2 * Copyright (c) 2012 The DragonFly Project. All rights reserved.
3 * Copyright (c) 1999 Peter Wemm <peter@FreeBSD.org>. All rights reserved.
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>
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
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
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.
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
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
41 #include <sys/queue.h>
43 #include <sys/rtprio.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>
55 #include <machine/cpu.h>
56 #include <machine/smp.h>
59 * Priorities. Note that with 32 run queues per scheduler each queue
60 * represents four priority levels.
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)
73 #define NQS 32 /* 32 run queues. */
74 #define PPQ (MAXPRI / NQS) /* priorities per queue */
75 #define PPQMASK (PPQ - 1)
78 * NICEPPQ - number of nice units per priority queue
79 * ESTCPUPPQ - number of estcpu units per priority queue
80 * ESTCPUMAX - number of estcpu units
84 #define ESTCPUMAX (ESTCPUPPQ * NQS)
85 #define BATCHMAX (ESTCPUFREQ * 30)
86 #define PRIO_RANGE (PRIO_MAX - PRIO_MIN + 1)
88 #define ESTCPULIM(v) min((v), ESTCPUMAX)
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
102 struct usched_dfly_pcpu {
103 struct spinlock spin;
104 struct thread helper_thread;
109 struct lwp *uschedcp;
110 struct rq queues[NQS];
111 struct rq rtqueues[NQS];
112 struct rq idqueues[NQS];
114 u_int32_t rtqueuebits;
115 u_int32_t idqueuebits;
122 typedef struct usched_dfly_pcpu *dfly_pcpu_t;
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,
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);
149 struct usched usched_dfly = {
151 "dfly", "Original DragonFly Scheduler",
152 NULL, /* default registration */
153 NULL, /* default deregistration */
154 dfly_acquire_curproc,
155 dfly_release_curproc,
158 dfly_recalculate_estcpu,
163 NULL, /* setcpumask not supported */
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
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;
188 /* Debug info exposed through debug.* sysctl */
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");
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");
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");
206 * Tunning usched_dfly - configurable through kern.usched_dfly.
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.
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).
216 * WARNING! Weight2 is a ridiculously sensitive parameter,
217 * a small value is recommended.
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.
225 * This value should be left fairly small relative to weight1
228 * weight4 - Weighting based on other cpu queues being available
229 * or running processes with higher lwp_priority's.
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
236 * features - These flags can be set or cleared to enable or disable various
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)
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;
260 /* KTR debug printings */
262 KTR_INFO_MASTER(usched);
264 #if !defined(KTR_USCHED_DFLY)
265 #define KTR_USCHED_DFLY KTR_ALL
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);
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.
277 * The kernel will not depress our LWKT priority until after we return,
278 * in case we have to shove over to another cpu.
280 * We must determine our thread's disposition before we switch away. This
281 * is very sensitive code.
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.
288 dfly_acquire_curproc(struct lwp *lp)
297 * Make sure we aren't sitting on a tsleep queue.
300 crit_enter_quick(td);
301 if (td->td_flags & TDF_TSLEEPQ)
303 dfly_recalculate_estcpu(lp);
306 dd = &dfly_pcpu[gd->gd_cpuid];
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.
314 if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
315 lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
319 if (user_resched_wanted()) {
320 if (dd->uschedcp == lp)
322 clear_user_resched();
323 dfly_release_curproc(lp);
327 * Loop until we are the current user thread.
329 * NOTE: dd spinlock not held at top of loop.
331 if (dd->uschedcp == lp)
334 while (dd->uschedcp != lp) {
337 spin_lock(&dd->spin);
340 (usched_dfly_features & 0x08) &&
341 (rdd = dfly_choose_best_queue(lp)) != dd) {
343 * We are not or are no longer the current lwp and a
344 * forced reschedule was requested. Figure out the
345 * best cpu to run on (our current cpu will be given
346 * significant weight).
348 * (if a reschedule was not requested we want to
349 * move this step after the uschedcp tests).
351 dfly_changeqcpu_locked(lp, dd, rdd);
352 spin_unlock(&dd->spin);
353 lwkt_deschedule(lp->lwp_thread);
354 dfly_setrunqueue_dd(rdd, lp);
357 dd = &dfly_pcpu[gd->gd_cpuid];
362 * Either no reschedule was requested or the best queue was
363 * dd, and no current process has been selected. We can
364 * trivially become the current lwp on the current cpu.
366 if (dd->uschedcp == NULL) {
367 atomic_clear_int(&lp->lwp_thread->td_mpflags,
369 atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
371 dd->upri = lp->lwp_priority;
372 KKASSERT(lp->lwp_qcpu == dd->cpuid);
373 spin_unlock(&dd->spin);
378 * Put us back on the same run queue unconditionally.
380 * Set rrinterval to force placement at end of queue.
381 * Select the worst queue to ensure we round-robin,
382 * but do not change estcpu.
384 if (lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) {
387 switch(lp->lwp_rqtype) {
388 case RTP_PRIO_NORMAL:
389 tsqbits = dd->queuebits;
390 spin_unlock(&dd->spin);
392 lp->lwp_rrcount = usched_dfly_rrinterval;
394 lp->lwp_rqindex = bsrl(tsqbits);
397 spin_unlock(&dd->spin);
400 lwkt_deschedule(lp->lwp_thread);
401 dfly_setrunqueue_dd(dd, lp);
402 atomic_clear_int(&lp->lwp_thread->td_mpflags,
406 dd = &dfly_pcpu[gd->gd_cpuid];
411 * Can we steal the current designated user thread?
413 * If we do the other thread will stall when it tries to
414 * return to userland, possibly rescheduling elsewhere.
416 * It is important to do a masked test to avoid the edge
417 * case where two near-equal-priority threads are constantly
418 * interrupting each other.
420 * In the exact match case another thread has already gained
421 * uschedcp and lowered its priority, if we steal it the
422 * other thread will stay stuck on the LWKT runq and not
423 * push to another cpu. So don't steal on equal-priority even
424 * though it might appear to be more beneficial due to not
425 * having to switch back to the other thread's context.
427 * usched_dfly_fast_resched requires that two threads be
428 * significantly far apart in priority in order to interrupt.
430 * If better but not sufficiently far apart, the current
431 * uschedcp will be interrupted at the next scheduler clock.
