xref: /dflybsd-src/sys/kern/usched_dfly.c (revision f0140465f072dacd8b485e76b254a999329a1d3c)
1 /*
2  * Copyright (c) 2012 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 #include <sys/mplock2.h>
52 
53 #include <sys/ktr.h>
54 
55 #include <machine/cpu.h>
56 #include <machine/smp.h>
57 
58 /*
59  * Priorities.  Note that with 32 run queues per scheduler each queue
60  * represents four priority levels.
61  */
62 
63 int dfly_rebalanced;
64 
65 #define MAXPRI			128
66 #define PRIMASK			(MAXPRI - 1)
67 #define PRIBASE_REALTIME	0
68 #define PRIBASE_NORMAL		MAXPRI
69 #define PRIBASE_IDLE		(MAXPRI * 2)
70 #define PRIBASE_THREAD		(MAXPRI * 3)
71 #define PRIBASE_NULL		(MAXPRI * 4)
72 
73 #define NQS	32			/* 32 run queues. */
74 #define PPQ	(MAXPRI / NQS)		/* priorities per queue */
75 #define PPQMASK	(PPQ - 1)
76 
77 /*
78  * NICEPPQ	- number of nice units per priority queue
79  * ESTCPUPPQ	- number of estcpu units per priority queue
80  * ESTCPUMAX	- number of estcpu units
81  */
82 #define NICEPPQ		2
83 #define ESTCPUPPQ	512
84 #define ESTCPUMAX	(ESTCPUPPQ * NQS)
85 #define BATCHMAX	(ESTCPUFREQ * 30)
86 #define PRIO_RANGE	(PRIO_MAX - PRIO_MIN + 1)
87 
88 #define ESTCPULIM(v)	min((v), ESTCPUMAX)
89 
90 TAILQ_HEAD(rq, lwp);
91 
92 #define lwp_priority	lwp_usdata.dfly.priority
93 #define lwp_forked	lwp_usdata.dfly.forked
94 #define lwp_rqindex	lwp_usdata.dfly.rqindex
95 #define lwp_estcpu	lwp_usdata.dfly.estcpu
96 #define lwp_estfast	lwp_usdata.dfly.estfast
97 #define lwp_uload	lwp_usdata.dfly.uload
98 #define lwp_rqtype	lwp_usdata.dfly.rqtype
99 #define lwp_qcpu	lwp_usdata.dfly.qcpu
100 #define lwp_rrcount	lwp_usdata.dfly.rrcount
101 
102 struct usched_dfly_pcpu {
103 	struct spinlock spin;
104 	struct thread	helper_thread;
105 	short		unusde01;
106 	short		upri;
107 	int		uload;
108 	int		ucount;
109 	struct lwp	*uschedcp;
110 	struct rq	queues[NQS];
111 	struct rq	rtqueues[NQS];
112 	struct rq	idqueues[NQS];
113 	u_int32_t	queuebits;
114 	u_int32_t	rtqueuebits;
115 	u_int32_t	idqueuebits;
116 	int		runqcount;
117 	int		cpuid;
118 	cpumask_t	cpumask;
119 #ifdef SMP
120 	cpu_node_t	*cpunode;
121 #endif
122 };
123 
124 typedef struct usched_dfly_pcpu	*dfly_pcpu_t;
125 
126 static void dfly_acquire_curproc(struct lwp *lp);
127 static void dfly_release_curproc(struct lwp *lp);
128 static void dfly_select_curproc(globaldata_t gd);
129 static void dfly_setrunqueue(struct lwp *lp);
130 static void dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp);
131 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period,
132 				sysclock_t cpstamp);
133 static void dfly_recalculate_estcpu(struct lwp *lp);
134 static void dfly_resetpriority(struct lwp *lp);
135 static void dfly_forking(struct lwp *plp, struct lwp *lp);
136 static void dfly_exiting(struct lwp *lp, struct proc *);
137 static void dfly_uload_update(struct lwp *lp);
138 static void dfly_yield(struct lwp *lp);
139 #ifdef SMP
140 static void dfly_changeqcpu_locked(struct lwp *lp,
141 				dfly_pcpu_t dd, dfly_pcpu_t rdd);
142 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
143 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
144 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
145 #endif
146 
147 #ifdef SMP
148 static void dfly_need_user_resched_remote(void *dummy);
149 #endif
150 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
151 					  struct lwp *chklp, int worst);
152 static void dfly_remrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
153 static void dfly_setrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
154 
155 struct usched usched_dfly = {
156 	{ NULL },
157 	"dfly", "Original DragonFly Scheduler",
158 	NULL,			/* default registration */
159 	NULL,			/* default deregistration */
160 	dfly_acquire_curproc,
161 	dfly_release_curproc,
162 	dfly_setrunqueue,
163 	dfly_schedulerclock,
164 	dfly_recalculate_estcpu,
165 	dfly_resetpriority,
166 	dfly_forking,
167 	dfly_exiting,
168 	dfly_uload_update,
169 	NULL,			/* setcpumask not supported */
170 	dfly_yield
171 };
172 
173 /*
174  * We have NQS (32) run queues per scheduling class.  For the normal
175  * class, there are 128 priorities scaled onto these 32 queues.  New
176  * processes are added to the last entry in each queue, and processes
177  * are selected for running by taking them from the head and maintaining
178  * a simple FIFO arrangement.  Realtime and Idle priority processes have
179  * and explicit 0-31 priority which maps directly onto their class queue
180  * index.  When a queue has something in it, the corresponding bit is
181  * set in the queuebits variable, allowing a single read to determine
182  * the state of all 32 queues and then a ffs() to find the first busy
183  * queue.
184  */
185 static cpumask_t dfly_curprocmask = -1;	/* currently running a user process */
186 static cpumask_t dfly_rdyprocmask;	/* ready to accept a user process */
187 #ifdef SMP
188 static volatile int dfly_scancpu;
189 #endif
190 static volatile int dfly_ucount;	/* total running on whole system */
191 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
192 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
193 static struct sysctl_oid *usched_dfly_sysctl_tree;
194 
195 /* Debug info exposed through debug.* sysctl */
196 
197 static int usched_dfly_debug = -1;
198 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
199 	   &usched_dfly_debug, 0,
200 	   "Print debug information for this pid");
201 
202 static int usched_dfly_pid_debug = -1;
203 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
204 	   &usched_dfly_pid_debug, 0,
205 	   "Print KTR debug information for this pid");
206 
207 static int usched_dfly_chooser = 0;
208 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
209 	   &usched_dfly_chooser, 0,
210 	   "Print KTR debug information for this pid");
211 
212 /*
213  * Tunning usched_dfly - configurable through kern.usched_dfly.
214  *
215  * weight1 - Tries to keep threads on their current cpu.  If you
216  *	     make this value too large the scheduler will not be
217  *	     able to load-balance large loads.
218  *
219  * weight2 - If non-zero, detects thread pairs undergoing synchronous
220  *	     communications and tries to move them closer together.
221  *	     Behavior is adjusted by bit 4 of features (0x10).
222  *
223  *	     WARNING!  Weight2 is a ridiculously sensitive parameter,
224  *	     a small value is recommended.
225  *
226  * weight3 - Weighting based on the number of recently runnable threads
227  *	     on the userland scheduling queue (ignoring their loads).
228  *	     A nominal value here prevents high-priority (low-load)
229  *	     threads from accumulating on one cpu core when other
230  *	     cores are available.
231  *
232  *	     This value should be left fairly small relative to weight1
233  *	     and weight4.
234  *
235  * weight4 - Weighting based on other cpu queues being available
236  *	     or running processes with higher lwp_priority's.
237  *
238  *	     This allows a thread to migrate to another nearby cpu if it
239  *	     is unable to run on the current cpu based on the other cpu
240  *	     being idle or running a lower priority (higher lwp_priority)
241  *	     thread.  This value should be large enough to override weight1
242  *
243  * features - These flags can be set or cleared to enable or disable various
244  *	      features.
245  *
246  *	      0x01	Enable idle-cpu pulling			(default)
247  *	      0x02	Enable proactive pushing		(default)
248  *	      0x04	Enable rebalancing rover		(default)
249  *	      0x08	Enable more proactive pushing		(default)
250  *	      0x10	(flip weight2 limit on same cpu)	(default)
251  *	      0x20	choose best cpu for forked process
252  *	      0x40	choose current cpu for forked process
253  *	      0x80	choose random cpu for forked process	(default)
254  */
255 #ifdef SMP
256 static int usched_dfly_smt = 0;
257 static int usched_dfly_cache_coherent = 0;
258 static int usched_dfly_weight1 = 200;	/* keep thread on current cpu */
259 static int usched_dfly_weight2 = 180;	/* synchronous peer's current cpu */
260 static int usched_dfly_weight3 = 40;	/* number of threads on queue */
261 static int usched_dfly_weight4 = 160;	/* availability of idle cores */
262 static int usched_dfly_fast_resched = 0;/* delta priority / resched */
263 static int usched_dfly_features = 0x8F;	/* allow pulls */
264 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */
265 #endif
266 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
267 static int usched_dfly_decay = 8;
268 
269 /* KTR debug printings */
270 
271 KTR_INFO_MASTER(usched);
272 
273 #if !defined(KTR_USCHED_DFLY)
274 #define	KTR_USCHED_DFLY	KTR_ALL
275 #endif
276 
277 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
278     "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
279     pid_t pid, int old_cpuid, int curr);
280 
281 /*
282  * This function is called when the kernel intends to return to userland.
283  * It is responsible for making the thread the current designated userland
284  * thread for this cpu, blocking if necessary.
285  *
286  * The kernel will not depress our LWKT priority until after we return,
287  * in case we have to shove over to another cpu.
288  *
289  * We must determine our thread's disposition before we switch away.  This
290  * is very sensitive code.
291  *
292  * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
293  * TO ANOTHER CPU!  Because most of the kernel assumes that no migration will
294  * occur, this function is called only under very controlled circumstances.
295  */
296 static void
297 dfly_acquire_curproc(struct lwp *lp)
298 {
299 	globaldata_t gd;
300 	dfly_pcpu_t dd;
301 #ifdef SMP
302 	dfly_pcpu_t rdd;
303 #endif
304 	thread_t td;
305 	int force_resched;
306 
307 	/*
308 	 * Make sure we aren't sitting on a tsleep queue.
