /= count;
1837 
1838 			/*
1839 			 * Prefer candidates which are somewhat closer to
1840 			 * our cpu.
1841 			 */
1842 			if (CPUMASK_TESTMASK(dd->cpumask, cpun->members))
1843 				load += usched_dfly_weight1;
1844 
1845 			/*
1846 			 * The best candidate is the one with the worst
1847 			 * (highest) load.
1848 			 */
1849 			if (cpub == NULL || highest_load < load ||
1850 			    (highest_load == load &&
1851 			     CPUMASK_TESTMASK(cpun->members, dd->cpumask))) {
1852 				highest_load = load;
1853 				cpub = cpun;
1854 			}
1855 		}
1856 		cpup = cpub;
1857 	}
1858 
1859 	/*
1860 	 * We never return our own node (dd), and only return a remote
1861 	 * node if it's load is significantly worse than ours (i.e. where
1862 	 * stealing a thread would be considered reasonable).
1863 	 *
1864 	 * This also helps us avoid breaking paired threads apart which
1865 	 * can have disastrous effects on performance.
1866 	 */
1867 	if (rdd == dd)
1868 		return(NULL);
1869 
1870 #if 0
1871 	hpri = 0;
1872 	if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1873 		hpri = pri;
1874 	if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1875 		hpri = pri;
1876 	if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1877 		hpri = pri;
1878 	hpri *= PPQ;
1879 	if (rdd->uload - hpri < dd->uload + hpri)
1880 		return(NULL);
1881 #endif
1882 	return (rdd);
1883 }
1884 
1885 static
1886 dfly_pcpu_t
1887 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1888 {
1889 	dfly_pcpu_t rdd;
1890 	cpumask_t tmpmask;
1891 	cpumask_t mask;
1892 	int cpubase;
1893 	int cpuid;
1894 
1895 	/*
1896 	 * Fallback to the original heuristic, select random cpu,
1897 	 * first checking the cpus not currently running a user thread.
1898 	 *
1899 	 * Use cpuid as the base cpu in our scan, first checking
1900 	 * cpuid...(ncpus-1), then 0...(cpuid-1).  This avoid favoring
1901 	 * lower-numbered cpus.
1902 	 */
1903 	++dd->scancpu;		/* SMP race ok */
1904 	mask = dfly_rdyprocmask;
1905 	CPUMASK_NANDMASK(mask, dfly_curprocmask);
1906 	CPUMASK_ANDMASK(mask, lp->lwp_cpumask);
1907 	CPUMASK_ANDMASK(mask, smp_active_mask);
1908 	CPUMASK_ANDMASK(mask, usched_global_cpumask);
1909 
1910 	cpubase = (int)(dd->scancpu % ncpus);
1911 	CPUMASK_ASSBMASK(tmpmask, cpubase);
1912 	CPUMASK_INVMASK(tmpmask);
1913 	CPUMASK_ANDMASK(tmpmask, mask);
1914 	while (CPUMASK_TESTNZERO(tmpmask)) {
1915 		cpuid = BSFCPUMASK(tmpmask);
1916 		rdd = &dfly_pcpu[cpuid];
1917 
1918 		if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1919 			goto found;
1920 		CPUMASK_NANDBIT(tmpmask, cpuid);
1921 	}
1922 
1923 	CPUMASK_ASSBMASK(tmpmask, cpubase);
1924 	CPUMASK_ANDMASK(tmpmask, mask);
1925 	while (CPUMASK_TESTNZERO(tmpmask)) {
1926 		cpuid = BSFCPUMASK(tmpmask);
1927 		rdd = &dfly_pcpu[cpuid];
1928 
1929 		if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1930 			goto found;
1931 		CPUMASK_NANDBIT(tmpmask, cpuid);
1932 	}
1933 
1934 	/*
1935 	 * Then cpus which might have a currently running lp
1936 	 */
1937 	mask = dfly_rdyprocmask;
1938 	CPUMASK_ANDMASK(mask, dfly_curprocmask);
1939 	CPUMASK_ANDMASK(mask, lp->lwp_cpumask);
1940 	CPUMASK_ANDMASK(mask, smp_active_mask);
1941 	CPUMASK_ANDMASK(mask, usched_global_cpumask);
1942 
1943 	CPUMASK_ASSBMASK(tmpmask, cpubase);
1944 	CPUMASK_INVMASK(tmpmask);
1945 	CPUMASK_ANDMASK(tmpmask, mask);
1946 	while (CPUMASK_TESTNZERO(tmpmask)) {
1947 		cpuid = BSFCPUMASK(tmpmask);
1948 		rdd = &dfly_pcpu[cpuid];
1949 
1950 		if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1951 			goto found;
1952 		CPUMASK_NANDBIT(tmpmask, cpuid);
1953 	}
1954 
1955 	CPUMASK_ASSBMASK(tmpmask, cpubase);
1956 	CPUMASK_ANDMASK(tmpmask, mask);
1957 	while (CPUMASK_TESTNZERO(tmpmask)) {
1958 		cpuid = BSFCPUMASK(tmpmask);
1959 		rdd = &dfly_pcpu[cpuid];
1960 
1961 		if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1962 			goto found;
1963 		CPUMASK_NANDBIT(tmpmask, cpuid);
1964 	}
1965 
1966 	/*
1967 	 * If we cannot find a suitable cpu we round-robin using scancpu.