434 (dd->upri & ~PPQMASK) >
435 (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
437 dd->upri = lp->lwp_priority;
438 KKASSERT(lp->lwp_qcpu == dd->cpuid);
439 spin_unlock(&dd->spin);
443 * We are not the current lwp, figure out the best cpu
444 * to run on (our current cpu will be given significant
445 * weight). Loop on cpu change.
447 if ((usched_dfly_features & 0x02) &&
448 force_resched == 0 &&
449 (rdd = dfly_choose_best_queue(lp)) != dd) {
450 dfly_changeqcpu_locked(lp, dd, rdd);
451 spin_unlock(&dd->spin);
452 lwkt_deschedule(lp->lwp_thread);
453 dfly_setrunqueue_dd(rdd, lp);
456 dd = &dfly_pcpu[gd->gd_cpuid];
461 * We cannot become the current lwp, place the lp on the
462 * run-queue of this or another cpu and deschedule ourselves.
464 * When we are reactivated we will have another chance.
466 * Reload after a switch or setrunqueue/switch possibly
467 * moved us to another cpu.
469 spin_unlock(&dd->spin);
470 lwkt_deschedule(lp->lwp_thread);
471 dfly_setrunqueue_dd(dd, lp);
474 dd = &dfly_pcpu[gd->gd_cpuid];
478 * Make sure upri is synchronized, then yield to LWKT threads as
479 * needed before returning. This could result in another reschedule.
484 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
488 * DFLY_RELEASE_CURPROC
490 * This routine detaches the current thread from the userland scheduler,
491 * usually because the thread needs to run or block in the kernel (at
492 * kernel priority) for a while.
494 * This routine is also responsible for selecting a new thread to
495 * make the current thread.
497 * NOTE: This implementation differs from the dummy example in that
498 * dfly_select_curproc() is able to select the current process, whereas
499 * dummy_select_curproc() is not able to select the current process.
500 * This means we have to NULL out uschedcp.
502 * Additionally, note that we may already be on a run queue if releasing
503 * via the lwkt_switch() in dfly_setrunqueue().
506 dfly_release_curproc(struct lwp *lp)
508 globaldata_t gd = mycpu;
509 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
512 * Make sure td_wakefromcpu is defaulted. This will be overwritten
515 if (dd->uschedcp == lp) {
516 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
517 spin_lock(&dd->spin);
518 if (dd->uschedcp == lp) {
519 dd->uschedcp = NULL; /* don't let lp be selected */
520 dd->upri = PRIBASE_NULL;
521 atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
522 spin_unlock(&dd->spin);
523 dfly_select_curproc(gd);
525 spin_unlock(&dd->spin);
531 * DFLY_SELECT_CURPROC
533 * Select a new current process for this cpu and clear any pending user
534 * reschedule request. The cpu currently has no current process.
536 * This routine is also responsible for equal-priority round-robining,
537 * typically triggered from dfly_schedulerclock(). In our dummy example
538 * all the 'user' threads are LWKT scheduled all at once and we just
539 * call lwkt_switch().
541 * The calling process is not on the queue and cannot be selected.
545 dfly_select_curproc(globaldata_t gd)
547 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
549 int cpuid = gd->gd_cpuid;
553 spin_lock(&dd->spin);
554 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
557 atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
558 dd->upri = nlp->lwp_priority;
561 dd->rrcount = 0; /* reset round robin */
563 spin_unlock(&dd->spin);
564 lwkt_acquire(nlp->lwp_thread);
565 lwkt_schedule(nlp->lwp_thread);
567 spin_unlock(&dd->spin);
573 * Place the specified lwp on the user scheduler's run queue. This routine
574 * must be called with the thread descheduled. The lwp must be runnable.
575 * It must not be possible for anyone else to explicitly schedule this thread.
577 * The thread may be the current thread as a special case.
580 dfly_setrunqueue(struct lwp *lp)
586 * First validate the process LWKT state.
588 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
589 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
590 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
591 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
592 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
595 * NOTE: dd/rdd do not necessarily represent the current cpu.
596 * Instead they may represent the cpu the thread was last
597 * scheduled on or inherited by its parent.
599 dd = &dfly_pcpu[lp->lwp_qcpu];
603 * This process is not supposed to be scheduled anywhere or assigned
604 * as the current process anywhere. Assert the condition.
606 KKASSERT(rdd->uschedcp != lp);
609 * Ok, we have to setrunqueue some target cpu and request a reschedule
612 * We have to choose the best target cpu. It might not be the current
613 * target even if the current cpu has no running user thread (for
614 * example, because the current cpu might be a hyperthread and its
615 * sibling has a thread assigned).
617 * If we just forked it is most optimal to run the child on the same
618 * cpu just in case the parent decides to wait for it (thus getting
619 * off that cpu). As long as there is nothing else runnable on the
620 * cpu, that is. If we did this unconditionally a parent forking
621 * multiple children before waiting (e.g. make -j N) leaves other
622 * cpus idle that could be working.
624 if (lp->lwp_forked) {
626 if (usched_dfly_features & 0x20)
627 rdd = dfly_choose_best_queue(lp);
628 else if (usched_dfly_features & 0x40)
629 rdd = &dfly_pcpu[lp->lwp_qcpu];
630 else if (usched_dfly_features & 0x80)
631 rdd = dfly_choose_queue_simple(rdd, lp);
632 else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
633 rdd = dfly_choose_best_queue(lp);
635 rdd = &dfly_pcpu[lp->lwp_qcpu];
637 rdd = dfly_choose_best_queue(lp);
638 /* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
640 if (lp->lwp_qcpu != rdd->cpuid) {
641 spin_lock(&dd->spin);
642 dfly_changeqcpu_locked(lp, dd, rdd);
643 spin_unlock(&dd->spin);
645 dfly_setrunqueue_dd(rdd, lp);
649 * Change qcpu to rdd->cpuid. The dd the lp is CURRENTLY on must be
650 * spin-locked on-call. rdd does not have to be.
653 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
655 if (lp->lwp_qcpu != rdd->cpuid) {
656 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
657 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
658 atomic_add_int(&dd->uload, -lp->lwp_uload);
659 atomic_add_int(&dd->ucount, -1);
660 atomic_add_int(&dfly_ucount, -1);
662 lp->lwp_qcpu = rdd->cpuid;
667 * Place lp on rdd's runqueue. Nothing is locked on call. This function
668 * also performs all necessary ancillary notification actions.