309 	 */
310 	td = lp->lwp_thread;
311 	crit_enter_quick(td);
312 	if (td->td_flags & TDF_TSLEEPQ)
313 		tsleep_remove(td);
314 	dfly_recalculate_estcpu(lp);
315 
316 	gd = mycpu;
317 	dd = &dfly_pcpu[gd->gd_cpuid];
318 
319 	/*
320 	 * Process any pending interrupts/ipi's, then handle reschedule
321 	 * requests.  dfly_release_curproc() will try to assign a new
322 	 * uschedcp that isn't us and otherwise NULL it out.
323 	 */
324 	force_resched = 0;
325 	if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
326 	    lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
327 		force_resched = 1;
328 	}
329 
330 	if (user_resched_wanted()) {
331 		if (dd->uschedcp == lp)
332 			force_resched = 1;
333 		clear_user_resched();
334 		dfly_release_curproc(lp);
335 	}
336 
337 	/*
338 	 * Loop until we are the current user thread.
339 	 *
340 	 * NOTE: dd spinlock not held at top of loop.
341 	 */
342 	if (dd->uschedcp == lp)
343 		lwkt_yield_quick();
344 
345 	while (dd->uschedcp != lp) {
346 		lwkt_yield_quick();
347 
348 		spin_lock(&dd->spin);
349 
350 		/*
351 		 * We are not or are no longer the current lwp and a forced
352 		 * reschedule was requested.  Figure out the best cpu to
353 		 * run on (our current cpu will be given significant weight).
354 		 *
355 		 * (if a reschedule was not requested we want to move this
356 		 *  step after the uschedcp tests).
357 		 */
358 #ifdef SMP
359 		if (force_resched &&
360 		    (usched_dfly_features & 0x08) &&
361 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
362 			dfly_changeqcpu_locked(lp, dd, rdd);
363 			spin_unlock(&dd->spin);
364 			lwkt_deschedule(lp->lwp_thread);
365 			dfly_setrunqueue_dd(rdd, lp);
366 			lwkt_switch();
367 			gd = mycpu;
368 			dd = &dfly_pcpu[gd->gd_cpuid];
369 			continue;
370 		}
371 #endif
372 
373 		/*
374 		 * Either no reschedule was requested or the best queue was
375 		 * dd, and no current process has been selected.  We can
376 		 * trivially become the current lwp on the current cpu.
377 		 */
378 		if (dd->uschedcp == NULL) {
379 			atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
380 			dd->uschedcp = lp;
381 			dd->upri = lp->lwp_priority;
382 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
383 			spin_unlock(&dd->spin);
384 			break;
385 		}
386 
387 		/*
388 		 * Can we steal the current designated user thread?
389 		 *
390 		 * If we do the other thread will stall when it tries to
391 		 * return to userland, possibly rescheduling elsewhere.
392 		 *
393 		 * It is important to do a masked test to avoid the edge
394 		 * case where two near-equal-priority threads are constantly
395 		 * interrupting each other.
396 		 *
397 		 * In the exact match case another thread has already gained
398 		 * uschedcp and lowered its priority, if we steal it the
399 		 * other thread will stay stuck on the LWKT runq and not
400 		 * push to another cpu.  So don't steal on equal-priority even
401 		 * though it might appear to be more beneficial due to not
402 		 * having to switch back to the other thread's context.
403 		 *
404 		 * usched_dfly_fast_resched requires that two threads be
405 		 * significantly far apart in priority in order to interrupt.
406 		 *
407 		 * If better but not sufficiently far apart, the current
408 		 * uschedcp will be interrupted at the next scheduler clock.
409 		 */
410 		if (dd->uschedcp &&
411 		   (dd->upri & ~PPQMASK) >
412 		   (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
413 			dd->uschedcp = lp;
414 			dd->upri = lp->lwp_priority;
415 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
416 			spin_unlock(&dd->spin);
417 			break;
418 		}
419 #ifdef SMP
420 		/*
421 		 * We are not the current lwp, figure out the best cpu
422 		 * to run on (our current cpu will be given significant
423 		 * weight).  Loop on cpu change.
424 		 */
425 		if ((usched_dfly_features & 0x02) &&
426 		    force_resched == 0 &&
427 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
428 			dfly_changeqcpu_locked(lp, dd, rdd);
429 			spin_unlock(&dd->spin);
430 			lwkt_deschedule(lp->lwp_thread);
431 			dfly_setrunqueue_dd(rdd, lp);
432 			lwkt_switch();
433 			gd = mycpu;
434 			dd = &dfly_pcpu[gd->gd_cpuid];
435 			continue;
436 		}
437 #endif
438 
439 		/*
440 		 * We cannot become the current lwp, place the lp on the
441 		 * run-queue of this or another cpu and deschedule ourselves.
442 		 *
443 		 * When we are reactivated we will have another chance.
444 		 *
445 		 * Reload after a switch or setrunqueue/switch possibly
446 		 * moved us to another cpu.
447 		 */
448 		spin_unlock(&dd->spin);
449 		lwkt_deschedule(lp->lwp_thread);
450 		dfly_setrunqueue_dd(dd, lp);
451 		lwkt_switch();
452 		gd = mycpu;
453 		dd = &dfly_pcpu[gd->gd_cpuid];
454 	}
455 
456 	/*
457 	 * Make sure upri is synchronized, then yield to LWKT threads as
458 	 * needed before returning.  This could result in another reschedule.
459 	 * XXX
460 	 */
461 	crit_exit_quick(td);
462 
463 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
464 }
465 
466 /*
467  * DFLY_RELEASE_CURPROC
468  *
469  * This routine detaches the current thread from the userland scheduler,
470  * usually because the thread needs to run or block in the kernel (at
471  * kernel priority) for a while.
472  *
473  * This routine is also responsible for selecting a new thread to
474  * make the current thread.
475  *
476  * NOTE: This implementation differs from the dummy example in that
477  * dfly_select_curproc() is able to select the current process, whereas
478  * dummy_select_curproc() is not able to select the current process.
479  * This means we have to NULL out uschedcp.
480  *
481  * Additionally, note that we may already be on a run queue if releasing
482  * via the lwkt_switch() in dfly_setrunqueue().
483  */
484 static void
485 dfly_release_curproc(struct lwp *lp)
486 {
487 	globaldata_t gd = mycpu;
488 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
489 
490 	/*
491 	 * Make sure td_wakefromcpu is defaulted.  This will be overwritten
492 	 * by wakeup().
493 	 */
494 	if (dd->uschedcp == lp) {
495 		KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
496 		spin_lock(&dd->spin);
497 		if (dd->uschedcp == lp) {
498 			dd->uschedcp = NULL;	/* don't let lp be selected */
499 			dd->upri = PRIBASE_NULL;
500 			atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
501 			spin_unlock(&dd->spin);
502 			dfly_select_curproc(gd);
503 		} else {
504 			spin_unlock(&dd->spin);
505 		}
506 	}
507 }
508 
509 /*
510  * DFLY_SELECT_CURPROC
511  *
512  * Select a new current process for this cpu and clear any pending user
513  * reschedule request.  The cpu currently has no current process.
514  *
515  * This routine is also responsible for equal-priority round-robining,
516  * typically triggered from dfly_schedulerclock().  In our dummy example
517  * all the 'user' threads are LWKT scheduled all at once and we just
518  * call lwkt_switch().
519  *
520  * The calling process is not on the queue and cannot be selected.
521  */
522 static
523 void
524 dfly_select_curproc(globaldata_t gd)
525 {
526 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
527 	struct lwp *nlp;
528 	int cpuid = gd->gd_cpuid;
529 
530 	crit_enter_gd(gd);
531 
532 	spin_lock(&dd->spin);
533 	nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
534 
535 	if (nlp) {
536 		atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
537 		dd->upri = nlp->lwp_priority;
538 		dd->uschedcp = nlp;
539 #if 0
540 		dd->rrcount = 0;		/* reset round robin */
541 #endif
542 		spin_unlock(&dd->spin);
543 #ifdef SMP
544 		lwkt_acquire(nlp->lwp_thread);
545 #endif
546 		lwkt_schedule(nlp->lwp_thread);
547 	} else {
548 		spin_unlock(&dd->spin);
549 	}
550 	crit_exit_gd(gd);
551 }
552 
553 /*
554  * Place the specified lwp on the user scheduler's run queue.  This routine
555  * must be called with the thread descheduled.  The lwp must be runnable.
556  * It must not be possible for anyone else to explicitly schedule this thread.
557  *
558  * The thread may be the current thread as a special case.
559  */
560 static void
561 dfly_setrunqueue(struct lwp *lp)
562 {
563 	dfly_pcpu_t dd;
564 	dfly_pcpu_t rdd;
565 
566 	/*
567 	 * First validate the process LWKT state.
568 	 */
569 	KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
570 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
571 	    ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
572 	     lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
573 	KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
574 
575 	/*
576 	 * NOTE: dd/rdd do not necessarily represent the current cpu.
577 	 *	 Instead they may represent the cpu the thread was last
578 	 *	 scheduled on or inherited by its parent.
579 	 */
580 	dd = &dfly_pcpu[lp->lwp_qcpu];
581 	rdd = dd;
582 
583 	/*
584 	 * This process is not supposed to be scheduled anywhere or assigned
585 	 * as the current process anywhere.  Assert the condition.
586 	 */
587 	KKASSERT(rdd->uschedcp != lp);
588 
589 #ifndef SMP
590 	/*
591 	 * If we are not SMP we do not have a scheduler helper to kick
592 	 * and must directly activate the process if none are scheduled.
593 	 *
594 	 * This is really only an issue when bootstrapping init since
595 	 * the caller in all other cases will be a user process, and
596 	 * even if released (rdd->uschedcp == NULL), that process will
597 	 * kickstart the scheduler when it returns to user mode from
598 	 * the kernel.
599 	 *
600 	 * NOTE: On SMP we can't just set some other cpu's uschedcp.