1968 	 * Other cpus will pickup as they release their current lwps or
1969 	 * become ready.
1970 	 *
1971 	 * Avoid a degenerate system lockup case if usched_global_cpumask
1972 	 * is set to 0 or otherwise does not cover lwp_cpumask.
1973 	 *
1974 	 * We only kick the target helper thread in this case, we do not
1975 	 * set the user resched flag because
1976 	 */
1977 	cpuid = cpubase;
1978 	if (CPUMASK_TESTBIT(lp->lwp_cpumask, cpuid) == 0)
1979 		cpuid = BSFCPUMASK(lp->lwp_cpumask);
1980 	else if (CPUMASK_TESTBIT(usched_global_cpumask, cpuid) == 0)
1981 		cpuid = 0;
1982 	rdd = &dfly_pcpu[cpuid];
1983 found:
1984 	return (rdd);
1985 }
1986 
1987 static
1988 void
1989 dfly_need_user_resched_remote(void *dummy)
1990 {
1991 	globaldata_t gd = mycpu;
1992 	dfly_pcpu_t  dd = &dfly_pcpu[gd->gd_cpuid];
1993 
1994 	/*
1995 	 * Flag reschedule needed
1996 	 */
1997 	need_user_resched();
1998 
1999 	/*
2000 	 * If no user thread is currently running we need to kick the helper
2001 	 * on our cpu to recover.  Otherwise the cpu will never schedule
2002 	 * anything again.
2003 	 *
2004 	 * We cannot schedule the process ourselves because this is an
2005 	 * IPI callback and we cannot acquire spinlocks in an IPI callback.
2006 	 *
2007 	 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
2008 	 */
2009 	if (dd->uschedcp == NULL &&
2010 	    CPUMASK_TESTBIT(dfly_rdyprocmask, gd->gd_cpuid)) {
2011 		ATOMIC_CPUMASK_NANDBIT(dfly_rdyprocmask, gd->gd_cpuid);
2012 		wakeup_mycpu(dd->helper_thread);
2013 	}
2014 }
2015 
2016 /*
2017  * dfly_remrunqueue_locked() removes a given process from the run queue
2018  * that it is on, clearing the queue busy bit if it becomes empty.
2019  *
2020  * Note that user process scheduler is different from the LWKT schedule.
2021  * The user process scheduler only manages user processes but it uses LWKT
2022  * underneath, and a user process operating in the kernel will often be
2023  * 'released' from our management.
2024  *
2025  * uload is NOT adjusted here.  It is only adjusted if the lwkt_thread goes
2026  * to sleep or the lwp is moved to a different runq.