671 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
676 * We might be moving the lp to another cpu's run queue, and once
677 * on the runqueue (even if it is our cpu's), another cpu can rip
680 * TDF_MIGRATING might already be set if this is part of a
681 * remrunqueue+setrunqueue sequence.
683 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
684 lwkt_giveaway(lp->lwp_thread);
686 rgd = globaldata_find(rdd->cpuid);
689 * We lose control of the lp the moment we release the spinlock
690 * after having placed it on the queue. i.e. another cpu could pick
691 * it up, or it could exit, or its priority could be further
692 * adjusted, or something like that.
694 * WARNING! rdd can point to a foreign cpu!
696 spin_lock(&rdd->spin);
697 dfly_setrunqueue_locked(rdd, lp);
700 * Potentially interrupt the currently-running thread
702 if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
704 * Currently running thread is better or same, do not
707 spin_unlock(&rdd->spin);
708 } else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
709 usched_dfly_fast_resched) {
711 * Currently running thread is not better, but not so bad
712 * that we need to interrupt it. Let it run for one more
716 rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
717 rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
719 spin_unlock(&rdd->spin);
720 } else if (rgd == mycpu) {
722 * We should interrupt the currently running thread, which
723 * is on the current cpu. However, if DIDYIELD is set we
724 * round-robin unconditionally and do not interrupt it.
726 spin_unlock(&rdd->spin);
727 if (rdd->uschedcp == NULL)
728 wakeup_mycpu(&rdd->helper_thread); /* XXX */
729 if ((lp->lwp_thread->td_mpflags & TDF_MP_DIDYIELD) == 0)
733 * We should interrupt the currently running thread, which
734 * is on a different cpu.
736 spin_unlock(&rdd->spin);
737 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
742 * This routine is called from a systimer IPI. It MUST be MP-safe and
743 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
748 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
750 globaldata_t gd = mycpu;
751 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
754 * Spinlocks also hold a critical section so there should not be
757 KKASSERT(gd->gd_spinlocks == 0);
763 * Do we need to round-robin? We round-robin 10 times a second.
764 * This should only occur for cpu-bound batch processes.
766 if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
767 lp->lwp_thread->td_wakefromcpu = -1;
772 * Adjust estcpu upward using a real time equivalent calculation,
773 * and recalculate lp's priority.
775 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
776 dfly_resetpriority(lp);
779 * Rebalance two cpus every 8 ticks, pulling the worst thread
780 * from the worst cpu's queue into a rotating cpu number.
782 * This mechanic is needed because the push algorithms can
783 * steady-state in an non-optimal configuration. We need to mix it
784 * up a little, even if it means breaking up a paired thread, so
785 * the push algorithms can rebalance the degenerate conditions.
786 * This portion of the algorithm exists to ensure stability at the
787 * selected weightings.
789 * Because we might be breaking up optimal conditions we do not want
790 * to execute this too quickly, hence we only rebalance approximately
791 * ~7-8 times per second. The push's, on the otherhand, are capable
792 * moving threads to other cpus at a much higher rate.
794 * We choose the most heavily loaded thread from the worst queue
795 * in order to ensure that multiple heavy-weight threads on the same
796 * queue get broken up, and also because these threads are the most
797 * likely to be able to remain in place. Hopefully then any pairings,
798 * if applicable, migrate to where these threads are.
800 if ((usched_dfly_features & 0x04) &&
801 ((u_int)sched_ticks & 7) == 0 &&
802 (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
809 rdd = dfly_choose_worst_queue(dd);
811 spin_lock(&dd->spin);
812 if (spin_trylock(&rdd->spin)) {
813 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
814 spin_unlock(&rdd->spin);
816 spin_unlock(&dd->spin);
818 spin_unlock(&dd->spin);
824 /* dd->spin held if nlp != NULL */
827 * Either schedule it or add it to our queue.
830 (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
831 atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
832 dd->upri = nlp->lwp_priority;
835 dd->rrcount = 0; /* reset round robin */
837 spin_unlock(&dd->spin);
838 lwkt_acquire(nlp->lwp_thread);
839 lwkt_schedule(nlp->lwp_thread);
841 dfly_setrunqueue_locked(dd, nlp);
842 spin_unlock(&dd->spin);
848 * Called from acquire and from kern_synch's one-second timer (one of the
849 * callout helper threads) with a critical section held.
851 * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
852 * overall system load.
854 * Note that no recalculation occurs for a process which sleeps and wakes
855 * up in the same tick. That is, a system doing thousands of context
856 * switches per second will still only do serious estcpu calculations
857 * ESTCPUFREQ times per second.
861 dfly_recalculate_estcpu(struct lwp *lp)
863 globaldata_t gd = mycpu;
871 * We have to subtract periodic to get the last schedclock
872 * timeout time, otherwise we would get the upcoming timeout.
873 * Keep in mind that a process can migrate between cpus and
874 * while the scheduler clock should be very close, boundary
875 * conditions could lead to a small negative delta.
877 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
879 if (lp->lwp_slptime > 1) {
881 * Too much time has passed, do a coarse correction.
883 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
884 dfly_resetpriority(lp);
885 lp->lwp_cpbase = cpbase;
888 } else if (lp->lwp_cpbase != cpbase) {
890 * Adjust estcpu if we are in a different tick. Don't waste
891 * time if we are in the same tick.
893 * First calculate the number of ticks in the measurement
894 * interval. The ttlticks calculation can wind up 0 due to
895 * a bug in the handling of lwp_slptime (as yet not found),
896 * so make sure we do not get a divide by 0 panic.
898 ttlticks = (cpbase - lp->lwp_cpbase) /
899 gd->gd_schedclock.periodic;
900 if ((ssysclock_t)ttlticks < 0) {
902 lp->lwp_cpbase = cpbase;
906 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
909 * Calculate the percentage of one cpu being used then
910 * compensate for any system load in excess of ncpus.
912 * For example, if we have 8 cores and 16 running cpu-bound
913 * processes then all things being equal each process will
914 * get 50% of one cpu. We need to pump this value back
915 * up to 100% so the estcpu calculation properly adjusts
916 * the process's dynamic priority.
918 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
920 estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
921 ucount = dfly_ucount;
922 if (ucount > ncpus) {
923 estcpu += estcpu * (ucount - ncpus) / ncpus;
926 if (usched_dfly_debug == lp->lwp_proc->p_pid) {
927 kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
928 lp->lwp_proc->p_pid, lp,
929 estcpu, lp->lwp_estcpu,
930 lp->lwp_cpticks, ttlticks);
934 * Adjust lp->lwp_esetcpu. The decay factor determines how
935 * quickly lwp_estcpu collapses to its realtime calculation.