601 	 */
602 	if (rdd->uschedcp == NULL) {
603 		spin_lock(&rdd->spin);
604 		if (rdd->uschedcp == NULL) {
605 			atomic_set_cpumask(&dfly_curprocmask, 1);
606 			rdd->uschedcp = lp;
607 			rdd->upri = lp->lwp_priority;
608 			spin_unlock(&rdd->spin);
609 			lwkt_schedule(lp->lwp_thread);
610 			return;
611 		}
612 		spin_unlock(&rdd->spin);
613 	}
614 #endif
615 
616 #ifdef SMP
617 	/*
618 	 * Ok, we have to setrunqueue some target cpu and request a reschedule
619 	 * if necessary.
620 	 *
621 	 * We have to choose the best target cpu.  It might not be the current
622 	 * target even if the current cpu has no running user thread (for
623 	 * example, because the current cpu might be a hyperthread and its
624 	 * sibling has a thread assigned).
625 	 *
626 	 * If we just forked it is most optimal to run the child on the same
627 	 * cpu just in case the parent decides to wait for it (thus getting
628 	 * off that cpu).  As long as there is nothing else runnable on the
629 	 * cpu, that is.  If we did this unconditionally a parent forking
630 	 * multiple children before waiting (e.g. make -j N) leaves other
631 	 * cpus idle that could be working.
632 	 */
633 	if (lp->lwp_forked) {
634 		lp->lwp_forked = 0;
635 		if (usched_dfly_features & 0x20)
636 			rdd = dfly_choose_best_queue(lp);
637 		else if (usched_dfly_features & 0x40)
638 			rdd = &dfly_pcpu[lp->lwp_qcpu];
639 		else if (usched_dfly_features & 0x80)
640 			rdd = dfly_choose_queue_simple(rdd, lp);
641 		else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
642 			rdd = dfly_choose_best_queue(lp);
643 		else
644 			rdd = &dfly_pcpu[lp->lwp_qcpu];
645 	} else {
646 		rdd = dfly_choose_best_queue(lp);
647 		/* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
648 	}
649 	if (lp->lwp_qcpu != rdd->cpuid) {
650 		spin_lock(&dd->spin);
651 		dfly_changeqcpu_locked(lp, dd, rdd);
652 		spin_unlock(&dd->spin);
653 	}
654 #endif
655 	dfly_setrunqueue_dd(rdd, lp);
656 }
657 
658 #ifdef SMP
659 
660 /*
661  * Change qcpu to rdd->cpuid.  The dd the lp is CURRENTLY on must be
662  * spin-locked on-call.  rdd does not have to be.
663  */
664 static void
665 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
666 {
667 	if (lp->lwp_qcpu != rdd->cpuid) {
668 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
669 			atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
670 			atomic_add_int(&dd->uload, -lp->lwp_uload);
671 			atomic_add_int(&dd->ucount, -1);
672 			atomic_add_int(&dfly_ucount, -1);
673 		}
674 		lp->lwp_qcpu = rdd->cpuid;
675 	}
676 }
677 
678 #endif
679 
680 /*
681  * Place lp on rdd's runqueue.  Nothing is locked on call.  This function
682  * also performs all necessary ancillary notification actions.
683  */
684 static void
685 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
686 {
687 #ifdef SMP
688 	globaldata_t rgd;
689 
690 	/*
691 	 * We might be moving the lp to another cpu's run queue, and once
692 	 * on the runqueue (even if it is our cpu's), another cpu can rip
693 	 * it away from us.
694 	 *
695 	 * TDF_MIGRATING might already be set if this is part of a
696 	 * remrunqueue+setrunqueue sequence.
697 	 */
698 	if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
699 		lwkt_giveaway(lp->lwp_thread);
700 
701 	rgd = globaldata_find(rdd->cpuid);
702 
703 	/*
704 	 * We lose control of the lp the moment we release the spinlock
705 	 * after having placed it on the queue.  i.e. another cpu could pick
706 	 * it up, or it could exit, or its priority could be further
707 	 * adjusted, or something like that.
708 	 *
709 	 * WARNING! rdd can point to a foreign cpu!
710 	 */
711 	spin_lock(&rdd->spin);
712 	dfly_setrunqueue_locked(rdd, lp);
713 
714 	/*
715 	 * Potentially interrupt the currently-running thread
716 	 */
717 	if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
718 		/*
719 		 * Currently running thread is better or same, do not
720 		 * interrupt.
721 		 */
722 		spin_unlock(&rdd->spin);
723 	} else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
724 		   usched_dfly_fast_resched) {
725 		/*
726 		 * Currently running thread is not better, but not so bad
727 		 * that we need to interrupt it.  Let it run for one more
728 		 * scheduler tick.
729 		 */
730 		if (rdd->uschedcp &&
731 		    rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
732 			rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
733 		}
734 		spin_unlock(&rdd->spin);
735 	} else if (rgd == mycpu) {
736 		/*
737 		 * We should interrupt the currently running thread, which
738 		 * is on the current cpu.
739 		 */
740 		spin_unlock(&rdd->spin);
741 		if (rdd->uschedcp == NULL) {
742 			wakeup_mycpu(&rdd->helper_thread); /* XXX */
743 			need_user_resched();
744 		} else {
745 			need_user_resched();
746 		}
747 	} else {
748 		/*
749 		 * We should interrupt the currently running thread, which
750 		 * is on a different cpu.
751 		 */
752 		spin_unlock(&rdd->spin);
753 		lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
754 	}
755 #else
756 	/*
757 	 * Request a reschedule if appropriate.
758 	 */
759 	spin_lock(&rdd->spin);
760 	dfly_setrunqueue_locked(rdd, lp);
761 	spin_unlock(&rdd->spin);
762 	if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
763 		need_user_resched();
764 	}
765 #endif
766 }
767 
768 /*
769  * This routine is called from a systimer IPI.  It MUST be MP-safe and
770  * the BGL IS NOT HELD ON ENTRY.  This routine is called at ESTCPUFREQ on
771  * each cpu.
772  */
773 static
774 void
775 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
776 {
777 	globaldata_t gd = mycpu;
778 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
779 
780 	/*
781 	 * Spinlocks also hold a critical section so there should not be
782 	 * any active.
783 	 */
784 	KKASSERT(gd->gd_spinlocks == 0);
785 
786 	if (lp == NULL)
787 		return;
788 
789 	/*
790 	 * Do we need to round-robin?  We round-robin 10 times a second.
791 	 * This should only occur for cpu-bound batch processes.
792 	 */
793 	if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
794 		lp->lwp_thread->td_wakefromcpu = -1;
795 		need_user_resched();
796 	}
797 
798 	/*
799 	 * Adjust estcpu upward using a real time equivalent calculation,
800 	 * and recalculate lp's priority.
801 	 */
802 	lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
803 	dfly_resetpriority(lp);
804 
805 	/*
806 	 * Rebalance two cpus every 8 ticks, pulling the worst thread
807 	 * from the worst cpu's queue into a rotating cpu number.
808 	 *
809 	 * This mechanic is needed because the push algorithms can
810 	 * steady-state in an non-optimal configuration.  We need to mix it
811 	 * up a little, even if it means breaking up a paired thread, so
812 	 * the push algorithms can rebalance the degenerate conditions.
813 	 * This portion of the algorithm exists to ensure stability at the
814 	 * selected weightings.
815 	 *
816 	 * Because we might be breaking up optimal conditions we do not want
817 	 * to execute this too quickly, hence we only rebalance approximately
818 	 * ~7-8 times per second.  The push's, on the otherhand, are capable
819 	 * moving threads to other cpus at a much higher rate.
820 	 *
821 	 * We choose the most heavily loaded thread from the worst queue
822 	 * in order to ensure that multiple heavy-weight threads on the same
823 	 * queue get broken up, and also because these threads are the most
824 	 * likely to be able to remain in place.  Hopefully then any pairings,
825 	 * if applicable, migrate to where these threads are.
826 	 */
827 #ifdef SMP
828 	if ((usched_dfly_features & 0x04) &&
829 	    ((u_int)sched_ticks & 7) == 0 &&
830 	    (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
831 		/*
832 		 * Our cpu is up.
833 		 */
834 		struct lwp *nlp;
835 		dfly_pcpu_t rdd;
836 
837 		rdd = dfly_choose_worst_queue(dd);
838 		if (rdd) {
839 			spin_lock(&dd->spin);
840 			if (spin_trylock(&rdd->spin)) {
841 				nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
842 				spin_unlock(&rdd->spin);
843 				if (nlp == NULL)
844 					spin_unlock(&dd->spin);
845 			} else {
846 				spin_unlock(&dd->spin);
847 				nlp = NULL;
848 			}
849 		} else {
850 			nlp = NULL;
851 		}
852 		/* dd->spin held if nlp != NULL */
853 
854 		/*
855 		 * Either schedule it or add it to our queue.
856 		 */
857 		if (nlp &&
858 		    (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
859 			atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
860 			dd->upri = nlp->lwp_priority;
861 			dd->uschedcp = nlp;
862 #if 0
863 			dd->rrcount = 0;	/* reset round robin */
864 #endif
865 			spin_unlock(&dd->spin);
866 			lwkt_acquire(nlp->lwp_thread);
867 			lwkt_schedule(nlp->lwp_thread);
868 		} else if (nlp) {
869 			dfly_setrunqueue_locked(dd, nlp);
870 			spin_unlock(&dd->spin);
871 		}
872 	}
873 #endif
874 }
875 
876 /*
877  * Called from acquire and from kern_synch's one-second timer (one of the
878  * callout helper threads) with a critical section held.
879  *
880  * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
881  * overall system load.
882  *
883  * Note that no recalculation occurs for a process which sleeps and wakes
884  * up in the same tick.  That is, a system doing thousands of context
885  * switches per second will still only do serious estcpu calculations
886  * ESTCPUFREQ times per second.
887  */
888 static
889 void
890 dfly_recalculate_estcpu(struct lwp *lp)
891 {
892 	globaldata_t gd = mycpu;
893 	sysclock_t cpbase;
894 	sysclock_t ttlticks;
895 	int estcpu;
896 	int decay_factor;
897 	int ucount;
898 
899 	/*
900 	 * We have to subtract periodic to get the last schedclock
901 	 * timeout time, otherwise we would get the upcoming timeout.
902 	 * Keep in mind that a process can migrate between cpus and
903 	 * while the scheduler clock should be very close, boundary
904 	 * conditions could lead to a small negative delta.