2027  */
2028 static void
2029 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
2030 {
2031 	struct rq *q;
2032 	u_int32_t *which;
2033 	u_int8_t pri;
2034 
2035 	KKASSERT(rdd->runqcount >= 0);
2036 
2037 	pri = lp->lwp_rqindex;
2038 
2039 	switch(lp->lwp_rqtype) {
2040 	case RTP_PRIO_NORMAL:
2041 		q = &rdd->queues[pri];
2042 		which = &rdd->queuebits;
2043 		break;
2044 	case RTP_PRIO_REALTIME:
2045 	case RTP_PRIO_FIFO:
2046 		q = &rdd->rtqueues[pri];
2047 		which = &rdd->rtqueuebits;
2048 		break;
2049 	case RTP_PRIO_IDLE:
2050 		q = &rdd->idqueues[pri];
2051 		which = &rdd->idqueuebits;
2052 		break;
2053 	default:
2054 		panic("remrunqueue: invalid rtprio type");
2055 		/* NOT REACHED */
2056 	}
2057 	KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
2058 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
2059 	TAILQ_REMOVE(q, lp, lwp_procq);
2060 	--rdd->runqcount;
2061 	if (TAILQ_EMPTY(q)) {
2062 		KASSERT((*which & (1 << pri)) != 0,
2063 			("remrunqueue: remove from empty queue"));
2064 		*which &= ~(1 << pri);
2065 	}
2066 }
2067 
2068 /*
2069  * dfly_setrunqueue_locked()
2070  *
2071  * Add a process whos rqtype and rqindex had previously been calculated
2072  * onto the appropriate run queue.   Determine if the addition requires
2073  * a reschedule on a cpu and return the cpuid or -1.
2074  *
2075  * NOTE: 	  Lower priorities are better priorities.
2076  *
2077  * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
2078  *		  sum of the rough lwp_priority for all running and runnable
2079  *		  processes.  Lower priority processes (higher lwp_priority
2080  *		  values) actually DO count as more load, not less, because
2081  *		  these are the programs which require the most care with
2082  *		  regards to cpu selection.
2083  */
2084 static void
2085 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
2086 {
2087 	u_int32_t *which;
2088 	struct rq *q;
2089 	int pri;
2090 
2091 	KKASSERT(lp->lwp_qcpu == rdd->cpuid);
2092 
2093 	if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
2094 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
2095 		atomic_add_long(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
2096 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
2097 	}
2098 
2099 	pri = lp->lwp_rqindex;
2100 
2101 	switch(lp->lwp_rqtype) {
2102 	case RTP_PRIO_NORMAL:
2103 		q = &rdd->queues[pri];
2104 		which = &rdd->queuebits;
2105 		break;
2106 	case RTP_PRIO_REALTIME:
2107 	case RTP_PRIO_FIFO:
2108 		q = &rdd->rtqueues[pri];
2109 		which = &rdd->rtqueuebits;
2110 		break;
2111 	case RTP_PRIO_IDLE:
2112 		q = &rdd->idqueues[pri];
2113 		which = &rdd->idqueuebits;
2114 		break;
2115 	default:
2116 		panic("remrunqueue: invalid rtprio type");
2117 		/* NOT REACHED */
2118 	}
2119 
2120 	/*
2121 	 * Place us on the selected queue.  Determine if we should be
2122 	 * placed at the head of the queue or at the end.
2123 	 *
2124 	 * We are placed at the tail if our round-robin count has expired,
2125 	 * or is about to expire and the system thinks its a good place to
2126 	 * round-robin, or there is already a next thread on the queue
2127 	 * (it might be trying to pick up where it left off and we don't
2128 	 * want to interfere).
2129 	 */
2130 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
2131 	atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
2132 	++rdd->runqcount;
2133 
2134 	if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
2135 	    (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
2136 	     (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC))
2137 	) {
2138 		/*
2139 		 * Place on tail
2140 		 */
2141 		atomic_clear_int(&lp->lwp_thread->td_mpflags,
2142 				 TDF_MP_BATCH_DEMARC);
2143 		lp->lwp_rrcount = 0;
2144 		TAILQ_INSERT_TAIL(q, lp, lwp_procq);
2145 	} else {
2146 		/*
2147 		 * Retain rrcount and place on head.  Count is retained
2148 		 * even if the queue is empty.
2149 		 */
2150 		TAILQ_INSERT_HEAD(q, lp, lwp_procq);
2151 	}
2152 	*which |= 1 << pri;
2153 }
2154 
2155 /*
2156  * For SMP systems a user scheduler helper thread is created for each
2157  * cpu and is used to allow one cpu to wakeup another for the purposes of
2158  * scheduling userland threads from setrunqueue().
2159  *
2160  * UP systems do not need the helper since there is only one cpu.
2161  *
2162  * We can't use the idle thread for this because we might block.
2163  * Additionally, doing things this way allows us to HLT idle cpus
2164  * on MP systems.
