xref: /dflybsd-src/sys/kern/usched_dfly.c (revision 7a27faded6bccbda68815bb82b3be97f731409b9)
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 	cpu_node_t	*cpunode;
120 };
121 
122 typedef struct usched_dfly_pcpu	*dfly_pcpu_t;
123 
124 static void dfly_acquire_curproc(struct lwp *lp);
125 static void dfly_release_curproc(struct lwp *lp);
126 static void dfly_select_curproc(globaldata_t gd);
127 static void dfly_setrunqueue(struct lwp *lp);
128 static void dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp);
129 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period,
130 				sysclock_t cpstamp);
131 static void dfly_recalculate_estcpu(struct lwp *lp);
132 static void dfly_resetpriority(struct lwp *lp);
133 static void dfly_forking(struct lwp *plp, struct lwp *lp);
134 static void dfly_exiting(struct lwp *lp, struct proc *);
135 static void dfly_uload_update(struct lwp *lp);
136 static void dfly_yield(struct lwp *lp);
137 static void dfly_changeqcpu_locked(struct lwp *lp,
138 				dfly_pcpu_t dd, dfly_pcpu_t rdd);
139 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
140 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
141 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
142 static void dfly_need_user_resched_remote(void *dummy);
143 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
144 					  struct lwp *chklp, int worst);
145 static void dfly_remrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
146 static void dfly_setrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
147 static void dfly_changedcpu(struct lwp *lp);
148 
149 struct usched usched_dfly = {
150 	{ NULL },
151 	"dfly", "Original DragonFly Scheduler",
152 	NULL,			/* default registration */
153 	NULL,			/* default deregistration */
154 	dfly_acquire_curproc,
155 	dfly_release_curproc,
156 	dfly_setrunqueue,
157 	dfly_schedulerclock,
158 	dfly_recalculate_estcpu,
159 	dfly_resetpriority,
160 	dfly_forking,
161 	dfly_exiting,
162 	dfly_uload_update,
163 	NULL,			/* setcpumask not supported */
164 	dfly_yield,
165 	dfly_changedcpu
166 };
167 
168 /*
169  * We have NQS (32) run queues per scheduling class.  For the normal
170  * class, there are 128 priorities scaled onto these 32 queues.  New
171  * processes are added to the last entry in each queue, and processes
172  * are selected for running by taking them from the head and maintaining
173  * a simple FIFO arrangement.  Realtime and Idle priority processes have
174  * and explicit 0-31 priority which maps directly onto their class queue
175  * index.  When a queue has something in it, the corresponding bit is
176  * set in the queuebits variable, allowing a single read to determine
177  * the state of all 32 queues and then a ffs() to find the first busy
178  * queue.
179  */
180 static cpumask_t dfly_curprocmask = -1;	/* currently running a user process */
181 static cpumask_t dfly_rdyprocmask;	/* ready to accept a user process */
182 static volatile int dfly_scancpu;
183 static volatile int dfly_ucount;	/* total running on whole system */
184 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
185 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
186 static struct sysctl_oid *usched_dfly_sysctl_tree;
187 
188 /* Debug info exposed through debug.* sysctl */
189 
190 static int usched_dfly_debug = -1;
191 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
192 	   &usched_dfly_debug, 0,
193 	   "Print debug information for this pid");
194 
195 static int usched_dfly_pid_debug = -1;
196 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
197 	   &usched_dfly_pid_debug, 0,
198 	   "Print KTR debug information for this pid");
199 
200 static int usched_dfly_chooser = 0;
201 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
202 	   &usched_dfly_chooser, 0,
203 	   "Print KTR debug information for this pid");
204 
205 /*
206  * Tunning usched_dfly - configurable through kern.usched_dfly.
207  *
208  * weight1 - Tries to keep threads on their current cpu.  If you
209  *	     make this value too large the scheduler will not be
210  *	     able to load-balance large loads.
211  *
212  * weight2 - If non-zero, detects thread pairs undergoing synchronous
213  *	     communications and tries to move them closer together.
214  *	     Behavior is adjusted by bit 4 of features (0x10).
215  *
216  *	     WARNING!  Weight2 is a ridiculously sensitive parameter,
217  *	     a small value is recommended.
218  *
219  * weight3 - Weighting based on the number of recently runnable threads
220  *	     on the userland scheduling queue (ignoring their loads).
221  *	     A nominal value here prevents high-priority (low-load)
222  *	     threads from accumulating on one cpu core when other
223  *	     cores are available.
224  *
225  *	     This value should be left fairly small relative to weight1
226  *	     and weight4.
227  *
228  * weight4 - Weighting based on other cpu queues being available
229  *	     or running processes with higher lwp_priority's.
230  *
231  *	     This allows a thread to migrate to another nearby cpu if it
232  *	     is unable to run on the current cpu based on the other cpu
233  *	     being idle or running a lower priority (higher lwp_priority)
234  *	     thread.  This value should be large enough to override weight1
235  *
236  * features - These flags can be set or cleared to enable or disable various
237  *	      features.
238  *
239  *	      0x01	Enable idle-cpu pulling			(default)
240  *	      0x02	Enable proactive pushing		(default)
241  *	      0x04	Enable rebalancing rover		(default)
242  *	      0x08	Enable more proactive pushing		(default)
243  *	      0x10	(flip weight2 limit on same cpu)	(default)
244  *	      0x20	choose best cpu for forked process
245  *	      0x40	choose current cpu for forked process
246  *	      0x80	choose random cpu for forked process	(default)
247  */
248 static int usched_dfly_smt = 0;
249 static int usched_dfly_cache_coherent = 0;
250 static int usched_dfly_weight1 = 200;	/* keep thread on current cpu */
251 static int usched_dfly_weight2 = 180;	/* synchronous peer's current cpu */
252 static int usched_dfly_weight3 = 40;	/* number of threads on queue */
253 static int usched_dfly_weight4 = 160;	/* availability of idle cores */
254 static int usched_dfly_features = 0x8F;	/* allow pulls */
255 static int usched_dfly_fast_resched = 0;/* delta priority / resched */
256 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */
257 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
258 static int usched_dfly_decay = 8;
259 
260 /* KTR debug printings */
261 
262 KTR_INFO_MASTER(usched);
263 
264 #if !defined(KTR_USCHED_DFLY)
265 #define	KTR_USCHED_DFLY	KTR_ALL
266 #endif
267 
268 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
269     "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
270     pid_t pid, int old_cpuid, int curr);
271 
272 /*
273  * This function is called when the kernel intends to return to userland.
274  * It is responsible for making the thread the current designated userland
275  * thread for this cpu, blocking if necessary.
276  *
277  * The kernel will not depress our LWKT priority until after we return,
278  * in case we have to shove over to another cpu.
279  *
280  * We must determine our thread's disposition before we switch away.  This
281  * is very sensitive code.
282  *
283  * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
284  * TO ANOTHER CPU!  Because most of the kernel assumes that no migration will
285  * occur, this function is called only under very controlled circumstances.
286  */
287 static void
288 dfly_acquire_curproc(struct lwp *lp)
289 {
290 	globaldata_t gd;
291 	dfly_pcpu_t dd;
292 	dfly_pcpu_t rdd;
293 	thread_t td;
294 	int force_resched;
295 
296 	/*
297 	 * Make sure we aren't sitting on a tsleep queue.
298 	 */
299 	td = lp->lwp_thread;
300 	crit_enter_quick(td);
301 	if (td->td_flags & TDF_TSLEEPQ)
302 		tsleep_remove(td);
303 	dfly_recalculate_estcpu(lp);
304 
305 	gd = mycpu;
306 	dd = &dfly_pcpu[gd->gd_cpuid];
307 
308 	/*
309 	 * Process any pending interrupts/ipi's, then handle reschedule
310 	 * requests.  dfly_release_curproc() will try to assign a new
311 	 * uschedcp that isn't us and otherwise NULL it out.
312 	 */
313 	force_resched = 0;
314 	if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
315 	    lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
316 		force_resched = 1;
317 	}
318 
319 	if (user_resched_wanted()) {
320 		if (dd->uschedcp == lp)
321 			force_resched = 1;
322 		clear_user_resched();
323 		dfly_release_curproc(lp);
324 	}
325 
326 	/*
327 	 * Loop until we are the current user thread.
328 	 *
329 	 * NOTE: dd spinlock not held at top of loop.
330 	 */
331 	if (dd->uschedcp == lp)
332 		lwkt_yield_quick();
333 
334 	while (dd->uschedcp != lp) {
335 		lwkt_yield_quick();
336 
337 		spin_lock(&dd->spin);
338 
339 		/*
340 		 * We are not or are no longer the current lwp and a forced
341 		 * reschedule was requested.  Figure out the best cpu to
342 		 * run on (our current cpu will be given significant weight).
343 		 *
344 		 * (if a reschedule was not requested we want to move this
345 		 *  step after the uschedcp tests).
346 		 */
347 		if (force_resched &&
348 		    (usched_dfly_features & 0x08) &&
349 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
350 			dfly_changeqcpu_locked(lp, dd, rdd);
351 			spin_unlock(&dd->spin);
352 			lwkt_deschedule(lp->lwp_thread);
353 			dfly_setrunqueue_dd(rdd, lp);
354 			lwkt_switch();
355 			gd = mycpu;
356 			dd = &dfly_pcpu[gd->gd_cpuid];
357 			continue;
358 		}
359 
360 		/*
361 		 * Either no reschedule was requested or the best queue was
362 		 * dd, and no current process has been selected.  We can
363 		 * trivially become the current lwp on the current cpu.