905 	 */
906 	cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
907 
908 	if (lp->lwp_slptime > 1) {
909 		/*
910 		 * Too much time has passed, do a coarse correction.
911 		 */
912 		lp->lwp_estcpu = lp->lwp_estcpu >> 1;
913 		dfly_resetpriority(lp);
914 		lp->lwp_cpbase = cpbase;
915 		lp->lwp_cpticks = 0;
916 		lp->lwp_estfast = 0;
917 	} else if (lp->lwp_cpbase != cpbase) {
918 		/*
919 		 * Adjust estcpu if we are in a different tick.  Don't waste
920 		 * time if we are in the same tick.
921 		 *
922 		 * First calculate the number of ticks in the measurement
923 		 * interval.  The ttlticks calculation can wind up 0 due to
924 		 * a bug in the handling of lwp_slptime  (as yet not found),
925 		 * so make sure we do not get a divide by 0 panic.
926 		 */
927 		ttlticks = (cpbase - lp->lwp_cpbase) /
928 			   gd->gd_schedclock.periodic;
929 		if (ttlticks < 0) {
930 			ttlticks = 0;
931 			lp->lwp_cpbase = cpbase;
932 		}
933 		if (ttlticks == 0)
934 			return;
935 		updatepcpu(lp, lp->lwp_cpticks, ttlticks);
936 
937 		/*
938 		 * Calculate the percentage of one cpu being used then
939 		 * compensate for any system load in excess of ncpus.
940 		 *
941 		 * For example, if we have 8 cores and 16 running cpu-bound
942 		 * processes then all things being equal each process will
943 		 * get 50% of one cpu.  We need to pump this value back
944 		 * up to 100% so the estcpu calculation properly adjusts
945 		 * the process's dynamic priority.
946 		 *
947 		 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
948 		 */
949 		estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
950 		ucount = dfly_ucount;
951 		if (ucount > ncpus) {
952 			estcpu += estcpu * (ucount - ncpus) / ncpus;
953 		}
954 
955 		if (usched_dfly_debug == lp->lwp_proc->p_pid) {
956 			kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
957 				lp->lwp_proc->p_pid, lp,
958 				estcpu, lp->lwp_estcpu,
959 				lp->lwp_cpticks, ttlticks);
960 		}
961 
962 		/*
963 		 * Adjust lp->lwp_esetcpu.  The decay factor determines how
964 		 * quickly lwp_estcpu collapses to its realtime calculation.
965 		 * A slower collapse gives us a more accurate number over
966 		 * the long term but can create problems with bursty threads
967 		 * or threads which become cpu hogs.
968 		 *
969 		 * To solve this problem, newly started lwps and lwps which
970 		 * are restarting after having been asleep for a while are
971 		 * given a much, much faster decay in order to quickly
972 		 * detect whether they become cpu-bound.
973 		 *
974 		 * NOTE: p_nice is accounted for in dfly_resetpriority(),
975 		 *	 and not here, but we must still ensure that a
976 		 *	 cpu-bound nice -20 process does not completely
977 		 *	 override a cpu-bound nice +20 process.
978 		 *
979 		 * NOTE: We must use ESTCPULIM() here to deal with any
980 		 *	 overshoot.
981 		 */
982 		decay_factor = usched_dfly_decay;
983 		if (decay_factor < 1)
984 			decay_factor = 1;
985 		if (decay_factor > 1024)
986 			decay_factor = 1024;
987 
988 		if (lp->lwp_estfast < usched_dfly_decay) {
989 			++lp->lwp_estfast;
990 			lp->lwp_estcpu = ESTCPULIM(
991 				(lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
992 				(lp->lwp_estfast + 1));
993 		} else {
994 			lp->lwp_estcpu = ESTCPULIM(
995 				(lp->lwp_estcpu * decay_factor + estcpu) /
996 				(decay_factor + 1));
997 		}
998 
999 		if (usched_dfly_debug == lp->lwp_proc->p_pid)
1000 			kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
1001 		dfly_resetpriority(lp);
1002 		lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
1003 		lp->lwp_cpticks = 0;
1004 	}
1005 }
1006 
1007 /*
1008  * Compute the priority of a process when running in user mode.
1009  * Arrange to reschedule if the resulting priority is better
1010  * than that of the current process.
1011  *
1012  * This routine may be called with any process.
1013  *
1014  * This routine is called by fork1() for initial setup with the process
1015  * of the run queue, and also may be called normally with the process on or
1016  * off the run queue.
1017  */
1018 static void
1019 dfly_resetpriority(struct lwp *lp)
1020 {
1021 	dfly_pcpu_t rdd;
1022 	int newpriority;
1023 	u_short newrqtype;
1024 	int rcpu;
1025 	int checkpri;
1026 	int estcpu;
1027 	int delta_uload;
1028 
1029 	crit_enter();
1030 
1031 	/*
1032 	 * Lock the scheduler (lp) belongs to.  This can be on a different
1033 	 * cpu.  Handle races.  This loop breaks out with the appropriate
1034 	 * rdd locked.
1035 	 */
1036 	for (;;) {
1037 		rcpu = lp->lwp_qcpu;
1038 		cpu_ccfence();
1039 		rdd = &dfly_pcpu[rcpu];
1040 		spin_lock(&rdd->spin);
1041 		if (rcpu == lp->lwp_qcpu)
1042 			break;
1043 		spin_unlock(&rdd->spin);
1044 	}
1045 
1046 	/*
1047 	 * Calculate the new priority and queue type
1048 	 */
1049 	newrqtype = lp->lwp_rtprio.type;
1050 
1051 	switch(newrqtype) {
1052 	case RTP_PRIO_REALTIME:
1053 	case RTP_PRIO_FIFO:
1054 		newpriority = PRIBASE_REALTIME +
1055 			     (lp->lwp_rtprio.prio & PRIMASK);
1056 		break;
1057 	case RTP_PRIO_NORMAL:
1058 		/*
1059 		 *
1060 		 */
1061 		estcpu = lp->lwp_estcpu;
1062 
1063 		/*
1064 		 * p_nice piece		Adds (0-40) * 2		0-80
1065 		 * estcpu		Adds 16384  * 4 / 512   0-128
1066 		 */
1067 		newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1068 		newpriority += estcpu * PPQ / ESTCPUPPQ;
1069 		newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1070 			      NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1071 		newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1072 		break;
1073 	case RTP_PRIO_IDLE:
1074 		newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1075 		break;
1076 	case RTP_PRIO_THREAD:
1077 		newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1078 		break;
1079 	default:
1080 		panic("Bad RTP_PRIO %d", newrqtype);
1081 		/* NOT REACHED */
1082 	}
1083 
1084 	/*
1085 	 * The LWKT scheduler doesn't dive usched structures, give it a hint
1086 	 * on the relative priority of user threads running in the kernel.
1087 	 * The LWKT scheduler will always ensure that a user thread running
1088 	 * in the kernel will get cpu some time, regardless of its upri,
1089 	 * but can decide not to instantly switch from one kernel or user
1090 	 * mode user thread to a kernel-mode user thread when it has a less
1091 	 * desireable user priority.
1092 	 *
1093 	 * td_upri has normal sense (higher values are more desireable), so
1094 	 * negate it.
1095 	 */
1096 	lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1097 
1098 	/*
1099 	 * The newpriority incorporates the queue type so do a simple masked
1100 	 * check to determine if the process has moved to another queue.  If
1101 	 * it has, and it is currently on a run queue, then move it.
1102 	 *
1103 	 * Since uload is ~PPQMASK masked, no modifications are necessary if
1104 	 * we end up in the same run queue.
1105 	 */
1106 	if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1107 		if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1108 			dfly_remrunqueue_locked(rdd, lp);
1109 			lp->lwp_priority = newpriority;
1110 			lp->lwp_rqtype = newrqtype;
1111 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1112 			dfly_setrunqueue_locked(rdd, lp);
1113 			checkpri = 1;
1114 		} else {
1115 			lp->lwp_priority = newpriority;
1116 			lp->lwp_rqtype = newrqtype;
1117 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1118 			checkpri = 0;
1119 		}
1120 	} else {
1121 		/*
1122 		 * In the same PPQ, uload cannot change.
1123 		 */
1124 		lp->lwp_priority = newpriority;
1125 		checkpri = 1;
1126 		rcpu = -1;
1127 	}
1128 
1129 	/*
1130 	 * Adjust effective load.
1131 	 *
1132 	 * Calculate load then scale up or down geometrically based on p_nice.
1133 	 * Processes niced up (positive) are less important, and processes
1134 	 * niced downard (negative) are more important.  The higher the uload,
1135 	 * the more important the thread.
1136 	 */
1137 	/* 0-511, 0-100% cpu */
1138 	delta_uload = lp->lwp_estcpu / NQS;
1139 	delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1140 
1141 
1142 	delta_uload -= lp->lwp_uload;
1143 	lp->lwp_uload += delta_uload;
1144 	if (lp->lwp_mpflags & LWP_MP_ULOAD)
1145 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1146 
1147 	/*
1148 	 * Determine if we need to reschedule the target cpu.  This only
1149 	 * occurs if the LWP is already on a scheduler queue, which means
1150 	 * that idle cpu notification has already occured.  At most we
1151 	 * need only issue a need_user_resched() on the appropriate cpu.
1152 	 *
1153 	 * The LWP may be owned by a CPU different from the current one,
1154 	 * in which case dd->uschedcp may be modified without an MP lock
1155 	 * or a spinlock held.  The worst that happens is that the code
1156 	 * below causes a spurious need_user_resched() on the target CPU
1157 	 * and dd->pri to be wrong for a short period of time, both of
1158 	 * which are harmless.
1159 	 *
1160 	 * If checkpri is 0 we are adjusting the priority of the current
1161 	 * process, possibly higher (less desireable), so ignore the upri
1162 	 * check which will fail in that case.