2165  */
2166 static void
2167 dfly_helper_thread(void *dummy)
2168 {
2169     globaldata_t gd;
2170     dfly_pcpu_t dd;
2171     dfly_pcpu_t rdd;
2172     struct lwp *nlp;
2173     cpumask_t mask;
2174     int cpuid;
2175 
2176     gd = mycpu;
2177     cpuid = gd->gd_cpuid;	/* doesn't change */
2178     mask = gd->gd_cpumask;	/* doesn't change */
2179     dd = &dfly_pcpu[cpuid];
2180 
2181     /*
2182      * Since we only want to be woken up only when no user processes
2183      * are scheduled on a cpu, run at an ultra low priority.
2184      */
2185     lwkt_setpri_self(TDPRI_USER_SCHEDULER);
2186 
2187     tsleep(dd->helper_thread, 0, "schslp", 0);
2188 
2189     for (;;) {
2190 	/*
2191 	 * We use the LWKT deschedule-interlock trick to avoid racing
2192 	 * dfly_rdyprocmask.  This means we cannot block through to the
2193 	 * manual lwkt_switch() call we make below.
2194 	 */
2195 	crit_enter_gd(gd);
2196 	tsleep_interlock(dd->helper_thread, 0);
2197 
2198 	spin_lock(&dd->spin);
2199 
2200 	ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask);
2201 	clear_user_resched();	/* This satisfied the reschedule request */
2202 #if 0
2203 	dd->rrcount = 0;	/* Reset the round-robin counter */
2204 #endif
2205 
2206 	if (dd->runqcount || dd->uschedcp != NULL) {
2207 		/*
2208 		 * Threads are available.  A thread may or may not be
2209 		 * currently scheduled.  Get the best thread already queued
2210 		 * to this cpu.
2211 		 */
2212 		nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
2213 		if (nlp) {
2214 			ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask);
2215 			dd->upri = nlp->lwp_priority;
2216 			dd->uschedcp = nlp;
2217 #if 0
2218 			dd->rrcount = 0;	/* reset round robin */
2219 #endif
2220 			spin_unlock(&dd->spin);
2221 			lwkt_acquire(nlp->lwp_thread);
2222 			lwkt_schedule(nlp->lwp_thread);
2223 		} else {
2224 			/*
2225 			 * This situation should not occur because we had
2226 			 * at least one thread available.
2227 			 */
2228 			spin_unlock(&dd->spin);
2229 		}
2230 	} else if (usched_dfly_features & 0x01) {
2231 		/*
2232 		 * This cpu is devoid of runnable threads, steal a thread
2233 		 * from another cpu.  Since we're stealing, might as well
2234 		 * load balance at the same time.
2235 		 *
2236 		 * We choose the highest-loaded thread from the worst queue.
2237 		 *
2238 		 * NOTE! This function only returns a non-NULL rdd when
2239 		 *	 another cpu's queue is obviously overloaded.  We
2240 		 *	 do not want to perform the type of rebalancing
2241 		 *	 the schedclock does here because it would result
2242 		 *	 in insane process pulling when 'steady' state is
2243 		 *	 partially unbalanced (e.g. 6 runnables and only
2244 		 *	 4 cores).
2245 		 */
2246 		rdd = dfly_choose_worst_queue(dd);
2247 		if (rdd && spin_trylock(&rdd->spin)) {
2248 			nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2249 			spin_unlock(&rdd->spin);
2250 		} else {
2251 			nlp = NULL;
2252 		}
2253 		if (nlp) {
2254 			ATOMIC_CPUMASK_ORMASK(dfly_curprocmask, mask);
2255 			dd->upri = nlp->lwp_priority;
2256 			dd->uschedcp = nlp;
2257 #if 0
2258 			dd->rrcount = 0;	/* reset round robin */
2259 #endif
2260 			spin_unlock(&dd->spin);
2261 			lwkt_acquire(nlp->lwp_thread);
2262 			lwkt_schedule(nlp->lwp_thread);
2263 		} else {
2264 			/*
2265 			 * Leave the thread on our run queue.  Another
2266 			 * scheduler will try to pull it later.
2267 			 */
2268 			spin_unlock(&dd->spin);
2269 		}
2270 	} else {
2271 		/*
2272 		 * devoid of runnable threads and not allowed to steal
2273 		 * any.
2274 		 */
2275 		spin_unlock(&dd->spin);
2276 	}
2277 
2278 	/*
2279 	 * We're descheduled unless someone scheduled us.  Switch away.