364 		 */
365 		if (dd->uschedcp == NULL) {
366 			atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
367 			dd->uschedcp = lp;
368 			dd->upri = lp->lwp_priority;
369 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
370 			spin_unlock(&dd->spin);
371 			break;
372 		}
373 
374 		/*
375 		 * Can we steal the current designated user thread?
376 		 *
377 		 * If we do the other thread will stall when it tries to
378 		 * return to userland, possibly rescheduling elsewhere.
379 		 *
380 		 * It is important to do a masked test to avoid the edge
381 		 * case where two near-equal-priority threads are constantly
382 		 * interrupting each other.
383 		 *
384 		 * In the exact match case another thread has already gained
385 		 * uschedcp and lowered its priority, if we steal it the
386 		 * other thread will stay stuck on the LWKT runq and not
387 		 * push to another cpu.  So don't steal on equal-priority even
388 		 * though it might appear to be more beneficial due to not
389 		 * having to switch back to the other thread's context.
390 		 *
391 		 * usched_dfly_fast_resched requires that two threads be
392 		 * significantly far apart in priority in order to interrupt.
393 		 *
394 		 * If better but not sufficiently far apart, the current
395 		 * uschedcp will be interrupted at the next scheduler clock.
396 		 */
397 		if (dd->uschedcp &&
398 		   (dd->upri & ~PPQMASK) >
399 		   (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
400 			dd->uschedcp = lp;
401 			dd->upri = lp->lwp_priority;
402 			KKASSERT(lp->lwp_qcpu == dd->cpuid);
403 			spin_unlock(&dd->spin);
404 			break;
405 		}
406 		/*
407 		 * We are not the current lwp, figure out the best cpu
408 		 * to run on (our current cpu will be given significant
409 		 * weight).  Loop on cpu change.
410 		 */
411 		if ((usched_dfly_features & 0x02) &&
412 		    force_resched == 0 &&
413 		    (rdd = dfly_choose_best_queue(lp)) != dd) {
414 			dfly_changeqcpu_locked(lp, dd, rdd);
415 			spin_unlock(&dd->spin);
416 			lwkt_deschedule(lp->lwp_thread);
417 			dfly_setrunqueue_dd(rdd, lp);
418 			lwkt_switch();
419 			gd = mycpu;
420 			dd = &dfly_pcpu[gd->gd_cpuid];
421 			continue;
422 		}
423 
424 		/*
425 		 * We cannot become the current lwp, place the lp on the
426 		 * run-queue of this or another cpu and deschedule ourselves.
427 		 *
428 		 * When we are reactivated we will have another chance.
429 		 *
430 		 * Reload after a switch or setrunqueue/switch possibly
431 		 * moved us to another cpu.
432 		 */
433 		spin_unlock(&dd->spin);
434 		lwkt_deschedule(lp->lwp_thread);
435 		dfly_setrunqueue_dd(dd, lp);
436 		lwkt_switch();
437 		gd = mycpu;
438 		dd = &dfly_pcpu[gd->gd_cpuid];
439 	}
440 
441 	/*
442 	 * Make sure upri is synchronized, then yield to LWKT threads as
443 	 * needed before returning.  This could result in another reschedule.
444 	 * XXX
445 	 */
446 	crit_exit_quick(td);
447 
448 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
449 }
450 
451 /*
452  * DFLY_RELEASE_CURPROC
453  *
454  * This routine detaches the current thread from the userland scheduler,
455  * usually because the thread needs to run or block in the kernel (at
456  * kernel priority) for a while.
457  *
458  * This routine is also responsible for selecting a new thread to
459  * make the current thread.
460  *
461  * NOTE: This implementation differs from the dummy example in that
462  * dfly_select_curproc() is able to select the current process, whereas
463  * dummy_select_curproc() is not able to select the current process.
464  * This means we have to NULL out uschedcp.
465  *
466  * Additionally, note that we may already be on a run queue if releasing
467  * via the lwkt_switch() in dfly_setrunqueue().
468  */
469 static void
470 dfly_release_curproc(struct lwp *lp)
471 {
472 	globaldata_t gd = mycpu;
473 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
474 
475 	/*
476 	 * Make sure td_wakefromcpu is defaulted.  This will be overwritten
477 	 * by wakeup().
478 	 */
479 	if (dd->uschedcp == lp) {
480 		KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
481 		spin_lock(&dd->spin);
482 		if (dd->uschedcp == lp) {
483 			dd->uschedcp = NULL;	/* don't let lp be selected */
484 			dd->upri = PRIBASE_NULL;
485 			atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
486 			spin_unlock(&dd->spin);
487 			dfly_select_curproc(gd);
488 		} else {
489 			spin_unlock(&dd->spin);
490 		}
491 	}
492 }
493 
494 /*
495  * DFLY_SELECT_CURPROC
496  *
497  * Select a new current process for this cpu and clear any pending user
498  * reschedule request.  The cpu currently has no current process.
499  *
500  * This routine is also responsible for equal-priority round-robining,
501  * typically triggered from dfly_schedulerclock().  In our dummy example
502  * all the 'user' threads are LWKT scheduled all at once and we just
503  * call lwkt_switch().
504  *
505  * The calling process is not on the queue and cannot be selected.
506  */
507 static
508 void
509 dfly_select_curproc(globaldata_t gd)
510 {
511 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
512 	struct lwp *nlp;
513 	int cpuid = gd->gd_cpuid;
514 
515 	crit_enter_gd(gd);
516 
517 	spin_lock(&dd->spin);
518 	nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
519 
520 	if (nlp) {
521 		atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
522 		dd->upri = nlp->lwp_priority;
523 		dd->uschedcp = nlp;
524 #if 0
525 		dd->rrcount = 0;		/* reset round robin */
526 #endif
527 		spin_unlock(&dd->spin);
528 		lwkt_acquire(nlp->lwp_thread);
529 		lwkt_schedule(nlp->lwp_thread);
530 	} else {
531 		spin_unlock(&dd->spin);
532 	}
533 	crit_exit_gd(gd);
534 }
535 
536 /*
537  * Place the specified lwp on the user scheduler's run queue.  This routine
538  * must be called with the thread descheduled.  The lwp must be runnable.
539  * It must not be possible for anyone else to explicitly schedule this thread.
540  *
541  * The thread may be the current thread as a special case.
542  */
543 static void
544 dfly_setrunqueue(struct lwp *lp)
545 {
546 	dfly_pcpu_t dd;
547 	dfly_pcpu_t rdd;
548 
549 	/*
550 	 * First validate the process LWKT state.
551 	 */
552 	KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
553 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
554 	    ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
555 	     lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
556 	KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
557 
558 	/*
559 	 * NOTE: dd/rdd do not necessarily represent the current cpu.
560 	 *	 Instead they may represent the cpu the thread was last
561 	 *	 scheduled on or inherited by its parent.
562 	 */
563 	dd = &dfly_pcpu[lp->lwp_qcpu];
564 	rdd = dd;
565 
566 	/*
567 	 * This process is not supposed to be scheduled anywhere or assigned
568 	 * as the current process anywhere.  Assert the condition.
569 	 */
570 	KKASSERT(rdd->uschedcp != lp);
571 
572 	/*
573 	 * Ok, we have to setrunqueue some target cpu and request a reschedule
574 	 * if necessary.
575 	 *
576 	 * We have to choose the best target cpu.  It might not be the current
577 	 * target even if the current cpu has no running user thread (for
578 	 * example, because the current cpu might be a hyperthread and its
579 	 * sibling has a thread assigned).
580 	 *
581 	 * If we just forked it is most optimal to run the child on the same
582 	 * cpu just in case the parent decides to wait for it (thus getting
583 	 * off that cpu).  As long as there is nothing else runnable on the
584 	 * cpu, that is.  If we did this unconditionally a parent forking
585 	 * multiple children before waiting (e.g. make -j N) leaves other
586 	 * cpus idle that could be working.
587 	 */
588 	if (lp->lwp_forked) {
589 		lp->lwp_forked = 0;
590 		if (usched_dfly_features & 0x20)
591 			rdd = dfly_choose_best_queue(lp);
592 		else if (usched_dfly_features & 0x40)
593 			rdd = &dfly_pcpu[lp->lwp_qcpu];
594 		else if (usched_dfly_features & 0x80)
595 			rdd = dfly_choose_queue_simple(rdd, lp);
596 		else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
597 			rdd = dfly_choose_best_queue(lp);
598 		else
599 			rdd = &dfly_pcpu[lp->lwp_qcpu];
600 	} else {
601 		rdd = dfly_choose_best_queue(lp);
602 		/* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
603 	}
604 	if (lp->lwp_qcpu != rdd->cpuid) {
605 		spin_lock(&dd->spin);
606 		dfly_changeqcpu_locked(lp, dd, rdd);
607 		spin_unlock(&dd->spin);
608 	}
609 	dfly_setrunqueue_dd(rdd, lp);
610 }
611 
612 /*
613  * Change qcpu to rdd->cpuid.  The dd the lp is CURRENTLY on must be
614  * spin-locked on-call.  rdd does not have to be.
615  */
616 static void
617 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
618 {
619 	if (lp->lwp_qcpu != rdd->cpuid) {
620 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
621 			atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
622 			atomic_add_int(&dd->uload, -lp->lwp_uload);
623 			atomic_add_int(&dd->ucount, -1);
624 			atomic_add_int(&dfly_ucount, -1);
625 		}
626 		lp->lwp_qcpu = rdd->cpuid;
627 	}
628 }
629 
630 /*
631  * Place lp on rdd's runqueue.  Nothing is locked on call.  This function
632  * also performs all necessary ancillary notification actions.