1163 	 */
1164 	if (rcpu >= 0) {
1165 		if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1166 		    (checkpri == 0 ||
1167 		     (rdd->upri & ~PRIMASK) >
1168 		     (lp->lwp_priority & ~PRIMASK))) {
1169 #ifdef SMP
1170 			if (rcpu == mycpu->gd_cpuid) {
1171 				spin_unlock(&rdd->spin);
1172 				need_user_resched();
1173 			} else {
1174 				spin_unlock(&rdd->spin);
1175 				lwkt_send_ipiq(globaldata_find(rcpu),
1176 					       dfly_need_user_resched_remote,
1177 					       NULL);
1178 			}
1179 #else
1180 			spin_unlock(&rdd->spin);
1181 			need_user_resched();
1182 #endif
1183 		} else {
1184 			spin_unlock(&rdd->spin);
1185 		}
1186 	} else {
1187 		spin_unlock(&rdd->spin);
1188 	}
1189 	crit_exit();
1190 }
1191 
1192 static
1193 void
1194 dfly_yield(struct lwp *lp)
1195 {
1196 #if 0
1197 	/* FUTURE (or something similar) */
1198 	switch(lp->lwp_rqtype) {
1199 	case RTP_PRIO_NORMAL:
1200 		lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
1201 		break;
1202 	default:
1203 		break;
1204 	}
1205 #endif
1206         need_user_resched();
1207 }
1208 
1209 /*
1210  * Called from fork1() when a new child process is being created.
1211  *
1212  * Give the child process an initial estcpu that is more batch then
1213  * its parent and dock the parent for the fork (but do not
1214  * reschedule the parent).
1215  *
1216  * fast
1217  *
1218  * XXX lwp should be "spawning" instead of "forking"
1219  */
1220 static void
1221 dfly_forking(struct lwp *plp, struct lwp *lp)
1222 {
1223 	/*
1224 	 * Put the child 4 queue slots (out of 32) higher than the parent
1225 	 * (less desireable than the parent).
1226 	 */
1227 	lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1228 	lp->lwp_forked = 1;
1229 	lp->lwp_estfast = 0;
1230 
1231 	/*
1232 	 * Dock the parent a cost for the fork, protecting us from fork
1233 	 * bombs.  If the parent is forking quickly make the child more
1234 	 * batchy.
1235 	 */
1236 	plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1237 }
1238 
1239 /*
1240  * Called when a lwp is being removed from this scheduler, typically
1241  * during lwp_exit().  We have to clean out any ULOAD accounting before
1242  * we can let the lp go.  The dd->spin lock is not needed for uload
1243  * updates.
1244  *
1245  * Scheduler dequeueing has already occurred, no further action in that
1246  * regard is needed.
1247  */
1248 static void
1249 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1250 {
1251 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1252 
1253 	if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1254 		atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1255 		atomic_add_int(&dd->uload, -lp->lwp_uload);
1256 		atomic_add_int(&dd->ucount, -1);
1257 		atomic_add_int(&dfly_ucount, -1);
1258 	}
1259 }
1260 
1261 /*
1262  * This function cannot block in any way, but spinlocks are ok.
1263  *
1264  * Update the uload based on the state of the thread (whether it is going
1265  * to sleep or running again).  The uload is meant to be a longer-term
1266  * load and not an instantanious load.
1267  */
1268 static void
1269 dfly_uload_update(struct lwp *lp)
1270 {
1271 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1272 
1273 	if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1274 		if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1275 			spin_lock(&dd->spin);
1276 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1277 				atomic_set_int(&lp->lwp_mpflags,
1278 					       LWP_MP_ULOAD);
1279 				atomic_add_int(&dd->uload, lp->lwp_uload);
1280 				atomic_add_int(&dd->ucount, 1);
1281 				atomic_add_int(&dfly_ucount, 1);
1282 			}
1283 			spin_unlock(&dd->spin);
1284 		}
1285 	} else if (lp->lwp_slptime > 0) {
1286 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1287 			spin_lock(&dd->spin);
1288 			if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1289 				atomic_clear_int(&lp->lwp_mpflags,
1290 						 LWP_MP_ULOAD);
1291 				atomic_add_int(&dd->uload, -lp->lwp_uload);
1292 				atomic_add_int(&dd->ucount, -1);
1293 				atomic_add_int(&dfly_ucount, -1);
1294 			}
1295 			spin_unlock(&dd->spin);
1296 		}
1297 	}
1298 }
1299 
1300 /*
1301  * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1302  * it selects a user process and returns it.  If chklp is non-NULL and chklp
1303  * has a better or equal priority then the process that would otherwise be
1304  * chosen, NULL is returned.
1305  *
1306  * Until we fix the RUNQ code the chklp test has to be strict or we may
1307  * bounce between processes trying to acquire the current process designation.
1308  *
1309  * Must be called with rdd->spin locked.  The spinlock is left intact through
1310  * the entire routine.  dd->spin does not have to be locked.
1311  *
1312  * If worst is non-zero this function finds the worst thread instead of the
1313  * best thread (used by the schedulerclock-based rover).
1314  */
1315 static
1316 struct lwp *
1317 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1318 		       struct lwp *chklp, int worst)
1319 {
1320 	struct lwp *lp;
1321 	struct rq *q;
1322 	u_int32_t *which, *which2;
1323 	u_int32_t pri;
1324 	u_int32_t rtqbits;
1325 	u_int32_t tsqbits;
1326 	u_int32_t idqbits;
1327 
1328 	rtqbits = rdd->rtqueuebits;
1329 	tsqbits = rdd->queuebits;
1330 	idqbits = rdd->idqueuebits;
1331 
1332 	if (worst) {
1333 		if (idqbits) {
1334 			pri = bsrl(idqbits);
1335 			q = &rdd->idqueues[pri];
1336 			which = &rdd->idqueuebits;
1337 			which2 = &idqbits;
1338 		} else if (tsqbits) {
1339 			pri = bsrl(tsqbits);
1340 			q = &rdd->queues[pri];
1341 			which = &rdd->queuebits;
1342 			which2 = &tsqbits;
1343 		} else if (rtqbits) {
1344 			pri = bsrl(rtqbits);
1345 			q = &rdd->rtqueues[pri];
1346 			which = &rdd->rtqueuebits;
1347 			which2 = &rtqbits;
1348 		} else {
1349 			return (NULL);
1350 		}
1351 		lp = TAILQ_LAST(q, rq);
1352 	} else {
1353 		if (rtqbits) {
1354 			pri = bsfl(rtqbits);
1355 			q = &rdd->rtqueues[pri];
1356 			which = &rdd->rtqueuebits;
1357 			which2 = &rtqbits;
1358 		} else if (tsqbits) {
1359 			pri = bsfl(tsqbits);
1360 			q = &rdd->queues[pri];
1361 			which = &rdd->queuebits;
1362 			which2 = &tsqbits;
1363 		} else if (idqbits) {
1364 			pri = bsfl(idqbits);
1365 			q = &rdd->idqueues[pri];
1366 			which = &rdd->idqueuebits;
1367 			which2 = &idqbits;
1368 		} else {
1369 			return (NULL);
1370 		}
1371 		lp = TAILQ_FIRST(q);
1372 	}
1373 	KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1374 
1375 	/*
1376 	 * If the passed lwp <chklp> is reasonably close to the selected
1377 	 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1378 	 *
1379 	 * Note that we must error on the side of <chklp> to avoid bouncing
1380 	 * between threads in the acquire code.
1381 	 */
1382 	if (chklp) {
1383 		if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1384 			return(NULL);
1385 	}
1386 
1387 	KTR_COND_LOG(usched_chooseproc,
1388 	    lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1389 	    lp->lwp_proc->p_pid,
1390 	    lp->lwp_thread->td_gd->gd_cpuid,
1391 	    mycpu->gd_cpuid);
1392 
1393 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1394 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1395 	TAILQ_REMOVE(q, lp, lwp_procq);
1396 	--rdd->runqcount;
1397 	if (TAILQ_EMPTY(q))
1398 		*which &= ~(1 << pri);
1399 
1400 	/*
1401 	 * If we are choosing a process from rdd with the intent to
1402 	 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1403 	 * is still held.
1404 	 */
1405 	if (rdd != dd) {
1406 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1407 			atomic_add_int(&rdd->uload, -lp->lwp_uload);
1408 			atomic_add_int(&rdd->ucount, -1);
1409 			atomic_add_int(&dfly_ucount, -1);
1410 		}
1411 		lp->lwp_qcpu = dd->cpuid;
1412 		atomic_add_int(&dd->uload, lp->lwp_uload);
1413 		atomic_add_int(&dd->ucount, 1);
1414 		atomic_add_int(&dfly_ucount, 1);
1415 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1416 	}
1417 	return lp;
1418 }
1419 
1420 #ifdef SMP
1421 
1422 /*
1423  * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1424  *
1425  * Choose a cpu node to schedule lp on, hopefully nearby its current
1426  * node.
1427  *
1428  * We give the current node a modest advantage for obvious reasons.
1429  *
1430  * We also give the node the thread was woken up FROM a slight advantage
1431  * in order to try to schedule paired threads which synchronize/block waiting
1432  * for each other fairly close to each other.  Similarly in a network setting
1433  * this feature will also attempt to place a user process near the kernel
1434  * protocol thread that is feeding it data.  THIS IS A CRITICAL PART of the
1435  * algorithm as it heuristically groups synchronizing processes for locality
1436  * of reference in multi-socket systems.
1437  *
1438  * We check against running processes and give a big advantage if there
1439  * are none running.
1440  *
1441  * The caller will normally dfly_setrunqueue() lp on the returned queue.
1442  *
1443  * When the topology is known choose a cpu whos group has, in aggregate,
1444  * has the lowest weighted load.
1445  */
1446 static
1447 dfly_pcpu_t
1448 dfly_choose_best_queue(struct lwp *lp)
1449 {
1450 	cpumask_t wakemask;
1451 	cpumask_t mask;
1452 	cpu_node_t *cpup;
1453 	cpu_node_t *cpun;
1454 	cpu_node_t *cpub;
1455 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1456 	dfly_pcpu_t rdd;
1457 	int wakecpu;
1458 	int cpuid;
1459 	int n;
1460 	int count;
1461 	int load;
1462 	int lowest_load;
1463 
1464 	/*
1465 	 * When the topology is unknown choose a random cpu that is hopefully
1466 	 * idle.
1467 	 */
1468 	if (dd->cpunode == NULL)
1469 		return (dfly_choose_queue_simple(dd, lp));
1470 
1471 	/*
1472 	 * Pairing mask
1473 	 */
1474 	if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1475 		wakemask = dfly_pcpu[wakecpu].cpumask;
1476 	else
1477 		wakemask = 0;
1478 
1479 	/*
1480 	 * When the topology is known choose a cpu whos group has, in
1481 	 * aggregate, has the lowest weighted load.