2280 	 * Exiting the critical section will cause splz() to be called
2281 	 * for us if interrupts and such are pending.
2282 	 */
2283 	crit_exit_gd(gd);
2284 	tsleep(dd->helper_thread, PINTERLOCKED, "schslp", 0);
2285     }
2286 }
2287 
2288 #if 0
2289 static int
2290 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2291 {
2292 	int error, new_val;
2293 
2294 	new_val = usched_dfly_stick_to_level;
2295 
2296 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2297         if (error != 0 || req->newptr == NULL)
2298 		return (error);
2299 	if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2300 		return (EINVAL);
2301 	usched_dfly_stick_to_level = new_val;
2302 	return (0);
2303 }
2304 #endif
2305 
2306 /*
2307  * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2308  * Note that curprocmask bit 0 has already been cleared by rqinit() and
2309  * we should not mess with it further.
2310  */
2311 static void
2312 usched_dfly_cpu_init(void)
2313 {
2314 	int i;
2315 	int j;
2316 	int smt_not_supported = 0;
2317 	int cache_coherent_not_supported = 0;
2318 
2319 	if (bootverbose)
2320 		kprintf("Start usched_dfly helpers on cpus:\n");
2321 
2322 	sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2323 	usched_dfly_sysctl_tree =
2324 		SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2325 				SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2326 				"usched_dfly", CTLFLAG_RD, 0, "");
2327 
2328 	for (i = 0; i < ncpus; ++i) {
2329 		dfly_pcpu_t dd = &dfly_pcpu[i];
2330 		cpumask_t mask;
2331 
2332 		CPUMASK_ASSBIT(mask, i);
2333 		if (CPUMASK_TESTMASK(mask, smp_active_mask) == 0)
2334 		    continue;
2335 
2336 		spin_init(&dd->spin, "uschedcpuinit");
2337 		dd->cpunode = get_cpu_node_by_cpuid(i);
2338 		dd->cpuid = i;
2339 		CPUMASK_ASSBIT(dd->cpumask, i);
2340 		for (j = 0; j < NQS; j++) {
2341 			TAILQ_INIT(&dd->queues[j]);
2342 			TAILQ_INIT(&dd->rtqueues[j]);
2343 			TAILQ_INIT(&dd->idqueues[j]);
2344 		}
2345 		ATOMIC_CPUMASK_NANDBIT(dfly_curprocmask, 0);
2346 
2347 		if (dd->cpunode == NULL) {
2348 			smt_not_supported = 1;
2349 			cache_coherent_not_supported = 1;
2350 			if (bootverbose)
2351 				kprintf ("    cpu%d - WARNING: No CPU NODE "
2352 					 "found for cpu\n", i);
2353 		} else {
2354 			switch (dd->cpunode->type) {
2355 			case THREAD_LEVEL:
2356 				if (bootverbose)
2357 					kprintf ("    cpu%d - HyperThreading "
2358 						 "available. Core siblings: ",
2359 						 i);
2360 				break;
2361 			case CORE_LEVEL:
2362 				smt_not_supported = 1;
2363 
2364 				if (bootverbose)
2365 					kprintf ("    cpu%d - No HT available, "
2366 						 "multi-core/physical "
2367 						 "cpu. Physical siblings: ",
2368 						 i);
2369 				break;
2370 			case CHIP_LEVEL:
2371 				smt_not_supported = 1;
2372 
2373 				if (bootverbose)
2374 					kprintf ("    cpu%d - No HT available, "
2375 						 "single-core/physical cpu. "
2376 						 "Package siblings: ",
2377 						 i);
2378 				break;
2379 			default:
2380 				/* Let's go for safe defaults here */
2381 				smt_not_supported = 1;
2382 				cache_coherent_not_supported = 1;
2383 				if (bootverbose)
2384 					kprintf ("    cpu%d - Unknown cpunode->"
2385 						 "type=%u. siblings: ",
2386 						 i,
2387 						 (u_int)dd->cpunode->type);
2388 				break;
2389 			}
2390 
2391 			if (bootverbose) {
2392 				if (dd->cpunode->parent_node != NULL) {
2393 					kprint_cpuset(&dd->cpunode->
2394 							parent_node->members);
2395 					kprintf("\n");
2396 				} else {
2397 					kprintf(" no siblings\n");
2398 				}
2399 			}
2400 		}
2401 
2402 		lwkt_create(dfly_helper_thread, NULL, &dd->helper_thread, NULL,
2403 			    0, i, "usched %d", i);
2404 
2405 		/*
2406 		 * Allow user scheduling on the target cpu.  cpu #0 has already
2407 		 * been enabled in rqinit().