936 * A slower collapse gives us a more accurate number over
937 * the long term but can create problems with bursty threads
938 * or threads which become cpu hogs.
940 * To solve this problem, newly started lwps and lwps which
941 * are restarting after having been asleep for a while are
942 * given a much, much faster decay in order to quickly
943 * detect whether they become cpu-bound.
945 * NOTE: p_nice is accounted for in dfly_resetpriority(),
946 * and not here, but we must still ensure that a
947 * cpu-bound nice -20 process does not completely
948 * override a cpu-bound nice +20 process.
950 * NOTE: We must use ESTCPULIM() here to deal with any
953 decay_factor = usched_dfly_decay;
954 if (decay_factor < 1)
956 if (decay_factor > 1024)
959 if (lp->lwp_estfast < usched_dfly_decay) {
961 lp->lwp_estcpu = ESTCPULIM(
962 (lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
963 (lp->lwp_estfast + 1));
965 lp->lwp_estcpu = ESTCPULIM(
966 (lp->lwp_estcpu * decay_factor + estcpu) /
970 if (usched_dfly_debug == lp->lwp_proc->p_pid)
971 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
972 dfly_resetpriority(lp);
973 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
979 * Compute the priority of a process when running in user mode.
980 * Arrange to reschedule if the resulting priority is better
981 * than that of the current process.
983 * This routine may be called with any process.
985 * This routine is called by fork1() for initial setup with the process
986 * of the run queue, and also may be called normally with the process on or
990 dfly_resetpriority(struct lwp *lp)
1003 * Lock the scheduler (lp) belongs to. This can be on a different
1004 * cpu. Handle races. This loop breaks out with the appropriate
1008 rcpu = lp->lwp_qcpu;
1010 rdd = &dfly_pcpu[rcpu];
1011 spin_lock(&rdd->spin);
1012 if (rcpu == lp->lwp_qcpu)
1014 spin_unlock(&rdd->spin);
1018 * Calculate the new priority and queue type
1020 newrqtype = lp->lwp_rtprio.type;
1023 case RTP_PRIO_REALTIME:
1025 newpriority = PRIBASE_REALTIME +
1026 (lp->lwp_rtprio.prio & PRIMASK);
1028 case RTP_PRIO_NORMAL:
1032 estcpu = lp->lwp_estcpu;
1035 * p_nice piece Adds (0-40) * 2 0-80
1036 * estcpu Adds 16384 * 4 / 512 0-128
1038 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1039 newpriority += estcpu * PPQ / ESTCPUPPQ;
1040 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1041 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1042 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1045 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1047 case RTP_PRIO_THREAD:
1048 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1051 panic("Bad RTP_PRIO %d", newrqtype);
1056 * The LWKT scheduler doesn't dive usched structures, give it a hint
1057 * on the relative priority of user threads running in the kernel.
1058 * The LWKT scheduler will always ensure that a user thread running
1059 * in the kernel will get cpu some time, regardless of its upri,
1060 * but can decide not to instantly switch from one kernel or user
1061 * mode user thread to a kernel-mode user thread when it has a less
1062 * desireable user priority.
1064 * td_upri has normal sense (higher values are more desireable), so
1067 lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1070 * The newpriority incorporates the queue type so do a simple masked
1071 * check to determine if the process has moved to another queue. If
1072 * it has, and it is currently on a run queue, then move it.
1074 * Since uload is ~PPQMASK masked, no modifications are necessary if
1075 * we end up in the same run queue.
1077 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1078 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1079 dfly_remrunqueue_locked(rdd, lp);
1080 lp->lwp_priority = newpriority;
1081 lp->lwp_rqtype = newrqtype;
1082 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1083 dfly_setrunqueue_locked(rdd, lp);
1086 lp->lwp_priority = newpriority;
1087 lp->lwp_rqtype = newrqtype;
1088 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1093 * In the same PPQ, uload cannot change.
1095 lp->lwp_priority = newpriority;
1101 * Adjust effective load.
1103 * Calculate load then scale up or down geometrically based on p_nice.
1104 * Processes niced up (positive) are less important, and processes
1105 * niced downard (negative) are more important. The higher the uload,
1106 * the more important the thread.
1108 /* 0-511, 0-100% cpu */
1109 delta_uload = lp->lwp_estcpu / NQS;
1110 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1113 delta_uload -= lp->lwp_uload;
1114 lp->lwp_uload += delta_uload;
1115 if (lp->lwp_mpflags & LWP_MP_ULOAD)
1116 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1119 * Determine if we need to reschedule the target cpu. This only
1120 * occurs if the LWP is already on a scheduler queue, which means
1121 * that idle cpu notification has already occured. At most we
1122 * need only issue a need_user_resched() on the appropriate cpu.
1124 * The LWP may be owned by a CPU different from the current one,
1125 * in which case dd->uschedcp may be modified without an MP lock
1126 * or a spinlock held. The worst that happens is that the code
1127 * below causes a spurious need_user_resched() on the target CPU
1128 * and dd->pri to be wrong for a short period of time, both of
1129 * which are harmless.
1131 * If checkpri is 0 we are adjusting the priority of the current
1132 * process, possibly higher (less desireable), so ignore the upri
1133 * check which will fail in that case.
1136 if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1138 (rdd->upri & ~PRIMASK) >
1139 (lp->lwp_priority & ~PRIMASK))) {
1140 if (rcpu == mycpu->gd_cpuid) {
1141 spin_unlock(&rdd->spin);
1142 need_user_resched();
1144 spin_unlock(&rdd->spin);
1145 lwkt_send_ipiq(globaldata_find(rcpu),
1146 dfly_need_user_resched_remote,
1150 spin_unlock(&rdd->spin);
1153 spin_unlock(&rdd->spin);
1160 dfly_yield(struct lwp *lp)
1162 if (lp->lwp_qcpu != mycpu->gd_cpuid)
1164 KKASSERT(lp == curthread->td_lwp);
1167 * Don't set need_user_resched() or mess with rrcount or anything.
1168 * the TDF flag will override everything as long as we release.