1482 	 */
1483 	cpup = root_cpu_node;
1484 	rdd = dd;
1485 
1486 	while (cpup) {
1487 		/*
1488 		 * Degenerate case super-root
1489 		 */
1490 		if (cpup->child_node && cpup->child_no == 1) {
1491 			cpup = cpup->child_node;
1492 			continue;
1493 		}
1494 
1495 		/*
1496 		 * Terminal cpunode
1497 		 */
1498 		if (cpup->child_node == NULL) {
1499 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1500 			break;
1501 		}
1502 
1503 		cpub = NULL;
1504 		lowest_load = 0x7FFFFFFF;
1505 
1506 		for (n = 0; n < cpup->child_no; ++n) {
1507 			/*
1508 			 * Accumulate load information for all cpus
1509 			 * which are members of this node.
1510 			 */
1511 			cpun = &cpup->child_node[n];
1512 			mask = cpun->members & usched_global_cpumask &
1513 			       smp_active_mask & lp->lwp_cpumask;
1514 			if (mask == 0)
1515 				continue;
1516 
1517 			count = 0;
1518 			load = 0;
1519 
1520 			while (mask) {
1521 				cpuid = BSFCPUMASK(mask);
1522 				rdd = &dfly_pcpu[cpuid];
1523 				load += rdd->uload;
1524 				load += rdd->ucount * usched_dfly_weight3;
1525 
1526 				if (rdd->uschedcp == NULL &&
1527 				    rdd->runqcount == 0 &&
1528 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1529 				) {
1530 					load -= usched_dfly_weight4;
1531 				}
1532 #if 0
1533 				else if (rdd->upri > lp->lwp_priority + PPQ) {
1534 					load -= usched_dfly_weight4 / 2;
1535 				}
1536 #endif
1537 				mask &= ~CPUMASK(cpuid);
1538 				++count;
1539 			}
1540 
1541 			/*
1542 			 * Compensate if the lp is already accounted for in
1543 			 * the aggregate uload for this mask set.  We want
1544 			 * to calculate the loads as if lp were not present,
1545 			 * otherwise the calculation is bogus.
1546 			 */
1547 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1548 			    (dd->cpumask & cpun->members)) {
1549 				load -= lp->lwp_uload;
1550 				load -= usched_dfly_weight3;
1551 			}
1552 
1553 			load /= count;
1554 
1555 			/*
1556 			 * Advantage the cpu group (lp) is already on.
1557 			 */
1558 			if (cpun->members & dd->cpumask)
1559 				load -= usched_dfly_weight1;
1560 
1561 			/*
1562 			 * Advantage the cpu group we want to pair (lp) to,
1563 			 * but don't let it go to the exact same cpu as
1564 			 * the wakecpu target.
1565 			 *
1566 			 * We do this by checking whether cpun is a
1567 			 * terminal node or not.  All cpun's at the same
1568 			 * level will either all be terminal or all not
1569 			 * terminal.
1570 			 *
1571 			 * If it is and we match we disadvantage the load.
1572 			 * If it is and we don't match we advantage the load.
1573 			 *
1574 			 * Also note that we are effectively disadvantaging
1575 			 * all-but-one by the same amount, so it won't effect
1576 			 * the weight1 factor for the all-but-one nodes.
1577 			 */
1578 			if (cpun->members & wakemask) {
1579 				if (cpun->child_node != NULL) {
1580 					/* advantage */
1581 					load -= usched_dfly_weight2;
1582 				} else {
1583 					if (usched_dfly_features & 0x10)
1584 						load += usched_dfly_weight2;
1585 					else
1586 						load -= usched_dfly_weight2;
1587 				}
1588 			}
1589 
1590 			/*
1591 			 * Calculate the best load
1592 			 */
1593 			if (cpub == NULL || lowest_load > load ||
1594 			    (lowest_load == load &&
1595 			     (cpun->members & dd->cpumask))
1596 			) {
1597 				lowest_load = load;
1598 				cpub = cpun;
1599 			}
1600 		}
1601 		cpup = cpub;
1602 	}
1603 	if (usched_dfly_chooser)
1604 		kprintf("lp %02d->%02d %s\n",
1605 			lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1606 	return (rdd);
1607 }
1608 
1609 /*
1610  * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1611  *
1612  * Choose the worst queue close to dd's cpu node with a non-empty runq
1613  * that is NOT dd.  Also require that the moving of the highest-load thread
1614  * from rdd to dd does not cause the uload's to cross each other.
1615  *
1616  * This is used by the thread chooser when the current cpu's queues are
1617  * empty to steal a thread from another cpu's queue.  We want to offload
1618  * the most heavily-loaded queue.
1619  */
1620 static
1621 dfly_pcpu_t
1622 dfly_choose_worst_queue(dfly_pcpu_t dd)
1623 {
1624 	cpumask_t mask;
1625 	cpu_node_t *cpup;
1626 	cpu_node_t *cpun;
1627 	cpu_node_t *cpub;
1628 	dfly_pcpu_t rdd;
1629 	int cpuid;
1630 	int n;
1631 	int count;
1632 	int load;
1633 #if 0
1634 	int pri;
1635 	int hpri;
1636 #endif
1637 	int highest_load;
1638 
1639 	/*
1640 	 * When the topology is unknown choose a random cpu that is hopefully
1641 	 * idle.
1642 	 */
1643 	if (dd->cpunode == NULL) {
1644 		return (NULL);
1645 	}
1646 
1647 	/*
1648 	 * When the topology is known choose a cpu whos group has, in
1649 	 * aggregate, has the lowest weighted load.
1650 	 */
1651 	cpup = root_cpu_node;
1652 	rdd = dd;
1653 	while (cpup) {
1654 		/*
1655 		 * Degenerate case super-root
1656 		 */
1657 		if (cpup->child_node && cpup->child_no == 1) {
1658 			cpup = cpup->child_node;
1659 			continue;
1660 		}
1661 
1662 		/*
1663 		 * Terminal cpunode
1664 		 */
1665 		if (cpup->child_node == NULL) {
1666 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1667 			break;
1668 		}
1669 
1670 		cpub = NULL;
1671 		highest_load = 0;
1672 
1673 		for (n = 0; n < cpup->child_no; ++n) {
1674 			/*
1675 			 * Accumulate load information for all cpus
1676 			 * which are members of this node.
1677 			 */
1678 			cpun = &cpup->child_node[n];
1679 			mask = cpun->members & usched_global_cpumask &
1680 			       smp_active_mask;
1681 			if (mask == 0)
1682 				continue;
1683 			count = 0;
1684 			load = 0;
1685 
1686 			while (mask) {
1687 				cpuid = BSFCPUMASK(mask);
1688 				rdd = &dfly_pcpu[cpuid];
1689 				load += rdd->uload;
1690 				load += rdd->ucount * usched_dfly_weight3;
1691 				if (rdd->uschedcp == NULL &&
1692 				    rdd->runqcount == 0 &&
1693 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1694 				) {
1695 					load -= usched_dfly_weight4;
1696 				}
1697 #if 0
1698 				else if (rdd->upri > dd->upri + PPQ) {
1699 					load -= usched_dfly_weight4 / 2;
1700 				}
1701 #endif
1702 				mask &= ~CPUMASK(cpuid);
1703 				++count;
1704 			}
1705 			load /= count;
1706 
1707 			/*
1708 			 * Prefer candidates which are somewhat closer to
1709 			 * our cpu.
1710 			 */
1711 			if (dd->cpumask & cpun->members)
1712 				load += usched_dfly_weight1;
1713 
1714 			/*
1715 			 * The best candidate is the one with the worst
1716 			 * (highest) load.
1717 			 */
1718 			if (cpub == NULL || highest_load < load) {
1719 				highest_load = load;
1720 				cpub = cpun;
1721 			}
1722 		}
1723 		cpup = cpub;
1724 	}
1725 
1726 	/*
1727 	 * We never return our own node (dd), and only return a remote
1728 	 * node if it's load is significantly worse than ours (i.e. where
1729 	 * stealing a thread would be considered reasonable).
1730 	 *
1731 	 * This also helps us avoid breaking paired threads apart which
1732 	 * can have disastrous effects on performance.
1733 	 */
1734 	if (rdd == dd)
1735 		return(NULL);
1736 
1737 #if 0
1738 	hpri = 0;
1739 	if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1740 		hpri = pri;
1741 	if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1742 		hpri = pri;
1743 	if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1744 		hpri = pri;
1745 	hpri *= PPQ;
1746 	if (rdd->uload - hpri < dd->uload + hpri)
1747 		return(NULL);
1748 #endif
1749 	return (rdd);
1750 }
1751 
1752 static
1753 dfly_pcpu_t
1754 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1755 {
1756 	dfly_pcpu_t rdd;
1757 	cpumask_t tmpmask;
1758 	cpumask_t mask;
1759 	int cpuid;
1760 
1761 	/*
1762 	 * Fallback to the original heuristic, select random cpu,
1763 	 * first checking cpus not currently running a user thread.
1764 	 */
1765 	++dfly_scancpu;
1766 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1767 	mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1768 	       smp_active_mask & usched_global_cpumask;
1769 
1770 	while (mask) {
1771 		tmpmask = ~(CPUMASK(cpuid) - 1);
1772 		if (mask & tmpmask)
1773 			cpuid = BSFCPUMASK(mask & tmpmask);
1774 		else
1775 			cpuid = BSFCPUMASK(mask);
1776 		rdd = &dfly_pcpu[cpuid];
1777 
1778 		if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1779 			goto found;
1780 		mask &= ~CPUMASK(cpuid);
1781 	}
1782 
1783 	/*
1784 	 * Then cpus which might have a currently running lp
1785 	 */
1786 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1787 	mask = dfly_curprocmask & dfly_rdyprocmask &
1788 	       lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1789 
1790 	while (mask) {
1791 		tmpmask = ~(CPUMASK(cpuid) - 1);
1792 		if (mask & tmpmask)
1793 			cpuid = BSFCPUMASK(mask & tmpmask);
1794 		else
1795 			cpuid = BSFCPUMASK(mask);
1796 		rdd = &dfly_pcpu[cpuid];
1797 
1798 		if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1799 			goto found;
1800 		mask &= ~CPUMASK(cpuid);
1801 	}
1802 
1803 	/*
1804 	 * If we cannot find a suitable cpu we reload from dfly_scancpu
1805 	 * and round-robin.  Other cpus will pickup as they release their
1806 	 * current lwps or become ready.