633  */
634 static void
635 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
636 {
637 	globaldata_t rgd;
638 
639 	/*
640 	 * We might be moving the lp to another cpu's run queue, and once
641 	 * on the runqueue (even if it is our cpu's), another cpu can rip
642 	 * it away from us.
643 	 *
644 	 * TDF_MIGRATING might already be set if this is part of a
645 	 * remrunqueue+setrunqueue sequence.
646 	 */
647 	if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
648 		lwkt_giveaway(lp->lwp_thread);
649 
650 	rgd = globaldata_find(rdd->cpuid);
651 
652 	/*
653 	 * We lose control of the lp the moment we release the spinlock
654 	 * after having placed it on the queue.  i.e. another cpu could pick
655 	 * it up, or it could exit, or its priority could be further
656 	 * adjusted, or something like that.
657 	 *
658 	 * WARNING! rdd can point to a foreign cpu!
659 	 */
660 	spin_lock(&rdd->spin);
661 	dfly_setrunqueue_locked(rdd, lp);
662 
663 	/*
664 	 * Potentially interrupt the currently-running thread
665 	 */
666 	if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
667 		/*
668 		 * Currently running thread is better or same, do not
669 		 * interrupt.
670 		 */
671 		spin_unlock(&rdd->spin);
672 	} else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
673 		   usched_dfly_fast_resched) {
674 		/*
675 		 * Currently running thread is not better, but not so bad
676 		 * that we need to interrupt it.  Let it run for one more
677 		 * scheduler tick.
678 		 */
679 		if (rdd->uschedcp &&
680 		    rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
681 			rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
682 		}
683 		spin_unlock(&rdd->spin);
684 	} else if (rgd == mycpu) {
685 		/*
686 		 * We should interrupt the currently running thread, which
687 		 * is on the current cpu.
688 		 */
689 		spin_unlock(&rdd->spin);
690 		if (rdd->uschedcp == NULL) {
691 			wakeup_mycpu(&rdd->helper_thread); /* XXX */
692 			need_user_resched();
693 		} else {
694 			need_user_resched();
695 		}
696 	} else {
697 		/*
698 		 * We should interrupt the currently running thread, which
699 		 * is on a different cpu.
700 		 */
701 		spin_unlock(&rdd->spin);
702 		lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
703 	}
704 }
705 
706 /*
707  * This routine is called from a systimer IPI.  It MUST be MP-safe and
708  * the BGL IS NOT HELD ON ENTRY.  This routine is called at ESTCPUFREQ on
709  * each cpu.
710  */
711 static
712 void
713 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
714 {
715 	globaldata_t gd = mycpu;
716 	dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
717 
718 	/*
719 	 * Spinlocks also hold a critical section so there should not be
720 	 * any active.
721 	 */
722 	KKASSERT(gd->gd_spinlocks == 0);
723 
724 	if (lp == NULL)
725 		return;
726 
727 	/*
728 	 * Do we need to round-robin?  We round-robin 10 times a second.
729 	 * This should only occur for cpu-bound batch processes.
730 	 */
731 	if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
732 		lp->lwp_thread->td_wakefromcpu = -1;
733 		need_user_resched();
734 	}
735 
736 	/*
737 	 * Adjust estcpu upward using a real time equivalent calculation,
738 	 * and recalculate lp's priority.
739 	 */
740 	lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
741 	dfly_resetpriority(lp);
742 
743 	/*
744 	 * Rebalance two cpus every 8 ticks, pulling the worst thread
745 	 * from the worst cpu's queue into a rotating cpu number.
746 	 *
747 	 * This mechanic is needed because the push algorithms can
748 	 * steady-state in an non-optimal configuration.  We need to mix it
749 	 * up a little, even if it means breaking up a paired thread, so
750 	 * the push algorithms can rebalance the degenerate conditions.
751 	 * This portion of the algorithm exists to ensure stability at the
752 	 * selected weightings.
753 	 *
754 	 * Because we might be breaking up optimal conditions we do not want
755 	 * to execute this too quickly, hence we only rebalance approximately
756 	 * ~7-8 times per second.  The push's, on the otherhand, are capable
757 	 * moving threads to other cpus at a much higher rate.
758 	 *
759 	 * We choose the most heavily loaded thread from the worst queue
760 	 * in order to ensure that multiple heavy-weight threads on the same
761 	 * queue get broken up, and also because these threads are the most
762 	 * likely to be able to remain in place.  Hopefully then any pairings,
763 	 * if applicable, migrate to where these threads are.
764 	 */
765 	if ((usched_dfly_features & 0x04) &&
766 	    ((u_int)sched_ticks & 7) == 0 &&
767 	    (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
768 		/*
769 		 * Our cpu is up.
770 		 */
771 		struct lwp *nlp;
772 		dfly_pcpu_t rdd;
773 
774 		rdd = dfly_choose_worst_queue(dd);
775 		if (rdd) {
776 			spin_lock(&dd->spin);
777 			if (spin_trylock(&rdd->spin)) {
778 				nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
779 				spin_unlock(&rdd->spin);
780 				if (nlp == NULL)
781 					spin_unlock(&dd->spin);
782 			} else {
783 				spin_unlock(&dd->spin);
784 				nlp = NULL;
785 			}
786 		} else {
787 			nlp = NULL;
788 		}
789 		/* dd->spin held if nlp != NULL */
790 
791 		/*
792 		 * Either schedule it or add it to our queue.
793 		 */
794 		if (nlp &&
795 		    (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
796 			atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
797 			dd->upri = nlp->lwp_priority;
798 			dd->uschedcp = nlp;
799 #if 0
800 			dd->rrcount = 0;	/* reset round robin */
801 #endif
802 			spin_unlock(&dd->spin);
803 			lwkt_acquire(nlp->lwp_thread);
804 			lwkt_schedule(nlp->lwp_thread);
805 		} else if (nlp) {
806 			dfly_setrunqueue_locked(dd, nlp);
807 			spin_unlock(&dd->spin);
808 		}
809 	}
810 }
811 
812 /*
813  * Called from acquire and from kern_synch's one-second timer (one of the
814  * callout helper threads) with a critical section held.
815  *
816  * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
817  * overall system load.
818  *
819  * Note that no recalculation occurs for a process which sleeps and wakes
820  * up in the same tick.  That is, a system doing thousands of context
821  * switches per second will still only do serious estcpu calculations
822  * ESTCPUFREQ times per second.
823  */
824 static
825 void
826 dfly_recalculate_estcpu(struct lwp *lp)
827 {
828 	globaldata_t gd = mycpu;
829 	sysclock_t cpbase;
830 	sysclock_t ttlticks;
831 	int estcpu;
832 	int decay_factor;
833 	int ucount;
834 
835 	/*
836 	 * We have to subtract periodic to get the last schedclock
837 	 * timeout time, otherwise we would get the upcoming timeout.
838 	 * Keep in mind that a process can migrate between cpus and
839 	 * while the scheduler clock should be very close, boundary
840 	 * conditions could lead to a small negative delta.
841 	 */
842 	cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
843 
844 	if (lp->lwp_slptime > 1) {
845 		/*
846 		 * Too much time has passed, do a coarse correction.
847 		 */
848 		lp->lwp_estcpu = lp->lwp_estcpu >> 1;
849 		dfly_resetpriority(lp);
850 		lp->lwp_cpbase = cpbase;
851 		lp->lwp_cpticks = 0;
852 		lp->lwp_estfast = 0;
853 	} else if (lp->lwp_cpbase != cpbase) {
854 		/*
855 		 * Adjust estcpu if we are in a different tick.  Don't waste
856 		 * time if we are in the same tick.
857 		 *
858 		 * First calculate the number of ticks in the measurement
859 		 * interval.  The ttlticks calculation can wind up 0 due to
860 		 * a bug in the handling of lwp_slptime  (as yet not found),
861 		 * so make sure we do not get a divide by 0 panic.
862 		 */
863 		ttlticks = (cpbase - lp->lwp_cpbase) /
864 			   gd->gd_schedclock.periodic;
865 		if ((ssysclock_t)ttlticks < 0) {
866 			ttlticks = 0;
867 			lp->lwp_cpbase = cpbase;
868 		}
869 		if (ttlticks == 0)
870 			return;
871 		updatepcpu(lp, lp->lwp_cpticks, ttlticks);
872 
873 		/*
874 		 * Calculate the percentage of one cpu being used then
875 		 * compensate for any system load in excess of ncpus.
876 		 *
877 		 * For example, if we have 8 cores and 16 running cpu-bound
878 		 * processes then all things being equal each process will
879 		 * get 50% of one cpu.  We need to pump this value back
880 		 * up to 100% so the estcpu calculation properly adjusts
881 		 * the process's dynamic priority.
882 		 *
883 		 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
884 		 */
885 		estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
886 		ucount = dfly_ucount;
887 		if (ucount > ncpus) {
888 			estcpu += estcpu * (ucount - ncpus) / ncpus;
889 		}
890 
891 		if (usched_dfly_debug == lp->lwp_proc->p_pid) {
892 			kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
893 				lp->lwp_proc->p_pid, lp,
894 				estcpu, lp->lwp_estcpu,
895 				lp->lwp_cpticks, ttlticks);
896 		}
897 
898 		/*
899 		 * Adjust lp->lwp_esetcpu.  The decay factor determines how
900 		 * quickly lwp_estcpu collapses to its realtime calculation.
901 		 * A slower collapse gives us a more accurate number over
902 		 * the long term but can create problems with bursty threads
903 		 * or threads which become cpu hogs.
904 		 *
905 		 * To solve this problem, newly started lwps and lwps which
906 		 * are restarting after having been asleep for a while are
907 		 * given a much, much faster decay in order to quickly
908 		 * detect whether they become cpu-bound.