2408 		 */
2409 		if (i)
2410 			ATOMIC_CPUMASK_NANDMASK(dfly_curprocmask, mask);
2411 		ATOMIC_CPUMASK_ORMASK(dfly_rdyprocmask, mask);
2412 		dd->upri = PRIBASE_NULL;
2413 
2414 	}
2415 
2416 	/* usched_dfly sysctl configurable parameters */
2417 
2418 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2419 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2420 		       OID_AUTO, "rrinterval", CTLFLAG_RW,
2421 		       &usched_dfly_rrinterval, 0, "");
2422 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2423 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2424 		       OID_AUTO, "decay", CTLFLAG_RW,
2425 		       &usched_dfly_decay, 0, "Extra decay when not running");
2426 
2427 	/* Add enable/disable option for SMT scheduling if supported */
2428 	if (smt_not_supported) {
2429 		usched_dfly_smt = 0;
2430 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2431 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2432 				  OID_AUTO, "smt", CTLFLAG_RD,
2433 				  "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2434 	} else {
2435 		usched_dfly_smt = 1;
2436 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2437 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2438 			       OID_AUTO, "smt", CTLFLAG_RW,
2439 			       &usched_dfly_smt, 0, "Enable SMT scheduling");
2440 	}
2441 
2442 	/*
2443 	 * Add enable/disable option for cache coherent scheduling
2444 	 * if supported
2445 	 */
2446 	if (cache_coherent_not_supported) {
2447 		usched_dfly_cache_coherent = 0;
2448 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2449 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2450 				  OID_AUTO, "cache_coherent", CTLFLAG_RD,
2451 				  "NOT SUPPORTED", 0,
2452 				  "Cache coherence NOT SUPPORTED");
2453 	} else {
2454 		usched_dfly_cache_coherent = 1;
2455 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2456 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2457 			       OID_AUTO, "cache_coherent", CTLFLAG_RW,
2458 			       &usched_dfly_cache_coherent, 0,
2459 			       "Enable/Disable cache coherent scheduling");
2460 
2461 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2462 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2463 			       OID_AUTO, "weight1", CTLFLAG_RW,
2464 			       &usched_dfly_weight1, 200,
2465 			       "Weight selection for current cpu");
2466 
2467 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2468 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2469 			       OID_AUTO, "weight2", CTLFLAG_RW,
2470 			       &usched_dfly_weight2, 180,
2471 			       "Weight selection for wakefrom cpu");
2472 
2473 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2474 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2475 			       OID_AUTO, "weight3", CTLFLAG_RW,
2476 			       &usched_dfly_weight3, 40,
2477 			       "Weight selection for num threads on queue");
2478 
2479 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2480 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2481 			       OID_AUTO, "weight4", CTLFLAG_RW,
2482 			       &usched_dfly_weight4, 160,
2483 			       "Availability of other idle cpus");
2484 
2485 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2486 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2487 			       OID_AUTO, "fast_resched", CTLFLAG_RW,
2488 			       &usched_dfly_fast_resched, 0,
2489 			       "Availability of other idle cpus");
2490 
2491 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2492 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2493 			       OID_AUTO, "features", CTLFLAG_RW,
2494 			       &usched_dfly_features, 0x8F,
2495 			       "Allow pulls into empty queues");
2496 
2497 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2498 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2499 			       OID_AUTO, "swmask", CTLFLAG_RW,
2500 			       &usched_dfly_swmask, ~PPQMASK,
2501 			       "Queue mask to force thread switch");
2502 
2503 #if 0
2504 		SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2505 				SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2506 				OID_AUTO, "stick_to_level",
2507 				CTLTYPE_INT | CTLFLAG_RW,
2508 				NULL, sizeof usched_dfly_stick_to_level,
2509 				sysctl_usched_dfly_stick_to_level, "I",
2510 				"Stick a process to this level. See sysctl"
2511 				"paremter hw.cpu_topology.level_description");
2512 #endif
2513 	}
2514 }
2515 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2516 	usched_dfly_cpu_init, NULL);
2517