1807 	 *
1808 	 * Avoid a degenerate system lockup case if usched_global_cpumask
1809 	 * is set to 0 or otherwise does not cover lwp_cpumask.
1810 	 *
1811 	 * We only kick the target helper thread in this case, we do not
1812 	 * set the user resched flag because
1813 	 */
1814 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1815 	if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1816 		cpuid = 0;
1817 	rdd = &dfly_pcpu[cpuid];
1818 found:
1819 	return (rdd);
1820 }
1821 
1822 static
1823 void
1824 dfly_need_user_resched_remote(void *dummy)
1825 {
1826 	globaldata_t gd = mycpu;
1827 	dfly_pcpu_t  dd = &dfly_pcpu[gd->gd_cpuid];
1828 
1829 	/*
1830 	 * Flag reschedule needed
1831 	 */
1832 	need_user_resched();
1833 
1834 	/*
1835 	 * If no user thread is currently running we need to kick the helper
1836 	 * on our cpu to recover.  Otherwise the cpu will never schedule
1837 	 * anything again.
1838 	 *
1839 	 * We cannot schedule the process ourselves because this is an
1840 	 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1841 	 *
1842 	 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1843 	 */
1844 	if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1845 		atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1846 		wakeup_mycpu(&dd->helper_thread);
1847 	}
1848 }
1849 
1850 #endif
1851 
1852 /*
1853  * dfly_remrunqueue_locked() removes a given process from the run queue
1854  * that it is on, clearing the queue busy bit if it becomes empty.
1855  *
1856  * Note that user process scheduler is different from the LWKT schedule.
1857  * The user process scheduler only manages user processes but it uses LWKT
1858  * underneath, and a user process operating in the kernel will often be
1859  * 'released' from our management.
1860  *
1861  * uload is NOT adjusted here.  It is only adjusted if the lwkt_thread goes
1862  * to sleep or the lwp is moved to a different runq.
1863  */
1864 static void
1865 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1866 {
1867 	struct rq *q;
1868 	u_int32_t *which;
1869 	u_int8_t pri;
1870 
1871 	KKASSERT(rdd->runqcount >= 0);
1872 
1873 	pri = lp->lwp_rqindex;
1874 
1875 	switch(lp->lwp_rqtype) {
1876 	case RTP_PRIO_NORMAL:
1877 		q = &rdd->queues[pri];
1878 		which = &rdd->queuebits;
1879 		break;
1880 	case RTP_PRIO_REALTIME:
1881 	case RTP_PRIO_FIFO:
1882 		q = &rdd->rtqueues[pri];
1883 		which = &rdd->rtqueuebits;
1884 		break;
1885 	case RTP_PRIO_IDLE:
1886 		q = &rdd->idqueues[pri];
1887 		which = &rdd->idqueuebits;
1888 		break;
1889 	default:
1890 		panic("remrunqueue: invalid rtprio type");
1891 		/* NOT REACHED */
1892 	}
1893 	KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1894 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1895 	TAILQ_REMOVE(q, lp, lwp_procq);
1896 	--rdd->runqcount;
1897 	if (TAILQ_EMPTY(q)) {
1898 		KASSERT((*which & (1 << pri)) != 0,
1899 			("remrunqueue: remove from empty queue"));
1900 		*which &= ~(1 << pri);
1901 	}
1902 }
1903 
1904 /*
1905  * dfly_setrunqueue_locked()
1906  *
1907  * Add a process whos rqtype and rqindex had previously been calculated
1908  * onto the appropriate run queue.   Determine if the addition requires
1909  * a reschedule on a cpu and return the cpuid or -1.
1910  *
1911  * NOTE: 	  Lower priorities are better priorities.
1912  *
1913  * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1914  *		  sum of the rough lwp_priority for all running and runnable
1915  *		  processes.  Lower priority processes (higher lwp_priority
1916  *		  values) actually DO count as more load, not less, because
1917  *		  these are the programs which require the most care with
1918  *		  regards to cpu selection.
1919  */
1920 static void
1921 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1922 {
1923 	struct rq *q;
1924 	u_int32_t *which;
1925 	int pri;
1926 
1927 	KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1928 
1929 	if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1930 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1931 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1932 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1933 		atomic_add_int(&dfly_ucount, 1);
1934 	}
1935 
1936 	pri = lp->lwp_rqindex;
1937 
1938 	switch(lp->lwp_rqtype) {
1939 	case RTP_PRIO_NORMAL:
1940 		q = &rdd->queues[pri];
1941 		which = &rdd->queuebits;
1942 		break;
1943 	case RTP_PRIO_REALTIME:
1944 	case RTP_PRIO_FIFO:
1945 		q = &rdd->rtqueues[pri];
1946 		which = &rdd->rtqueuebits;
1947 		break;
1948 	case RTP_PRIO_IDLE:
1949 		q = &rdd->idqueues[pri];
1950 		which = &rdd->idqueuebits;
1951 		break;
1952 	default:
1953 		panic("remrunqueue: invalid rtprio type");
1954 		/* NOT REACHED */
1955 	}
1956 
1957 	/*
1958 	 * Place us on the selected queue.  Determine if we should be
1959 	 * placed at the head of the queue or at the end.
1960 	 *
1961 	 * We are placed at the tail if our round-robin count has expired,
1962 	 * or is about to expire and the system thinks its a good place to
1963 	 * round-robin, or there is already a next thread on the queue
1964 	 * (it might be trying to pick up where it left off and we don't
1965 	 * want to interfere).
1966 	 */
1967 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1968 	atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1969 	++rdd->runqcount;
1970 
1971 	if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1972 	    (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1973 	     (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1974 	    !TAILQ_EMPTY(q)
1975 	) {
1976 		atomic_clear_int(&lp->lwp_thread->td_mpflags,
1977 				 TDF_MP_BATCH_DEMARC);
1978 		lp->lwp_rrcount = 0;
1979 		TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1980 	} else {
1981 		if (TAILQ_EMPTY(q))
1982 			lp->lwp_rrcount = 0;
1983 		TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1984 	}
1985 	*which |= 1 << pri;
1986 }
1987 
1988 #ifdef SMP
1989 
1990 /*
1991  * For SMP systems a user scheduler helper thread is created for each
1992  * cpu and is used to allow one cpu to wakeup another for the purposes of
1993  * scheduling userland threads from setrunqueue().
1994  *
1995  * UP systems do not need the helper since there is only one cpu.
1996  *
1997  * We can't use the idle thread for this because we might block.
1998  * Additionally, doing things this way allows us to HLT idle cpus
1999  * on MP systems.
2000  */
2001 static void
2002 dfly_helper_thread(void *dummy)
2003 {
2004     globaldata_t gd;
2005     dfly_pcpu_t dd;
2006     dfly_pcpu_t rdd;
2007     struct lwp *nlp;
2008     cpumask_t mask;
2009     int cpuid;
2010 
2011     gd = mycpu;
2012     cpuid = gd->gd_cpuid;	/* doesn't change */
2013     mask = gd->gd_cpumask;	/* doesn't change */
2014     dd = &dfly_pcpu[cpuid];
2015 
2016     /*
2017      * Since we only want to be woken up only when no user processes
2018      * are scheduled on a cpu, run at an ultra low priority.
2019      */
2020     lwkt_setpri_self(TDPRI_USER_SCHEDULER);
2021 
2022     tsleep(&dd->helper_thread, 0, "schslp", 0);
2023 
2024     for (;;) {
2025 	/*
2026 	 * We use the LWKT deschedule-interlock trick to avoid racing
2027 	 * dfly_rdyprocmask.  This means we cannot block through to the
2028 	 * manual lwkt_switch() call we make below.
2029 	 */
2030 	crit_enter_gd(gd);
2031 	tsleep_interlock(&dd->helper_thread, 0);
2032 
2033 	spin_lock(&dd->spin);
2034 
2035 	atomic_set_cpumask(&dfly_rdyprocmask, mask);
2036 	clear_user_resched();	/* This satisfied the reschedule request */
2037 #if 0
2038 	dd->rrcount = 0;	/* Reset the round-robin counter */
2039 #endif
2040 
2041 	if (dd->runqcount || dd->uschedcp != NULL) {
2042 		/*
2043 		 * Threads are available.  A thread may or may not be
2044 		 * currently scheduled.  Get the best thread already queued
2045 		 * to this cpu.
2046 		 */
2047 		nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
2048 		if (nlp) {
2049 			atomic_set_cpumask(&dfly_curprocmask, mask);
2050 			dd->upri = nlp->lwp_priority;
2051 			dd->uschedcp = nlp;
2052 #if 0
2053 			dd->rrcount = 0;	/* reset round robin */
2054 #endif
2055 			spin_unlock(&dd->spin);
2056 			lwkt_acquire(nlp->lwp_thread);
2057 			lwkt_schedule(nlp->lwp_thread);
2058 		} else {
2059 			/*
2060 			 * This situation should not occur because we had
2061 			 * at least one thread available.
2062 			 */
2063 			spin_unlock(&dd->spin);
2064 		}
2065 	} else if (usched_dfly_features & 0x01) {
2066 		/*
2067 		 * This cpu is devoid of runnable threads, steal a thread
2068 		 * from another cpu.  Since we're stealing, might as well
2069 		 * load balance at the same time.
2070 		 *
2071 		 * We choose the highest-loaded thread from the worst queue.
2072 		 *
2073 		 * NOTE! This function only returns a non-NULL rdd when
2074 		 *	 another cpu's queue is obviously overloaded.  We
2075 		 *	 do not want to perform the type of rebalancing
2076 		 *	 the schedclock does here because it would result
2077 		 *	 in insane process pulling when 'steady' state is
2078 		 *	 partially unbalanced (e.g. 6 runnables and only
2079 		 *	 4 cores).