909 		 *
910 		 * NOTE: p_nice is accounted for in dfly_resetpriority(),
911 		 *	 and not here, but we must still ensure that a
912 		 *	 cpu-bound nice -20 process does not completely
913 		 *	 override a cpu-bound nice +20 process.
914 		 *
915 		 * NOTE: We must use ESTCPULIM() here to deal with any
916 		 *	 overshoot.
917 		 */
918 		decay_factor = usched_dfly_decay;
919 		if (decay_factor < 1)
920 			decay_factor = 1;
921 		if (decay_factor > 1024)
922 			decay_factor = 1024;
923 
924 		if (lp->lwp_estfast < usched_dfly_decay) {
925 			++lp->lwp_estfast;
926 			lp->lwp_estcpu = ESTCPULIM(
927 				(lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
928 				(lp->lwp_estfast + 1));
929 		} else {
930 			lp->lwp_estcpu = ESTCPULIM(
931 				(lp->lwp_estcpu * decay_factor + estcpu) /
932 				(decay_factor + 1));
933 		}
934 
935 		if (usched_dfly_debug == lp->lwp_proc->p_pid)
936 			kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
937 		dfly_resetpriority(lp);
938 		lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
939 		lp->lwp_cpticks = 0;
940 	}
941 }
942 
943 /*
944  * Compute the priority of a process when running in user mode.
945  * Arrange to reschedule if the resulting priority is better
946  * than that of the current process.
947  *
948  * This routine may be called with any process.
949  *
950  * This routine is called by fork1() for initial setup with the process
951  * of the run queue, and also may be called normally with the process on or
952  * off the run queue.
953  */
954 static void
955 dfly_resetpriority(struct lwp *lp)
956 {
957 	dfly_pcpu_t rdd;
958 	int newpriority;
959 	u_short newrqtype;
960 	int rcpu;
961 	int checkpri;
962 	int estcpu;
963 	int delta_uload;
964 
965 	crit_enter();
966 
967 	/*
968 	 * Lock the scheduler (lp) belongs to.  This can be on a different
969 	 * cpu.  Handle races.  This loop breaks out with the appropriate
970 	 * rdd locked.
971 	 */
972 	for (;;) {
973 		rcpu = lp->lwp_qcpu;
974 		cpu_ccfence();
975 		rdd = &dfly_pcpu[rcpu];
976 		spin_lock(&rdd->spin);
977 		if (rcpu == lp->lwp_qcpu)
978 			break;
979 		spin_unlock(&rdd->spin);
980 	}
981 
982 	/*
983 	 * Calculate the new priority and queue type
984 	 */
985 	newrqtype = lp->lwp_rtprio.type;
986 
987 	switch(newrqtype) {
988 	case RTP_PRIO_REALTIME:
989 	case RTP_PRIO_FIFO:
990 		newpriority = PRIBASE_REALTIME +
991 			     (lp->lwp_rtprio.prio & PRIMASK);
992 		break;
993 	case RTP_PRIO_NORMAL:
994 		/*
995 		 *
996 		 */
997 		estcpu = lp->lwp_estcpu;
998 
999 		/*
1000 		 * p_nice piece		Adds (0-40) * 2		0-80
1001 		 * estcpu		Adds 16384  * 4 / 512   0-128
1002 		 */
1003 		newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1004 		newpriority += estcpu * PPQ / ESTCPUPPQ;
1005 		newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1006 			      NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1007 		newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1008 		break;
1009 	case RTP_PRIO_IDLE:
1010 		newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1011 		break;
1012 	case RTP_PRIO_THREAD:
1013 		newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1014 		break;
1015 	default:
1016 		panic("Bad RTP_PRIO %d", newrqtype);
1017 		/* NOT REACHED */
1018 	}
1019 
1020 	/*
1021 	 * The LWKT scheduler doesn't dive usched structures, give it a hint
1022 	 * on the relative priority of user threads running in the kernel.
1023 	 * The LWKT scheduler will always ensure that a user thread running
1024 	 * in the kernel will get cpu some time, regardless of its upri,
1025 	 * but can decide not to instantly switch from one kernel or user
1026 	 * mode user thread to a kernel-mode user thread when it has a less
1027 	 * desireable user priority.
1028 	 *
1029 	 * td_upri has normal sense (higher values are more desireable), so
1030 	 * negate it.
1031 	 */
1032 	lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1033 
1034 	/*
1035 	 * The newpriority incorporates the queue type so do a simple masked
1036 	 * check to determine if the process has moved to another queue.  If
1037 	 * it has, and it is currently on a run queue, then move it.
1038 	 *
1039 	 * Since uload is ~PPQMASK masked, no modifications are necessary if
1040 	 * we end up in the same run queue.
1041 	 */
1042 	if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1043 		if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1044 			dfly_remrunqueue_locked(rdd, lp);
1045 			lp->lwp_priority = newpriority;
1046 			lp->lwp_rqtype = newrqtype;
1047 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1048 			dfly_setrunqueue_locked(rdd, lp);
1049 			checkpri = 1;
1050 		} else {
1051 			lp->lwp_priority = newpriority;
1052 			lp->lwp_rqtype = newrqtype;
1053 			lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1054 			checkpri = 0;
1055 		}
1056 	} else {
1057 		/*
1058 		 * In the same PPQ, uload cannot change.
1059 		 */
1060 		lp->lwp_priority = newpriority;
1061 		checkpri = 1;
1062 		rcpu = -1;
1063 	}
1064 
1065 	/*
1066 	 * Adjust effective load.
1067 	 *
1068 	 * Calculate load then scale up or down geometrically based on p_nice.
1069 	 * Processes niced up (positive) are less important, and processes
1070 	 * niced downard (negative) are more important.  The higher the uload,
1071 	 * the more important the thread.
1072 	 */
1073 	/* 0-511, 0-100% cpu */
1074 	delta_uload = lp->lwp_estcpu / NQS;
1075 	delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1076 
1077 
1078 	delta_uload -= lp->lwp_uload;
1079 	lp->lwp_uload += delta_uload;
1080 	if (lp->lwp_mpflags & LWP_MP_ULOAD)
1081 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1082 
1083 	/*
1084 	 * Determine if we need to reschedule the target cpu.  This only
1085 	 * occurs if the LWP is already on a scheduler queue, which means
1086 	 * that idle cpu notification has already occured.  At most we
1087 	 * need only issue a need_user_resched() on the appropriate cpu.
1088 	 *
1089 	 * The LWP may be owned by a CPU different from the current one,
1090 	 * in which case dd->uschedcp may be modified without an MP lock
1091 	 * or a spinlock held.  The worst that happens is that the code
1092 	 * below causes a spurious need_user_resched() on the target CPU
1093 	 * and dd->pri to be wrong for a short period of time, both of
1094 	 * which are harmless.
1095 	 *
1096 	 * If checkpri is 0 we are adjusting the priority of the current
1097 	 * process, possibly higher (less desireable), so ignore the upri
1098 	 * check which will fail in that case.
1099 	 */
1100 	if (rcpu >= 0) {
1101 		if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1102 		    (checkpri == 0 ||
1103 		     (rdd->upri & ~PRIMASK) >
1104 		     (lp->lwp_priority & ~PRIMASK))) {
1105 			if (rcpu == mycpu->gd_cpuid) {
1106 				spin_unlock(&rdd->spin);
1107 				need_user_resched();
1108 			} else {
1109 				spin_unlock(&rdd->spin);
1110 				lwkt_send_ipiq(globaldata_find(rcpu),
1111 					       dfly_need_user_resched_remote,
1112 					       NULL);
1113 			}
1114 		} else {
1115 			spin_unlock(&rdd->spin);
1116 		}
1117 	} else {
1118 		spin_unlock(&rdd->spin);
1119 	}
1120 	crit_exit();
1121 }
1122 
1123 static
1124 void
1125 dfly_yield(struct lwp *lp)
1126 {
1127 #if 0
1128 	/* FUTURE (or something similar) */
1129 	switch(lp->lwp_rqtype) {
1130 	case RTP_PRIO_NORMAL:
1131 		lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
1132 		break;
1133 	default:
1134 		break;
1135 	}
1136 #endif
1137         need_user_resched();
1138 }
1139 
1140 /*
1141  * Thread was forcefully migrated to another cpu.  Normally forced migrations
1142  * are used for iterations and the kernel returns to the original cpu before
1143  * returning and this is not needed.  However, if the kernel migrates a
1144  * thread to another cpu and wants to leave it there, it has to call this
1145  * scheduler helper.
1146  *
1147  * Note that the lwkt_migratecpu() function also released the thread, so
1148  * we don't have to worry about that.
1149  */
1150 static
1151 void
1152 dfly_changedcpu(struct lwp *lp)
1153 {
1154 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1155 	dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid];
1156 
1157 	if (dd != rdd) {
1158 		spin_lock(&dd->spin);
1159 		dfly_changeqcpu_locked(lp, dd, rdd);
1160 		spin_unlock(&dd->spin);
1161 	}
1162 }
1163 
1164 /*
1165  * Called from fork1() when a new child process is being created.
1166  *
1167  * Give the child process an initial estcpu that is more batch then
1168  * its parent and dock the parent for the fork (but do not
1169  * reschedule the parent).
1170  *
1171  * fast
1172  *
1173  * XXX lwp should be "spawning" instead of "forking"
1174  */
1175 static void
1176 dfly_forking(struct lwp *plp, struct lwp *lp)
1177 {
1178 	/*
1179 	 * Put the child 4 queue slots (out of 32) higher than the parent
1180 	 * (less desireable than the parent).