1170 atomic_set_int(&lp->lwp_thread->td_mpflags, TDF_MP_DIDYIELD);
1171 dfly_release_curproc(lp);
1175 * Thread was forcefully migrated to another cpu. Normally forced migrations
1176 * are used for iterations and the kernel returns to the original cpu before
1177 * returning and this is not needed. However, if the kernel migrates a
1178 * thread to another cpu and wants to leave it there, it has to call this
1181 * Note that the lwkt_migratecpu() function also released the thread, so
1182 * we don't have to worry about that.
1186 dfly_changedcpu(struct lwp *lp)
1188 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1189 dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid];
1192 spin_lock(&dd->spin);
1193 dfly_changeqcpu_locked(lp, dd, rdd);
1194 spin_unlock(&dd->spin);
1199 * Called from fork1() when a new child process is being created.
1201 * Give the child process an initial estcpu that is more batch then
1202 * its parent and dock the parent for the fork (but do not
1203 * reschedule the parent).
1207 * XXX lwp should be "spawning" instead of "forking"
1210 dfly_forking(struct lwp *plp, struct lwp *lp)
1213 * Put the child 4 queue slots (out of 32) higher than the parent
1214 * (less desireable than the parent).
1216 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1218 lp->lwp_estfast = 0;
1221 * Dock the parent a cost for the fork, protecting us from fork
1222 * bombs. If the parent is forking quickly make the child more
1225 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1229 * Called when a lwp is being removed from this scheduler, typically
1230 * during lwp_exit(). We have to clean out any ULOAD accounting before
1231 * we can let the lp go. The dd->spin lock is not needed for uload
1234 * Scheduler dequeueing has already occurred, no further action in that
1238 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1240 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1242 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1243 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1244 atomic_add_int(&dd->uload, -lp->lwp_uload);
1245 atomic_add_int(&dd->ucount, -1);
1246 atomic_add_int(&dfly_ucount, -1);
1251 * This function cannot block in any way, but spinlocks are ok.
1253 * Update the uload based on the state of the thread (whether it is going
1254 * to sleep or running again). The uload is meant to be a longer-term
1255 * load and not an instantanious load.
1258 dfly_uload_update(struct lwp *lp)
1260 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1262 if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1263 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1264 spin_lock(&dd->spin);
1265 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1266 atomic_set_int(&lp->lwp_mpflags,
1268 atomic_add_int(&dd->uload, lp->lwp_uload);
1269 atomic_add_int(&dd->ucount, 1);
1270 atomic_add_int(&dfly_ucount, 1);
1272 spin_unlock(&dd->spin);
1274 } else if (lp->lwp_slptime > 0) {
1275 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1276 spin_lock(&dd->spin);
1277 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1278 atomic_clear_int(&lp->lwp_mpflags,
1280 atomic_add_int(&dd->uload, -lp->lwp_uload);
1281 atomic_add_int(&dd->ucount, -1);
1282 atomic_add_int(&dfly_ucount, -1);
1284 spin_unlock(&dd->spin);
1290 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1291 * it selects a user process and returns it. If chklp is non-NULL and chklp
1292 * has a better or equal priority then the process that would otherwise be
1293 * chosen, NULL is returned.
1295 * Until we fix the RUNQ code the chklp test has to be strict or we may
1296 * bounce between processes trying to acquire the current process designation.
1298 * Must be called with rdd->spin locked. The spinlock is left intact through
1299 * the entire routine. dd->spin does not have to be locked.
1301 * If worst is non-zero this function finds the worst thread instead of the
1302 * best thread (used by the schedulerclock-based rover).
1306 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1307 struct lwp *chklp, int worst)
1317 rtqbits = rdd->rtqueuebits;
1318 tsqbits = rdd->queuebits;
1319 idqbits = rdd->idqueuebits;
1323 pri = bsrl(idqbits);
1324 q = &rdd->idqueues[pri];
1325 which = &rdd->idqueuebits;
1326 } else if (tsqbits) {
1327 pri = bsrl(tsqbits);
1328 q = &rdd->queues[pri];
1329 which = &rdd->queuebits;
1330 } else if (rtqbits) {
1331 pri = bsrl(rtqbits);
1332 q = &rdd->rtqueues[pri];
1333 which = &rdd->rtqueuebits;
1337 lp = TAILQ_LAST(q, rq);
1340 pri = bsfl(rtqbits);
1341 q = &rdd->rtqueues[pri];
1342 which = &rdd->rtqueuebits;
1343 } else if (tsqbits) {
1344 pri = bsfl(tsqbits);
1345 q = &rdd->queues[pri];
1346 which = &rdd->queuebits;
1347 } else if (idqbits) {
1348 pri = bsfl(idqbits);
1349 q = &rdd->idqueues[pri];
1350 which = &rdd->idqueuebits;
1354 lp = TAILQ_FIRST(q);
1356 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1359 * If the passed lwp <chklp> is reasonably close to the selected
1360 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1362 * Note that we must error on the side of <chklp> to avoid bouncing
1363 * between threads in the acquire code.
1366 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1370 KTR_COND_LOG(usched_chooseproc,
1371 lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1372 lp->lwp_proc->p_pid,
1373 lp->lwp_thread->td_gd->gd_cpuid,
1376 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1377 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1378 TAILQ_REMOVE(q, lp, lwp_procq);
1381 *which &= ~(1 << pri);
1384 * If we are choosing a process from rdd with the intent to
1385 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1389 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1390 atomic_add_int(&rdd->uload, -lp->lwp_uload);
1391 atomic_add_int(&rdd->ucount, -1);
1392 atomic_add_int(&dfly_ucount, -1);
1394 lp->lwp_qcpu = dd->cpuid;
1395 atomic_add_int(&dd->uload, lp->lwp_uload);
1396 atomic_add_int(&dd->ucount, 1);
1397 atomic_add_int(&dfly_ucount, 1);
1398 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1404 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1406 * Choose a cpu node to schedule lp on, hopefully nearby its current
1409 * We give the current node a modest advantage for obvious reasons.
1411 * We also give the node the thread was woken up FROM a slight advantage
1412 * in order to try to schedule paired threads which synchronize/block waiting
1413 * for each other fairly close to each other. Similarly in a network setting
1414 * this feature will also attempt to place a user process near the kernel
1415 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the
1416 * algorithm as it heuristically groups synchronizing processes for locality
1417 * of reference in multi-socket systems.
1419 * We check against running processes and give a big advantage if there
1422 * The caller will normally dfly_setrunqueue() lp on the returned queue.