2080 		 */
2081 		rdd = dfly_choose_worst_queue(dd);
2082 		if (rdd && spin_trylock(&rdd->spin)) {
2083 			nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2084 			spin_unlock(&rdd->spin);
2085 		} else {
2086 			nlp = NULL;
2087 		}
2088 		if (nlp) {
2089 			atomic_set_cpumask(&dfly_curprocmask, mask);
2090 			dd->upri = nlp->lwp_priority;
2091 			dd->uschedcp = nlp;
2092 #if 0
2093 			dd->rrcount = 0;	/* reset round robin */
2094 #endif
2095 			spin_unlock(&dd->spin);
2096 			lwkt_acquire(nlp->lwp_thread);
2097 			lwkt_schedule(nlp->lwp_thread);
2098 		} else {
2099 			/*
2100 			 * Leave the thread on our run queue.  Another
2101 			 * scheduler will try to pull it later.
2102 			 */
2103 			spin_unlock(&dd->spin);
2104 		}
2105 	} else {
2106 		/*
2107 		 * devoid of runnable threads and not allowed to steal
2108 		 * any.
2109 		 */
2110 		spin_unlock(&dd->spin);
2111 	}
2112 
2113 	/*
2114 	 * We're descheduled unless someone scheduled us.  Switch away.
2115 	 * Exiting the critical section will cause splz() to be called
2116 	 * for us if interrupts and such are pending.
2117 	 */
2118 	crit_exit_gd(gd);
2119 	tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2120     }
2121 }
2122 
2123 #if 0
2124 static int
2125 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2126 {
2127 	int error, new_val;
2128 
2129 	new_val = usched_dfly_stick_to_level;
2130 
2131 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2132         if (error != 0 || req->newptr == NULL)
2133 		return (error);
2134 	if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2135 		return (EINVAL);
2136 	usched_dfly_stick_to_level = new_val;
2137 	return (0);
2138 }
2139 #endif
2140 
2141 /*
2142  * Setup our scheduler helpers.  Note that curprocmask bit 0 has already
2143  * been cleared by rqinit() and we should not mess with it further.
2144  */
2145 static void
2146 dfly_helper_thread_cpu_init(void)
2147 {
2148 	int i;
2149 	int j;
2150 	int cpuid;
2151 	int smt_not_supported = 0;
2152 	int cache_coherent_not_supported = 0;
2153 
2154 	if (bootverbose)
2155 		kprintf("Start scheduler helpers on cpus:\n");
2156 
2157 	sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2158 	usched_dfly_sysctl_tree =
2159 		SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2160 				SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2161 				"usched_dfly", CTLFLAG_RD, 0, "");
2162 
2163 	for (i = 0; i < ncpus; ++i) {
2164 		dfly_pcpu_t dd = &dfly_pcpu[i];
2165 		cpumask_t mask = CPUMASK(i);
2166 
2167 		if ((mask & smp_active_mask) == 0)
2168 		    continue;
2169 
2170 		spin_init(&dd->spin);
2171 		dd->cpunode = get_cpu_node_by_cpuid(i);
2172 		dd->cpuid = i;
2173 		dd->cpumask = CPUMASK(i);
2174 		for (j = 0; j < NQS; j++) {
2175 			TAILQ_INIT(&dd->queues[j]);
2176 			TAILQ_INIT(&dd->rtqueues[j]);
2177 			TAILQ_INIT(&dd->idqueues[j]);
2178 		}
2179 		atomic_clear_cpumask(&dfly_curprocmask, 1);
2180 
2181 		if (dd->cpunode == NULL) {
2182 			smt_not_supported = 1;
2183 			cache_coherent_not_supported = 1;
2184 			if (bootverbose)
2185 				kprintf ("\tcpu%d - WARNING: No CPU NODE "
2186 					 "found for cpu\n", i);
2187 		} else {
2188 			switch (dd->cpunode->type) {
2189 			case THREAD_LEVEL:
2190 				if (bootverbose)
2191 					kprintf ("\tcpu%d - HyperThreading "
2192 						 "available. Core siblings: ",
2193 						 i);
2194 				break;
2195 			case CORE_LEVEL:
2196 				smt_not_supported = 1;
2197 
2198 				if (bootverbose)
2199 					kprintf ("\tcpu%d - No HT available, "
2200 						 "multi-core/physical "
2201 						 "cpu. Physical siblings: ",
2202 						 i);
2203 				break;
2204 			case CHIP_LEVEL:
2205 				smt_not_supported = 1;
2206 
2207 				if (bootverbose)
2208 					kprintf ("\tcpu%d - No HT available, "
2209 						 "single-core/physical cpu. "
2210 						 "Package Siblings: ",
2211 						 i);
2212 				break;
2213 			default:
2214 				/* Let's go for safe defaults here */
2215 				smt_not_supported = 1;
2216 				cache_coherent_not_supported = 1;
2217 				if (bootverbose)
2218 					kprintf ("\tcpu%d - Unknown cpunode->"
2219 						 "type=%u. Siblings: ",
2220 						 i,
2221 						 (u_int)dd->cpunode->type);
2222 				break;
2223 			}
2224 
2225 			if (bootverbose) {
2226 				if (dd->cpunode->parent_node != NULL) {
2227 					CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2228 						kprintf("cpu%d ", cpuid);
2229 					kprintf("\n");
2230 				} else {
2231 					kprintf(" no siblings\n");
2232 				}
2233 			}
2234 		}
2235 
2236 		lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2237 			    0, i, "usched %d", i);
2238 
2239 		/*
2240 		 * Allow user scheduling on the target cpu.  cpu #0 has already
2241 		 * been enabled in rqinit().
2242 		 */
2243 		if (i)
2244 		    atomic_clear_cpumask(&dfly_curprocmask, mask);
2245 		atomic_set_cpumask(&dfly_rdyprocmask, mask);
2246 		dd->upri = PRIBASE_NULL;
2247 
2248 	}
2249 
2250 	/* usched_dfly sysctl configurable parameters */
2251 
2252 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2253 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2254 		       OID_AUTO, "rrinterval", CTLFLAG_RW,
2255 		       &usched_dfly_rrinterval, 0, "");
2256 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2257 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2258 		       OID_AUTO, "decay", CTLFLAG_RW,
2259 		       &usched_dfly_decay, 0, "Extra decay when not running");
2260 
2261 	/* Add enable/disable option for SMT scheduling if supported */
2262 	if (smt_not_supported) {
2263 		usched_dfly_smt = 0;
2264 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2265 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2266 				  OID_AUTO, "smt", CTLFLAG_RD,
2267 				  "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2268 	} else {
2269 		usched_dfly_smt = 1;
2270 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2271 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2272 			       OID_AUTO, "smt", CTLFLAG_RW,
2273 			       &usched_dfly_smt, 0, "Enable SMT scheduling");
2274 	}
2275 
2276 	/*
2277 	 * Add enable/disable option for cache coherent scheduling
2278 	 * if supported
2279 	 */
2280 	if (cache_coherent_not_supported) {
2281 		usched_dfly_cache_coherent = 0;
2282 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2283 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2284 				  OID_AUTO, "cache_coherent", CTLFLAG_RD,
2285 				  "NOT SUPPORTED", 0,
2286 				  "Cache coherence NOT SUPPORTED");
2287 	} else {
2288 		usched_dfly_cache_coherent = 1;
2289 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2290 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2291 			       OID_AUTO, "cache_coherent", CTLFLAG_RW,
2292 			       &usched_dfly_cache_coherent, 0,
2293 			       "Enable/Disable cache coherent scheduling");
2294 
2295 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2296 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2297 			       OID_AUTO, "weight1", CTLFLAG_RW,
2298 			       &usched_dfly_weight1, 200,
2299 			       "Weight selection for current cpu");
2300 
2301 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2302 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2303 			       OID_AUTO, "weight2", CTLFLAG_RW,
2304 			       &usched_dfly_weight2, 180,
2305 			       "Weight selection for wakefrom cpu");
2306 
2307 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2308 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2309 			       OID_AUTO, "weight3", CTLFLAG_RW,
2310 			       &usched_dfly_weight3, 40,
2311 			       "Weight selection for num threads on queue");
2312 
2313 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2314 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2315 			       OID_AUTO, "weight4", CTLFLAG_RW,
2316 			       &usched_dfly_weight4, 160,
2317 			       "Availability of other idle cpus");
2318 
2319 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2320 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2321 			       OID_AUTO, "fast_resched", CTLFLAG_RW,
2322 			       &usched_dfly_fast_resched, 0,
2323 			       "Availability of other idle cpus");
2324 
2325 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2326 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2327 			       OID_AUTO, "features", CTLFLAG_RW,
2328 			       &usched_dfly_features, 0x8F,
2329 			       "Allow pulls into empty queues");
2330 
2331 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2332 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2333 			       OID_AUTO, "swmask", CTLFLAG_RW,
2334 			       &usched_dfly_swmask, ~PPQMASK,
2335 			       "Queue mask to force thread switch");
2336 
2337 
2338 #if 0
2339 		SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2340 				SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2341 				OID_AUTO, "stick_to_level",
2342 				CTLTYPE_INT | CTLFLAG_RW,
2343 				NULL, sizeof usched_dfly_stick_to_level,
2344 				sysctl_usched_dfly_stick_to_level, "I",
2345 				"Stick a process to this level. See sysctl"
2346 				"paremter hw.cpu_topology.level_description");
2347 #endif
2348 	}
2349 }
2350 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2351 	dfly_helper_thread_cpu_init, NULL)
2352 
2353 #else /* No SMP options - just add the configurable parameters to sysctl */
2354 
2355 static void
2356 sched_sysctl_tree_init(void)
2357 {
2358 	sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2359 	usched_dfly_sysctl_tree =
2360 		SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2361 				SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2362 				"usched_dfly", CTLFLAG_RD, 0, "");
2363 
2364 	/* usched_dfly sysctl configurable parameters */
2365 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2366 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2367 		       OID_AUTO, "rrinterval", CTLFLAG_RW,
2368 		       &usched_dfly_rrinterval, 0, "");
2369 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2370 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2371 		       OID_AUTO, "decay", CTLFLAG_RW,
2372 		       &usched_dfly_decay, 0, "Extra decay when not running");
2373 }
2374 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2375 	sched_sysctl_tree_init, NULL)
2376 #endif
2377