1181 	 */
1182 	lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1183 	lp->lwp_forked = 1;
1184 	lp->lwp_estfast = 0;
1185 
1186 	/*
1187 	 * Dock the parent a cost for the fork, protecting us from fork
1188 	 * bombs.  If the parent is forking quickly make the child more
1189 	 * batchy.
1190 	 */
1191 	plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1192 }
1193 
1194 /*
1195  * Called when a lwp is being removed from this scheduler, typically
1196  * during lwp_exit().  We have to clean out any ULOAD accounting before
1197  * we can let the lp go.  The dd->spin lock is not needed for uload
1198  * updates.
1199  *
1200  * Scheduler dequeueing has already occurred, no further action in that
1201  * regard is needed.
1202  */
1203 static void
1204 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1205 {
1206 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1207 
1208 	if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1209 		atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1210 		atomic_add_int(&dd->uload, -lp->lwp_uload);
1211 		atomic_add_int(&dd->ucount, -1);
1212 		atomic_add_int(&dfly_ucount, -1);
1213 	}
1214 }
1215 
1216 /*
1217  * This function cannot block in any way, but spinlocks are ok.
1218  *
1219  * Update the uload based on the state of the thread (whether it is going
1220  * to sleep or running again).  The uload is meant to be a longer-term
1221  * load and not an instantanious load.
1222  */
1223 static void
1224 dfly_uload_update(struct lwp *lp)
1225 {
1226 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1227 
1228 	if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1229 		if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1230 			spin_lock(&dd->spin);
1231 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1232 				atomic_set_int(&lp->lwp_mpflags,
1233 					       LWP_MP_ULOAD);
1234 				atomic_add_int(&dd->uload, lp->lwp_uload);
1235 				atomic_add_int(&dd->ucount, 1);
1236 				atomic_add_int(&dfly_ucount, 1);
1237 			}
1238 			spin_unlock(&dd->spin);
1239 		}
1240 	} else if (lp->lwp_slptime > 0) {
1241 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1242 			spin_lock(&dd->spin);
1243 			if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1244 				atomic_clear_int(&lp->lwp_mpflags,
1245 						 LWP_MP_ULOAD);
1246 				atomic_add_int(&dd->uload, -lp->lwp_uload);
1247 				atomic_add_int(&dd->ucount, -1);
1248 				atomic_add_int(&dfly_ucount, -1);
1249 			}
1250 			spin_unlock(&dd->spin);
1251 		}
1252 	}
1253 }
1254 
1255 /*
1256  * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1257  * it selects a user process and returns it.  If chklp is non-NULL and chklp
1258  * has a better or equal priority then the process that would otherwise be
1259  * chosen, NULL is returned.
1260  *
1261  * Until we fix the RUNQ code the chklp test has to be strict or we may
1262  * bounce between processes trying to acquire the current process designation.
1263  *
1264  * Must be called with rdd->spin locked.  The spinlock is left intact through
1265  * the entire routine.  dd->spin does not have to be locked.
1266  *
1267  * If worst is non-zero this function finds the worst thread instead of the
1268  * best thread (used by the schedulerclock-based rover).
1269  */
1270 static
1271 struct lwp *
1272 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1273 		       struct lwp *chklp, int worst)
1274 {
1275 	struct lwp *lp;
1276 	struct rq *q;
1277 	u_int32_t *which;
1278 	u_int32_t pri;
1279 	u_int32_t rtqbits;
1280 	u_int32_t tsqbits;
1281 	u_int32_t idqbits;
1282 
1283 	rtqbits = rdd->rtqueuebits;
1284 	tsqbits = rdd->queuebits;
1285 	idqbits = rdd->idqueuebits;
1286 
1287 	if (worst) {
1288 		if (idqbits) {
1289 			pri = bsrl(idqbits);
1290 			q = &rdd->idqueues[pri];
1291 			which = &rdd->idqueuebits;
1292 		} else if (tsqbits) {
1293 			pri = bsrl(tsqbits);
1294 			q = &rdd->queues[pri];
1295 			which = &rdd->queuebits;
1296 		} else if (rtqbits) {
1297 			pri = bsrl(rtqbits);
1298 			q = &rdd->rtqueues[pri];
1299 			which = &rdd->rtqueuebits;
1300 		} else {
1301 			return (NULL);
1302 		}
1303 		lp = TAILQ_LAST(q, rq);
1304 	} else {
1305 		if (rtqbits) {
1306 			pri = bsfl(rtqbits);
1307 			q = &rdd->rtqueues[pri];
1308 			which = &rdd->rtqueuebits;
1309 		} else if (tsqbits) {
1310 			pri = bsfl(tsqbits);
1311 			q = &rdd->queues[pri];
1312 			which = &rdd->queuebits;
1313 		} else if (idqbits) {
1314 			pri = bsfl(idqbits);
1315 			q = &rdd->idqueues[pri];
1316 			which = &rdd->idqueuebits;
1317 		} else {
1318 			return (NULL);
1319 		}
1320 		lp = TAILQ_FIRST(q);
1321 	}
1322 	KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1323 
1324 	/*
1325 	 * If the passed lwp <chklp> is reasonably close to the selected
1326 	 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1327 	 *
1328 	 * Note that we must error on the side of <chklp> to avoid bouncing
1329 	 * between threads in the acquire code.
1330 	 */
1331 	if (chklp) {
1332 		if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1333 			return(NULL);
1334 	}
1335 
1336 	KTR_COND_LOG(usched_chooseproc,
1337 	    lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1338 	    lp->lwp_proc->p_pid,
1339 	    lp->lwp_thread->td_gd->gd_cpuid,
1340 	    mycpu->gd_cpuid);
1341 
1342 	KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1343 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1344 	TAILQ_REMOVE(q, lp, lwp_procq);
1345 	--rdd->runqcount;
1346 	if (TAILQ_EMPTY(q))
1347 		*which &= ~(1 << pri);
1348 
1349 	/*
1350 	 * If we are choosing a process from rdd with the intent to
1351 	 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1352 	 * is still held.
1353 	 */
1354 	if (rdd != dd) {
1355 		if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1356 			atomic_add_int(&rdd->uload, -lp->lwp_uload);
1357 			atomic_add_int(&rdd->ucount, -1);
1358 			atomic_add_int(&dfly_ucount, -1);
1359 		}
1360 		lp->lwp_qcpu = dd->cpuid;
1361 		atomic_add_int(&dd->uload, lp->lwp_uload);
1362 		atomic_add_int(&dd->ucount, 1);
1363 		atomic_add_int(&dfly_ucount, 1);
1364 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1365 	}
1366 	return lp;
1367 }
1368 
1369 /*
1370  * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1371  *
1372  * Choose a cpu node to schedule lp on, hopefully nearby its current
1373  * node.
1374  *
1375  * We give the current node a modest advantage for obvious reasons.
1376  *
1377  * We also give the node the thread was woken up FROM a slight advantage
1378  * in order to try to schedule paired threads which synchronize/block waiting
1379  * for each other fairly close to each other.  Similarly in a network setting
1380  * this feature will also attempt to place a user process near the kernel
1381  * protocol thread that is feeding it data.  THIS IS A CRITICAL PART of the
1382  * algorithm as it heuristically groups synchronizing processes for locality
1383  * of reference in multi-socket systems.
1384  *
1385  * We check against running processes and give a big advantage if there
1386  * are none running.
1387  *
1388  * The caller will normally dfly_setrunqueue() lp on the returned queue.
1389  *
1390  * When the topology is known choose a cpu whos group has, in aggregate,
1391  * has the lowest weighted load.
1392  */
1393 static
1394 dfly_pcpu_t
1395 dfly_choose_best_queue(struct lwp *lp)
1396 {
1397 	cpumask_t wakemask;
1398 	cpumask_t mask;
1399 	cpu_node_t *cpup;
1400 	cpu_node_t *cpun;
1401 	cpu_node_t *cpub;
1402 	dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1403 	dfly_pcpu_t rdd;
1404 	int wakecpu;
1405 	int cpuid;
1406 	int n;
1407 	int count;
1408 	int load;
1409 	int lowest_load;
1410 
1411 	/*
1412 	 * When the topology is unknown choose a random cpu that is hopefully
1413 	 * idle.
1414 	 */
1415 	if (dd->cpunode == NULL)
1416 		return (dfly_choose_queue_simple(dd, lp));
1417 
1418 	/*
1419 	 * Pairing mask
1420 	 */
1421 	if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1422 		wakemask = dfly_pcpu[wakecpu].cpumask;
1423 	else
1424 		wakemask = 0;
1425 
1426 	/*
1427 	 * When the topology is known choose a cpu whos group has, in
1428 	 * aggregate, has the lowest weighted load.
1429 	 */
1430 	cpup = root_cpu_node;
1431 	rdd = dd;
1432 
1433 	while (cpup) {
1434 		/*
1435 		 * Degenerate case super-root
1436 		 */
1437 		if (cpup->child_node && cpup->child_no == 1) {
1438 			cpup = cpup->child_node;
1439 			continue;
1440 		}
1441 
1442 		/*
1443 		 * Terminal cpunode
1444 		 */
1445 		if (cpup->child_node == NULL) {
1446 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1447 			break;
1448 		}
1449 
1450 		cpub = NULL;
1451 		lowest_load = 0x7FFFFFFF;
1452 
1453 		for (n = 0; n < cpup->child_no; ++n) {
1454 			/*
1455 			 * Accumulate load information for all cpus
1456 			 * which are members of this node.