1424 * When the topology is known choose a cpu whos group has, in aggregate,
1425 * has the lowest weighted load.
1429 dfly_choose_best_queue(struct lwp *lp)
1436 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1446 * When the topology is unknown choose a random cpu that is hopefully
1449 if (dd->cpunode == NULL)
1450 return (dfly_choose_queue_simple(dd, lp));
1455 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1456 wakemask = dfly_pcpu[wakecpu].cpumask;
1461 * When the topology is known choose a cpu whos group has, in
1462 * aggregate, has the lowest weighted load.
1464 cpup = root_cpu_node;
1469 * Degenerate case super-root
1471 if (cpup->child_no == 1) {
1472 cpup = cpup->child_node[0];
1479 if (cpup->child_no == 0) {
1480 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1485 lowest_load = 0x7FFFFFFF;
1487 for (n = 0; n < cpup->child_no; ++n) {
1489 * Accumulate load information for all cpus
1490 * which are members of this node.
1492 cpun = cpup->child_node[n];
1493 mask = cpun->members & usched_global_cpumask &
1494 smp_active_mask & lp->lwp_cpumask;
1502 cpuid = BSFCPUMASK(mask);
1503 rdd = &dfly_pcpu[cpuid];
1505 load += rdd->ucount * usched_dfly_weight3;
1507 if (rdd->uschedcp == NULL &&
1508 rdd->runqcount == 0 &&
1509 globaldata_find(cpuid)->gd_tdrunqcount == 0
1511 load -= usched_dfly_weight4;
1514 else if (rdd->upri > lp->lwp_priority + PPQ) {
1515 load -= usched_dfly_weight4 / 2;
1518 mask &= ~CPUMASK(cpuid);
1523 * Compensate if the lp is already accounted for in
1524 * the aggregate uload for this mask set. We want
1525 * to calculate the loads as if lp were not present,
1526 * otherwise the calculation is bogus.
1528 if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1529 (dd->cpumask & cpun->members)) {
1530 load -= lp->lwp_uload;
1531 load -= usched_dfly_weight3;
1537 * Advantage the cpu group (lp) is already on.
1539 if (cpun->members & dd->cpumask)
1540 load -= usched_dfly_weight1;
1543 * Advantage the cpu group we want to pair (lp) to,
1544 * but don't let it go to the exact same cpu as
1545 * the wakecpu target.
1547 * We do this by checking whether cpun is a
1548 * terminal node or not. All cpun's at the same
1549 * level will either all be terminal or all not
1552 * If it is and we match we disadvantage the load.
1553 * If it is and we don't match we advantage the load.
1555 * Also note that we are effectively disadvantaging
1556 * all-but-one by the same amount, so it won't effect
1557 * the weight1 factor for the all-but-one nodes.
1559 if (cpun->members & wakemask) {
1560 if (cpun->child_no != 0) {
1562 load -= usched_dfly_weight2;
1564 if (usched_dfly_features & 0x10)
1565 load += usched_dfly_weight2;
1567 load -= usched_dfly_weight2;
1572 * Calculate the best load
1574 if (cpub == NULL || lowest_load > load ||
1575 (lowest_load == load &&
1576 (cpun->members & dd->cpumask))
1584 if (usched_dfly_chooser)
1585 kprintf("lp %02d->%02d %s\n",
1586 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1591 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1593 * Choose the worst queue close to dd's cpu node with a non-empty runq
1594 * that is NOT dd. Also require that the moving of the highest-load thread
1595 * from rdd to dd does not cause the uload's to cross each other.
1597 * This is used by the thread chooser when the current cpu's queues are
1598 * empty to steal a thread from another cpu's queue. We want to offload
1599 * the most heavily-loaded queue.
1603 dfly_choose_worst_queue(dfly_pcpu_t dd)
1621 * When the topology is unknown choose a random cpu that is hopefully
1624 if (dd->cpunode == NULL) {
1629 * When the topology is known choose a cpu whos group has, in
1630 * aggregate, has the lowest weighted load.
1632 cpup = root_cpu_node;
1636 * Degenerate case super-root
1638 if (cpup->child_no == 1) {
1639 cpup = cpup->child_node[0];
1646 if (cpup->child_no == 0) {
1647 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1654 for (n = 0; n < cpup->child_no; ++n) {
1656 * Accumulate load information for all cpus
1657 * which are members of this node.
1659 cpun = cpup->child_node[n];
1660 mask = cpun->members & usched_global_cpumask &
1668 cpuid = BSFCPUMASK(mask);
1669 rdd = &dfly_pcpu[cpuid];
1671 load += rdd->ucount * usched_dfly_weight3;
1672 if (rdd->uschedcp == NULL &&
1673 rdd->runqcount == 0 &&
1674 globaldata_find(cpuid)->gd_tdrunqcount == 0
1676 load -= usched_dfly_weight4;
1679 else if (rdd->upri > dd->upri + PPQ) {
1680 load -= usched_dfly_weight4 / 2;
1683 mask &= ~CPUMASK(cpuid);
1689 * Prefer candidates which are somewhat closer to
1692 if (dd->cpumask & cpun->members)
1693 load += usched_dfly_weight1;
1696 * The best candidate is the one with the worst
1699 if (cpub == NULL || highest_load < load) {
1700 highest_load = load;
1708 * We never return our own node (dd), and only return a remote
1709 * node if it's load is significantly worse than ours (i.e. where
1710 * stealing a thread would be considered reasonable).
1712 * This also helps us avoid breaking paired threads apart which
1713 * can have disastrous effects on performance.
1720 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1722 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1724 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1727 if (rdd->uload - hpri < dd->uload + hpri)
1735 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1743 * Fallback to the original heuristic, select random cpu,
1744 * first checking cpus not currently running a user thread.
1747 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1748 mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1749 smp_active_mask & usched_global_cpumask;
1752 tmpmask = ~(CPUMASK(cpuid) - 1);
1754 cpuid = BSFCPUMASK(mask & tmpmask);
1756 cpuid = BSFCPUMASK(mask);
1757 rdd = &dfly_pcpu[cpuid];
1759 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1761 mask &= ~CPUMASK(cpuid);
1765 * Then cpus which might have a currently running lp
1767 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1768 mask = dfly_curprocmask & dfly_rdyprocmask &
1769 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1772 tmpmask = ~(CPUMASK(cpuid) - 1);
1774 cpuid = BSFCPUMASK(mask & tmpmask);
1776 cpuid = BSFCPUMASK(mask);
1777 rdd = &dfly_pcpu[cpuid];
1779 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1781 mask &= ~CPUMASK(cpuid);
1785 * If we cannot find a suitable cpu we reload from dfly_scancpu
1786 * and round-robin. Other cpus will pickup as they release their
1787 * current lwps or become ready.