1457 			 */
1458 			cpun = &cpup->child_node[n];
1459 			mask = cpun->members & usched_global_cpumask &
1460 			       smp_active_mask & lp->lwp_cpumask;
1461 			if (mask == 0)
1462 				continue;
1463 
1464 			count = 0;
1465 			load = 0;
1466 
1467 			while (mask) {
1468 				cpuid = BSFCPUMASK(mask);
1469 				rdd = &dfly_pcpu[cpuid];
1470 				load += rdd->uload;
1471 				load += rdd->ucount * usched_dfly_weight3;
1472 
1473 				if (rdd->uschedcp == NULL &&
1474 				    rdd->runqcount == 0 &&
1475 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1476 				) {
1477 					load -= usched_dfly_weight4;
1478 				}
1479 #if 0
1480 				else if (rdd->upri > lp->lwp_priority + PPQ) {
1481 					load -= usched_dfly_weight4 / 2;
1482 				}
1483 #endif
1484 				mask &= ~CPUMASK(cpuid);
1485 				++count;
1486 			}
1487 
1488 			/*
1489 			 * Compensate if the lp is already accounted for in
1490 			 * the aggregate uload for this mask set.  We want
1491 			 * to calculate the loads as if lp were not present,
1492 			 * otherwise the calculation is bogus.
1493 			 */
1494 			if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1495 			    (dd->cpumask & cpun->members)) {
1496 				load -= lp->lwp_uload;
1497 				load -= usched_dfly_weight3;
1498 			}
1499 
1500 			load /= count;
1501 
1502 			/*
1503 			 * Advantage the cpu group (lp) is already on.
1504 			 */
1505 			if (cpun->members & dd->cpumask)
1506 				load -= usched_dfly_weight1;
1507 
1508 			/*
1509 			 * Advantage the cpu group we want to pair (lp) to,
1510 			 * but don't let it go to the exact same cpu as
1511 			 * the wakecpu target.
1512 			 *
1513 			 * We do this by checking whether cpun is a
1514 			 * terminal node or not.  All cpun's at the same
1515 			 * level will either all be terminal or all not
1516 			 * terminal.
1517 			 *
1518 			 * If it is and we match we disadvantage the load.
1519 			 * If it is and we don't match we advantage the load.
1520 			 *
1521 			 * Also note that we are effectively disadvantaging
1522 			 * all-but-one by the same amount, so it won't effect
1523 			 * the weight1 factor for the all-but-one nodes.
1524 			 */
1525 			if (cpun->members & wakemask) {
1526 				if (cpun->child_node != NULL) {
1527 					/* advantage */
1528 					load -= usched_dfly_weight2;
1529 				} else {
1530 					if (usched_dfly_features & 0x10)
1531 						load += usched_dfly_weight2;
1532 					else
1533 						load -= usched_dfly_weight2;
1534 				}
1535 			}
1536 
1537 			/*
1538 			 * Calculate the best load
1539 			 */
1540 			if (cpub == NULL || lowest_load > load ||
1541 			    (lowest_load == load &&
1542 			     (cpun->members & dd->cpumask))
1543 			) {
1544 				lowest_load = load;
1545 				cpub = cpun;
1546 			}
1547 		}
1548 		cpup = cpub;
1549 	}
1550 	if (usched_dfly_chooser)
1551 		kprintf("lp %02d->%02d %s\n",
1552 			lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1553 	return (rdd);
1554 }
1555 
1556 /*
1557  * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1558  *
1559  * Choose the worst queue close to dd's cpu node with a non-empty runq
1560  * that is NOT dd.  Also require that the moving of the highest-load thread
1561  * from rdd to dd does not cause the uload's to cross each other.
1562  *
1563  * This is used by the thread chooser when the current cpu's queues are
1564  * empty to steal a thread from another cpu's queue.  We want to offload
1565  * the most heavily-loaded queue.
1566  */
1567 static
1568 dfly_pcpu_t
1569 dfly_choose_worst_queue(dfly_pcpu_t dd)
1570 {
1571 	cpumask_t mask;
1572 	cpu_node_t *cpup;
1573 	cpu_node_t *cpun;
1574 	cpu_node_t *cpub;
1575 	dfly_pcpu_t rdd;
1576 	int cpuid;
1577 	int n;
1578 	int count;
1579 	int load;
1580 #if 0
1581 	int pri;
1582 	int hpri;
1583 #endif
1584 	int highest_load;
1585 
1586 	/*
1587 	 * When the topology is unknown choose a random cpu that is hopefully
1588 	 * idle.
1589 	 */
1590 	if (dd->cpunode == NULL) {
1591 		return (NULL);
1592 	}
1593 
1594 	/*
1595 	 * When the topology is known choose a cpu whos group has, in
1596 	 * aggregate, has the lowest weighted load.
1597 	 */
1598 	cpup = root_cpu_node;
1599 	rdd = dd;
1600 	while (cpup) {
1601 		/*
1602 		 * Degenerate case super-root
1603 		 */
1604 		if (cpup->child_node && cpup->child_no == 1) {
1605 			cpup = cpup->child_node;
1606 			continue;
1607 		}
1608 
1609 		/*
1610 		 * Terminal cpunode
1611 		 */
1612 		if (cpup->child_node == NULL) {
1613 			rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1614 			break;
1615 		}
1616 
1617 		cpub = NULL;
1618 		highest_load = 0;
1619 
1620 		for (n = 0; n < cpup->child_no; ++n) {
1621 			/*
1622 			 * Accumulate load information for all cpus
1623 			 * which are members of this node.
1624 			 */
1625 			cpun = &cpup->child_node[n];
1626 			mask = cpun->members & usched_global_cpumask &
1627 			       smp_active_mask;
1628 			if (mask == 0)
1629 				continue;
1630 			count = 0;
1631 			load = 0;
1632 
1633 			while (mask) {
1634 				cpuid = BSFCPUMASK(mask);
1635 				rdd = &dfly_pcpu[cpuid];
1636 				load += rdd->uload;
1637 				load += rdd->ucount * usched_dfly_weight3;
1638 				if (rdd->uschedcp == NULL &&
1639 				    rdd->runqcount == 0 &&
1640 				    globaldata_find(cpuid)->gd_tdrunqcount == 0
1641 				) {
1642 					load -= usched_dfly_weight4;
1643 				}
1644 #if 0
1645 				else if (rdd->upri > dd->upri + PPQ) {
1646 					load -= usched_dfly_weight4 / 2;
1647 				}
1648 #endif
1649 				mask &= ~CPUMASK(cpuid);
1650 				++count;
1651 			}
1652 			load /= count;
1653 
1654 			/*
1655 			 * Prefer candidates which are somewhat closer to
1656 			 * our cpu.
1657 			 */
1658 			if (dd->cpumask & cpun->members)
1659 				load += usched_dfly_weight1;
1660 
1661 			/*
1662 			 * The best candidate is the one with the worst
1663 			 * (highest) load.
1664 			 */
1665 			if (cpub == NULL || highest_load < load) {
1666 				highest_load = load;
1667 				cpub = cpun;
1668 			}
1669 		}
1670 		cpup = cpub;
1671 	}
1672 
1673 	/*
1674 	 * We never return our own node (dd), and only return a remote
1675 	 * node if it's load is significantly worse than ours (i.e. where
1676 	 * stealing a thread would be considered reasonable).
1677 	 *
1678 	 * This also helps us avoid breaking paired threads apart which
1679 	 * can have disastrous effects on performance.
1680 	 */
1681 	if (rdd == dd)
1682 		return(NULL);
1683 
1684 #if 0
1685 	hpri = 0;
1686 	if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1687 		hpri = pri;
1688 	if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1689 		hpri = pri;
1690 	if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1691 		hpri = pri;
1692 	hpri *= PPQ;
1693 	if (rdd->uload - hpri < dd->uload + hpri)
1694 		return(NULL);
1695 #endif
1696 	return (rdd);
1697 }
1698 
1699 static
1700 dfly_pcpu_t
1701 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1702 {
1703 	dfly_pcpu_t rdd;
1704 	cpumask_t tmpmask;
1705 	cpumask_t mask;
1706 	int cpuid;
1707 
1708 	/*
1709 	 * Fallback to the original heuristic, select random cpu,
1710 	 * first checking cpus not currently running a user thread.
1711 	 */
1712 	++dfly_scancpu;
1713 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1714 	mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1715 	       smp_active_mask & usched_global_cpumask;
1716 
1717 	while (mask) {
1718 		tmpmask = ~(CPUMASK(cpuid) - 1);
1719 		if (mask & tmpmask)
1720 			cpuid = BSFCPUMASK(mask & tmpmask);
1721 		else
1722 			cpuid = BSFCPUMASK(mask);
1723 		rdd = &dfly_pcpu[cpuid];
1724 
1725 		if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1726 			goto found;
1727 		mask &= ~CPUMASK(cpuid);
1728 	}
1729 
1730 	/*
1731 	 * Then cpus which might have a currently running lp
1732 	 */
1733 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1734 	mask = dfly_curprocmask & dfly_rdyprocmask &
1735 	       lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1736 
1737 	while (mask) {
1738 		tmpmask = ~(CPUMASK(cpuid) - 1);
1739 		if (mask & tmpmask)
1740 			cpuid = BSFCPUMASK(mask & tmpmask);
1741 		else
1742 			cpuid = BSFCPUMASK(mask);
1743 		rdd = &dfly_pcpu[cpuid];
1744 
1745 		if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1746 			goto found;
1747 		mask &= ~CPUMASK(cpuid);
1748 	}
1749 
1750 	/*
1751 	 * If we cannot find a suitable cpu we reload from dfly_scancpu
1752 	 * and round-robin.  Other cpus will pickup as they release their
1753 	 * current lwps or become ready.
1754 	 *
1755 	 * Avoid a degenerate system lockup case if usched_global_cpumask
1756 	 * is set to 0 or otherwise does not cover lwp_cpumask.
1757 	 *
1758 	 * We only kick the target helper thread in this case, we do not
1759 	 * set the user resched flag because
1760 	 */
1761 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1762 	if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1763 		cpuid = 0;
1764 	rdd = &dfly_pcpu[cpuid];
1765 found:
1766 	return (rdd);
1767 }
1768 
1769 static
1770 void
1771 dfly_need_user_resched_remote(void *dummy)
1772 {
1773 	globaldata_t gd = mycpu;
1774 	dfly_pcpu_t  dd = &dfly_pcpu[gd->gd_cpuid];
1775 
1776 	/*
1777 	 * Flag reschedule needed
1778 	 */
1779 	need_user_resched();
1780 
1781 	/*
1782 	 * If no user thread is currently running we need to kick the helper
1783 	 * on our cpu to recover.  Otherwise the cpu will never schedule
1784 	 * anything again.
1785 	 *
1786 	 * We cannot schedule the process ourselves because this is an
1787 	 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1788 	 *
1789 	 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1790 	 */
1791 	if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1792 		atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1793 		wakeup_mycpu(&dd->helper_thread);
1794 	}
1795 }
1796 
1797 /*
1798  * dfly_remrunqueue_locked() removes a given process from the run queue
1799  * that it is on, clearing the queue busy bit if it becomes empty.
1800  *
1801  * Note that user process scheduler is different from the LWKT schedule.
1802  * The user process scheduler only manages user processes but it uses LWKT
1803  * underneath, and a user process operating in the kernel will often be
1804  * 'released' from our management.
1805  *
1806  * uload is NOT adjusted here.  It is only adjusted if the lwkt_thread goes
1807  * to sleep or the lwp is moved to a different runq.
1808  */
1809 static void
1810 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1811 {
1812 	struct rq *q;
1813 	u_int32_t *which;
1814 	u_int8_t pri;
1815 
1816 	KKASSERT(rdd->runqcount >= 0);
1817 
1818 	pri = lp->lwp_rqindex;
1819 
1820 	switch(lp->lwp_rqtype) {
1821 	case RTP_PRIO_NORMAL:
1822 		q = &rdd->queues[pri];
1823 		which = &rdd->queuebits;
1824 		break;
1825 	case RTP_PRIO_REALTIME:
1826 	case RTP_PRIO_FIFO:
1827 		q = &rdd->rtqueues[pri];
1828 		which = &rdd->rtqueuebits;
1829 		break;
1830 	case RTP_PRIO_IDLE:
1831 		q = &rdd->idqueues[pri];
1832 		which = &rdd->idqueuebits;
1833 		break;
1834 	default:
1835 		panic("remrunqueue: invalid rtprio type");
1836 		/* NOT REACHED */
1837 	}
1838 	KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1839 	atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1840 	TAILQ_REMOVE(q, lp, lwp_procq);
1841 	--rdd->runqcount;
1842 	if (TAILQ_EMPTY(q)) {
1843 		KASSERT((*which & (1 << pri)) != 0,
1844 			("remrunqueue: remove from empty queue"));
1845 		*which &= ~(1 << pri);
1846 	}
1847 }
1848 
1849 /*
1850  * dfly_setrunqueue_locked()
1851  *
1852  * Add a process whos rqtype and rqindex had previously been calculated
1853  * onto the appropriate run queue.   Determine if the addition requires
1854  * a reschedule on a cpu and return the cpuid or -1.
1855  *
1856  * NOTE: 	  Lower priorities are better priorities.
1857  *
1858  * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1859  *		  sum of the rough lwp_priority for all running and runnable
1860  *		  processes.  Lower priority processes (higher lwp_priority
1861  *		  values) actually DO count as more load, not less, because
1862  *		  these are the programs which require the most care with
1863  *		  regards to cpu selection.
1864  */
1865 static void
1866 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1867 {
1868 	struct rq *q;
1869 	u_int32_t *which;
1870 	int pri;
1871 
1872 	KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1873 
1874 	if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1875 		atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1876 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1877 		atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1878 		atomic_add_int(&dfly_ucount, 1);
1879 	}
1880 
1881 	pri = lp->lwp_rqindex;
1882 
1883 	switch(lp->lwp_rqtype) {
1884 	case RTP_PRIO_NORMAL:
1885 		q = &rdd->queues[pri];
1886 		which = &rdd->queuebits;
1887 		break;
1888 	case RTP_PRIO_REALTIME:
1889 	case RTP_PRIO_FIFO:
1890 		q = &rdd->rtqueues[pri];
1891 		which = &rdd->rtqueuebits;
1892 		break;
1893 	case RTP_PRIO_IDLE:
1894 		q = &rdd->idqueues[pri];
1895 		which = &rdd->idqueuebits;
1896 		break;
1897 	default:
1898 		panic("remrunqueue: invalid rtprio type");
1899 		/* NOT REACHED */
1900 	}
1901 
1902 	/*
1903 	 * Place us on the selected queue.  Determine if we should be
1904 	 * placed at the head of the queue or at the end.
1905 	 *
1906 	 * We are placed at the tail if our round-robin count has expired,
1907 	 * or is about to expire and the system thinks its a good place to
1908 	 * round-robin, or there is already a next thread on the queue
1909 	 * (it might be trying to pick up where it left off and we don't
1910 	 * want to interfere).
1911 	 */
1912 	KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1913 	atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1914 	++rdd->runqcount;
1915 
1916 	if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1917 	    (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1918 	     (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1919 	    !TAILQ_EMPTY(q)
1920 	) {
1921 		atomic_clear_int(&lp->lwp_thread->td_mpflags,
1922 				 TDF_MP_BATCH_DEMARC);
1923 		lp->lwp_rrcount = 0;
1924 		TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1925 	} else {
1926 		if (TAILQ_EMPTY(q))
1927 			lp->lwp_rrcount = 0;
1928 		TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1929 	}
1930 	*which |= 1 << pri;
1931 }
1932 
1933 /*
1934  * For SMP systems a user scheduler helper thread is created for each
1935  * cpu and is used to allow one cpu to wakeup another for the purposes of
1936  * scheduling userland threads from setrunqueue().
1937  *
1938  * UP systems do not need the helper since there is only one cpu.
1939  *
1940  * We can't use the idle thread for this because we might block.
1941  * Additionally, doing things this way allows us to HLT idle cpus
1942  * on MP systems.
1943  */
1944 static void
1945 dfly_helper_thread(void *dummy)
1946 {
1947     globaldata_t gd;
1948     dfly_pcpu_t dd;
1949     dfly_pcpu_t rdd;
1950     struct lwp *nlp;
1951     cpumask_t mask;
1952     int cpuid;
1953 
1954     gd = mycpu;
1955     cpuid = gd->gd_cpuid;	/* doesn't change */
1956     mask = gd->gd_cpumask;	/* doesn't change */
1957     dd = &dfly_pcpu[cpuid];
1958 
1959     /*
1960      * Since we only want to be woken up only when no user processes
1961      * are scheduled on a cpu, run at an ultra low priority.
1962      */
1963     lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1964 
1965     tsleep(&dd->helper_thread, 0, "schslp", 0);
1966 
1967     for (;;) {
1968 	/*
1969 	 * We use the LWKT deschedule-interlock trick to avoid racing
1970 	 * dfly_rdyprocmask.  This means we cannot block through to the
1971 	 * manual lwkt_switch() call we make below.
1972 	 */
1973 	crit_enter_gd(gd);
1974 	tsleep_interlock(&dd->helper_thread, 0);
1975 
1976 	spin_lock(&dd->spin);
1977 
1978 	atomic_set_cpumask(&dfly_rdyprocmask, mask);
1979 	clear_user_resched();	/* This satisfied the reschedule request */
1980 #if 0
1981 	dd->rrcount = 0;	/* Reset the round-robin counter */
1982 #endif
1983 
1984 	if (dd->runqcount || dd->uschedcp != NULL) {
1985 		/*
1986 		 * Threads are available.  A thread may or may not be
1987 		 * currently scheduled.  Get the best thread already queued
1988 		 * to this cpu.
1989 		 */
1990 		nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
1991 		if (nlp) {
1992 			atomic_set_cpumask(&dfly_curprocmask, mask);
1993 			dd->upri = nlp->lwp_priority;
1994 			dd->uschedcp = nlp;
1995 #if 0
1996 			dd->rrcount = 0;	/* reset round robin */
1997 #endif
1998 			spin_unlock(&dd->spin);
1999 			lwkt_acquire(nlp->lwp_thread);
2000 			lwkt_schedule(nlp->lwp_thread);
2001 		} else {
2002 			/*
2003 			 * This situation should not occur because we had
2004 			 * at least one thread available.
2005 			 */
2006 			spin_unlock(&dd->spin);
2007 		}
2008 	} else if (usched_dfly_features & 0x01) {
2009 		/*
2010 		 * This cpu is devoid of runnable threads, steal a thread
2011 		 * from another cpu.  Since we're stealing, might as well
2012 		 * load balance at the same time.
2013 		 *
2014 		 * We choose the highest-loaded thread from the worst queue.
2015 		 *
2016 		 * NOTE! This function only returns a non-NULL rdd when
2017 		 *	 another cpu's queue is obviously overloaded.  We
2018 		 *	 do not want to perform the type of rebalancing
2019 		 *	 the schedclock does here because it would result
2020 		 *	 in insane process pulling when 'steady' state is
2021 		 *	 partially unbalanced (e.g. 6 runnables and only
2022 		 *	 4 cores).