1789 * Avoid a degenerate system lockup case if usched_global_cpumask
1790 * is set to 0 or otherwise does not cover lwp_cpumask.
1792 * We only kick the target helper thread in this case, we do not
1793 * set the user resched flag because
1795 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1796 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1798 rdd = &dfly_pcpu[cpuid];
1805 dfly_need_user_resched_remote(void *dummy)
1807 globaldata_t gd = mycpu;
1808 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
1811 * Flag reschedule needed
1813 need_user_resched();
1816 * If no user thread is currently running we need to kick the helper
1817 * on our cpu to recover. Otherwise the cpu will never schedule
1820 * We cannot schedule the process ourselves because this is an
1821 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1823 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1825 if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1826 atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1827 wakeup_mycpu(&dd->helper_thread);
1832 * dfly_remrunqueue_locked() removes a given process from the run queue
1833 * that it is on, clearing the queue busy bit if it becomes empty.
1835 * Note that user process scheduler is different from the LWKT schedule.
1836 * The user process scheduler only manages user processes but it uses LWKT
1837 * underneath, and a user process operating in the kernel will often be
1838 * 'released' from our management.
1840 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes
1841 * to sleep or the lwp is moved to a different runq.
1844 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1850 KKASSERT(rdd->runqcount >= 0);
1852 pri = lp->lwp_rqindex;
1854 switch(lp->lwp_rqtype) {
1855 case RTP_PRIO_NORMAL:
1856 q = &rdd->queues[pri];
1857 which = &rdd->queuebits;
1859 case RTP_PRIO_REALTIME:
1861 q = &rdd->rtqueues[pri];
1862 which = &rdd->rtqueuebits;
1865 q = &rdd->idqueues[pri];
1866 which = &rdd->idqueuebits;
1869 panic("remrunqueue: invalid rtprio type");
1872 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1873 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1874 TAILQ_REMOVE(q, lp, lwp_procq);
1876 if (TAILQ_EMPTY(q)) {
1877 KASSERT((*which & (1 << pri)) != 0,
1878 ("remrunqueue: remove from empty queue"));
1879 *which &= ~(1 << pri);
1884 * dfly_setrunqueue_locked()
1886 * Add a process whos rqtype and rqindex had previously been calculated
1887 * onto the appropriate run queue. Determine if the addition requires
1888 * a reschedule on a cpu and return the cpuid or -1.
1890 * NOTE: Lower priorities are better priorities.
1892 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1893 * sum of the rough lwp_priority for all running and runnable
1894 * processes. Lower priority processes (higher lwp_priority
1895 * values) actually DO count as more load, not less, because
1896 * these are the programs which require the most care with
1897 * regards to cpu selection.
1900 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1906 KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1908 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1909 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1910 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1911 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1912 atomic_add_int(&dfly_ucount, 1);
1915 pri = lp->lwp_rqindex;
1917 switch(lp->lwp_rqtype) {
1918 case RTP_PRIO_NORMAL:
1919 q = &rdd->queues[pri];
1920 which = &rdd->queuebits;
1922 case RTP_PRIO_REALTIME:
1924 q = &rdd->rtqueues[pri];
1925 which = &rdd->rtqueuebits;
1928 q = &rdd->idqueues[pri];
1929 which = &rdd->idqueuebits;
1932 panic("remrunqueue: invalid rtprio type");
1937 * Place us on the selected queue. Determine if we should be
1938 * placed at the head of the queue or at the end.
1940 * We are placed at the tail if our round-robin count has expired,
1941 * or is about to expire and the system thinks its a good place to
1942 * round-robin, or there is already a next thread on the queue
1943 * (it might be trying to pick up where it left off and we don't
1944 * want to interfere).
1946 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1947 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1950 if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1951 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1952 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1955 atomic_clear_int(&lp->lwp_thread->td_mpflags,
1956 TDF_MP_BATCH_DEMARC);
1957 lp->lwp_rrcount = 0;
1958 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1961 lp->lwp_rrcount = 0;
1962 TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1968 * For SMP systems a user scheduler helper thread is created for each
1969 * cpu and is used to allow one cpu to wakeup another for the purposes of
1970 * scheduling userland threads from setrunqueue().
1972 * UP systems do not need the helper since there is only one cpu.
1974 * We can't use the idle thread for this because we might block.
1975 * Additionally, doing things this way allows us to HLT idle cpus
1979 dfly_helper_thread(void *dummy)
1989 cpuid = gd->gd_cpuid; /* doesn't change */
1990 mask = gd->gd_cpumask; /* doesn't change */
1991 dd = &dfly_pcpu[cpuid];
1994 * Since we only want to be woken up only when no user processes
1995 * are scheduled on a cpu, run at an ultra low priority.
1997 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1999 tsleep(&dd->helper_thread, 0, "schslp", 0);
2003 * We use the LWKT deschedule-interlock trick to avoid racing
2004 * dfly_rdyprocmask. This means we cannot block through to the
2005 * manual lwkt_switch() call we make below.
2008 tsleep_interlock(&dd->helper_thread, 0);
2010 spin_lock(&dd->spin);
2012 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2013 clear_user_resched(); /* This satisfied the reschedule request */
2015 dd->rrcount = 0; /* Reset the round-robin counter */
2018 if (dd->runqcount || dd->uschedcp != NULL) {
2020 * Threads are available. A thread may or may not be
2021 * currently scheduled. Get the best thread already queued
2024 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
2026 atomic_set_cpumask(&dfly_curprocmask, mask);
2027 dd->upri = nlp->lwp_priority;
2030 dd->rrcount = 0; /* reset round robin */
2032 spin_unlock(&dd->spin);
2033 lwkt_acquire(nlp->lwp_thread);
2034 lwkt_schedule(nlp->lwp_thread);
2037 * This situation should not occur because we had
2038 * at least one thread available.
2040 spin_unlock(&dd->spin);
2042 } else if (usched_dfly_features & 0x01) {
2044 * This cpu is devoid of runnable threads, steal a thread
2045 * from another cpu. Since we're stealing, might as well
2046 * load balance at the same time.