2023 		 */
2024 		rdd = dfly_choose_worst_queue(dd);
2025 		if (rdd && spin_trylock(&rdd->spin)) {
2026 			nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2027 			spin_unlock(&rdd->spin);
2028 		} else {
2029 			nlp = NULL;
2030 		}
2031 		if (nlp) {
2032 			atomic_set_cpumask(&dfly_curprocmask, mask);
2033 			dd->upri = nlp->lwp_priority;
2034 			dd->uschedcp = nlp;
2035 #if 0
2036 			dd->rrcount = 0;	/* reset round robin */
2037 #endif
2038 			spin_unlock(&dd->spin);
2039 			lwkt_acquire(nlp->lwp_thread);
2040 			lwkt_schedule(nlp->lwp_thread);
2041 		} else {
2042 			/*
2043 			 * Leave the thread on our run queue.  Another
2044 			 * scheduler will try to pull it later.
2045 			 */
2046 			spin_unlock(&dd->spin);
2047 		}
2048 	} else {
2049 		/*
2050 		 * devoid of runnable threads and not allowed to steal
2051 		 * any.
2052 		 */
2053 		spin_unlock(&dd->spin);
2054 	}
2055 
2056 	/*
2057 	 * We're descheduled unless someone scheduled us.  Switch away.
2058 	 * Exiting the critical section will cause splz() to be called
2059 	 * for us if interrupts and such are pending.
2060 	 */
2061 	crit_exit_gd(gd);
2062 	tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2063     }
2064 }
2065 
2066 #if 0
2067 static int
2068 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2069 {
2070 	int error, new_val;
2071 
2072 	new_val = usched_dfly_stick_to_level;
2073 
2074 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2075         if (error != 0 || req->newptr == NULL)
2076 		return (error);
2077 	if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2078 		return (EINVAL);
2079 	usched_dfly_stick_to_level = new_val;
2080 	return (0);
2081 }
2082 #endif
2083 
2084 /*
2085  * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2086  * Note that curprocmask bit 0 has already been cleared by rqinit() and
2087  * we should not mess with it further.
2088  */
2089 static void
2090 usched_dfly_cpu_init(void)
2091 {
2092 	int i;
2093 	int j;
2094 	int cpuid;
2095 	int smt_not_supported = 0;
2096 	int cache_coherent_not_supported = 0;
2097 
2098 	if (bootverbose)
2099 		kprintf("Start scheduler helpers on cpus:\n");
2100 
2101 	sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2102 	usched_dfly_sysctl_tree =
2103 		SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2104 				SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2105 				"usched_dfly", CTLFLAG_RD, 0, "");
2106 
2107 	for (i = 0; i < ncpus; ++i) {
2108 		dfly_pcpu_t dd = &dfly_pcpu[i];
2109 		cpumask_t mask = CPUMASK(i);
2110 
2111 		if ((mask & smp_active_mask) == 0)
2112 		    continue;
2113 
2114 		spin_init(&dd->spin);
2115 		dd->cpunode = get_cpu_node_by_cpuid(i);
2116 		dd->cpuid = i;
2117 		dd->cpumask = CPUMASK(i);
2118 		for (j = 0; j < NQS; j++) {
2119 			TAILQ_INIT(&dd->queues[j]);
2120 			TAILQ_INIT(&dd->rtqueues[j]);
2121 			TAILQ_INIT(&dd->idqueues[j]);
2122 		}
2123 		atomic_clear_cpumask(&dfly_curprocmask, 1);
2124 
2125 		if (dd->cpunode == NULL) {
2126 			smt_not_supported = 1;
2127 			cache_coherent_not_supported = 1;
2128 			if (bootverbose)
2129 				kprintf ("\tcpu%d - WARNING: No CPU NODE "
2130 					 "found for cpu\n", i);
2131 		} else {
2132 			switch (dd->cpunode->type) {
2133 			case THREAD_LEVEL:
2134 				if (bootverbose)
2135 					kprintf ("\tcpu%d - HyperThreading "
2136 						 "available. Core siblings: ",
2137 						 i);
2138 				break;
2139 			case CORE_LEVEL:
2140 				smt_not_supported = 1;
2141 
2142 				if (bootverbose)
2143 					kprintf ("\tcpu%d - No HT available, "
2144 						 "multi-core/physical "
2145 						 "cpu. Physical siblings: ",
2146 						 i);
2147 				break;
2148 			case CHIP_LEVEL:
2149 				smt_not_supported = 1;
2150 
2151 				if (bootverbose)
2152 					kprintf ("\tcpu%d - No HT available, "
2153 						 "single-core/physical cpu. "
2154 						 "Package Siblings: ",
2155 						 i);
2156 				break;
2157 			default:
2158 				/* Let's go for safe defaults here */
2159 				smt_not_supported = 1;
2160 				cache_coherent_not_supported = 1;
2161 				if (bootverbose)
2162 					kprintf ("\tcpu%d - Unknown cpunode->"
2163 						 "type=%u. Siblings: ",
2164 						 i,
2165 						 (u_int)dd->cpunode->type);
2166 				break;
2167 			}
2168 
2169 			if (bootverbose) {
2170 				if (dd->cpunode->parent_node != NULL) {
2171 					CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2172 						kprintf("cpu%d ", cpuid);
2173 					kprintf("\n");
2174 				} else {
2175 					kprintf(" no siblings\n");
2176 				}
2177 			}
2178 		}
2179 
2180 		lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2181 			    0, i, "usched %d", i);
2182 
2183 		/*
2184 		 * Allow user scheduling on the target cpu.  cpu #0 has already
2185 		 * been enabled in rqinit().
2186 		 */
2187 		if (i)
2188 		    atomic_clear_cpumask(&dfly_curprocmask, mask);
2189 		atomic_set_cpumask(&dfly_rdyprocmask, mask);
2190 		dd->upri = PRIBASE_NULL;
2191 
2192 	}
2193 
2194 	/* usched_dfly sysctl configurable parameters */
2195 
2196 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2197 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2198 		       OID_AUTO, "rrinterval", CTLFLAG_RW,
2199 		       &usched_dfly_rrinterval, 0, "");
2200 	SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2201 		       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2202 		       OID_AUTO, "decay", CTLFLAG_RW,
2203 		       &usched_dfly_decay, 0, "Extra decay when not running");
2204 
2205 	/* Add enable/disable option for SMT scheduling if supported */
2206 	if (smt_not_supported) {
2207 		usched_dfly_smt = 0;
2208 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2209 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2210 				  OID_AUTO, "smt", CTLFLAG_RD,
2211 				  "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2212 	} else {
2213 		usched_dfly_smt = 1;
2214 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2215 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2216 			       OID_AUTO, "smt", CTLFLAG_RW,
2217 			       &usched_dfly_smt, 0, "Enable SMT scheduling");
2218 	}
2219 
2220 	/*
2221 	 * Add enable/disable option for cache coherent scheduling
2222 	 * if supported
2223 	 */
2224 	if (cache_coherent_not_supported) {
2225 		usched_dfly_cache_coherent = 0;
2226 		SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2227 				  SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2228 				  OID_AUTO, "cache_coherent", CTLFLAG_RD,
2229 				  "NOT SUPPORTED", 0,
2230 				  "Cache coherence NOT SUPPORTED");
2231 	} else {
2232 		usched_dfly_cache_coherent = 1;
2233 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2234 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2235 			       OID_AUTO, "cache_coherent", CTLFLAG_RW,
2236 			       &usched_dfly_cache_coherent, 0,
2237 			       "Enable/Disable cache coherent scheduling");
2238 
2239 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2240 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2241 			       OID_AUTO, "weight1", CTLFLAG_RW,
2242 			       &usched_dfly_weight1, 200,
2243 			       "Weight selection for current cpu");
2244 
2245 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2246 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2247 			       OID_AUTO, "weight2", CTLFLAG_RW,
2248 			       &usched_dfly_weight2, 180,
2249 			       "Weight selection for wakefrom cpu");
2250 
2251 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2252 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2253 			       OID_AUTO, "weight3", CTLFLAG_RW,
2254 			       &usched_dfly_weight3, 40,
2255 			       "Weight selection for num threads on queue");
2256 
2257 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2258 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2259 			       OID_AUTO, "weight4", CTLFLAG_RW,
2260 			       &usched_dfly_weight4, 160,
2261 			       "Availability of other idle cpus");
2262 
2263 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2264 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2265 			       OID_AUTO, "fast_resched", CTLFLAG_RW,
2266 			       &usched_dfly_fast_resched, 0,
2267 			       "Availability of other idle cpus");
2268 
2269 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2270 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2271 			       OID_AUTO, "features", CTLFLAG_RW,
2272 			       &usched_dfly_features, 0x8F,
2273 			       "Allow pulls into empty queues");
2274 
2275 		SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2276 			       SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2277 			       OID_AUTO, "swmask", CTLFLAG_RW,
2278 			       &usched_dfly_swmask, ~PPQMASK,
2279 			       "Queue mask to force thread switch");
2280 
2281 #if 0
2282 		SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2283 				SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2284 				OID_AUTO, "stick_to_level",
2285 				CTLTYPE_INT | CTLFLAG_RW,
2286 				NULL, sizeof usched_dfly_stick_to_level,
2287 				sysctl_usched_dfly_stick_to_level, "I",
2288 				"Stick a process to this level. See sysctl"
2289 				"paremter hw.cpu_topology.level_description");
2290 #endif
2291 	}
2292 }
2293 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2294 	usched_dfly_cpu_init, NULL)
2295