2048 * We choose the highest-loaded thread from the worst queue.
2050 * NOTE! This function only returns a non-NULL rdd when
2051 * another cpu's queue is obviously overloaded. We
2052 * do not want to perform the type of rebalancing
2053 * the schedclock does here because it would result
2054 * in insane process pulling when 'steady' state is
2055 * partially unbalanced (e.g. 6 runnables and only
2058 rdd = dfly_choose_worst_queue(dd);
2059 if (rdd && spin_trylock(&rdd->spin)) {
2060 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2061 spin_unlock(&rdd->spin);
2066 atomic_set_cpumask(&dfly_curprocmask, mask);
2067 dd->upri = nlp->lwp_priority;
2070 dd->rrcount = 0; /* reset round robin */
2072 spin_unlock(&dd->spin);
2073 lwkt_acquire(nlp->lwp_thread);
2074 lwkt_schedule(nlp->lwp_thread);
2077 * Leave the thread on our run queue. Another
2078 * scheduler will try to pull it later.
2080 spin_unlock(&dd->spin);
2084 * devoid of runnable threads and not allowed to steal
2087 spin_unlock(&dd->spin);
2091 * We're descheduled unless someone scheduled us. Switch away.
2092 * Exiting the critical section will cause splz() to be called
2093 * for us if interrupts and such are pending.
2096 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2102 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2106 new_val = usched_dfly_stick_to_level;
2108 error = sysctl_handle_int(oidp, &new_val, 0, req);
2109 if (error != 0 || req->newptr == NULL)
2111 if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2113 usched_dfly_stick_to_level = new_val;
2119 * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2120 * Note that curprocmask bit 0 has already been cleared by rqinit() and
2121 * we should not mess with it further.
2124 usched_dfly_cpu_init(void)
2129 int smt_not_supported = 0;
2130 int cache_coherent_not_supported = 0;
2133 kprintf("Start scheduler helpers on cpus:\n");
2135 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2136 usched_dfly_sysctl_tree =
2137 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2138 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2139 "usched_dfly", CTLFLAG_RD, 0, "");
2141 for (i = 0; i < ncpus; ++i) {
2142 dfly_pcpu_t dd = &dfly_pcpu[i];
2143 cpumask_t mask = CPUMASK(i);
2145 if ((mask & smp_active_mask) == 0)
2148 spin_init(&dd->spin);
2149 dd->cpunode = get_cpu_node_by_cpuid(i);
2151 dd->cpumask = CPUMASK(i);
2152 for (j = 0; j < NQS; j++) {
2153 TAILQ_INIT(&dd->queues[j]);
2154 TAILQ_INIT(&dd->rtqueues[j]);
2155 TAILQ_INIT(&dd->idqueues[j]);
2157 atomic_clear_cpumask(&dfly_curprocmask, 1);
2159 if (dd->cpunode == NULL) {
2160 smt_not_supported = 1;
2161 cache_coherent_not_supported = 1;
2163 kprintf ("\tcpu%d - WARNING: No CPU NODE "
2164 "found for cpu\n", i);
2166 switch (dd->cpunode->type) {
2169 kprintf ("\tcpu%d - HyperThreading "
2170 "available. Core siblings: ",
2174 smt_not_supported = 1;
2177 kprintf ("\tcpu%d - No HT available, "
2178 "multi-core/physical "
2179 "cpu. Physical siblings: ",
2183 smt_not_supported = 1;
2186 kprintf ("\tcpu%d - No HT available, "
2187 "single-core/physical cpu. "
2188 "Package Siblings: ",
2192 /* Let's go for safe defaults here */
2193 smt_not_supported = 1;
2194 cache_coherent_not_supported = 1;
2196 kprintf ("\tcpu%d - Unknown cpunode->"
2197 "type=%u. Siblings: ",
2199 (u_int)dd->cpunode->type);
2204 if (dd->cpunode->parent_node != NULL) {
2205 CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2206 kprintf("cpu%d ", cpuid);
2209 kprintf(" no siblings\n");
2214 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2215 0, i, "usched %d", i);
2218 * Allow user scheduling on the target cpu. cpu #0 has already
2219 * been enabled in rqinit().
2222 atomic_clear_cpumask(&dfly_curprocmask, mask);
2223 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2224 dd->upri = PRIBASE_NULL;
2228 /* usched_dfly sysctl configurable parameters */
2230 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2231 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2232 OID_AUTO, "rrinterval", CTLFLAG_RW,
2233 &usched_dfly_rrinterval, 0, "");
2234 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2235 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2236 OID_AUTO, "decay", CTLFLAG_RW,
2237 &usched_dfly_decay, 0, "Extra decay when not running");
2239 /* Add enable/disable option for SMT scheduling if supported */
2240 if (smt_not_supported) {
2241 usched_dfly_smt = 0;
2242 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2243 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2244 OID_AUTO, "smt", CTLFLAG_RD,
2245 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2247 usched_dfly_smt = 1;
2248 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2249 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2250 OID_AUTO, "smt", CTLFLAG_RW,
2251 &usched_dfly_smt, 0, "Enable SMT scheduling");
2255 * Add enable/disable option for cache coherent scheduling
2258 if (cache_coherent_not_supported) {
2259 usched_dfly_cache_coherent = 0;
2260 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2261 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2262 OID_AUTO, "cache_coherent", CTLFLAG_RD,
2264 "Cache coherence NOT SUPPORTED");
2266 usched_dfly_cache_coherent = 1;
2267 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2268 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2269 OID_AUTO, "cache_coherent", CTLFLAG_RW,
2270 &usched_dfly_cache_coherent, 0,
2271 "Enable/Disable cache coherent scheduling");
2273 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2274 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2275 OID_AUTO, "weight1", CTLFLAG_RW,
2276 &usched_dfly_weight1, 200,
2277 "Weight selection for current cpu");
2279 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2280 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2281 OID_AUTO, "weight2", CTLFLAG_RW,
2282 &usched_dfly_weight2, 180,
2283 "Weight selection for wakefrom cpu");
2285 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2286 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2287 OID_AUTO, "weight3", CTLFLAG_RW,
2288 &usched_dfly_weight3, 40,
2289 "Weight selection for num threads on queue");
2291 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2292 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2293 OID_AUTO, "weight4", CTLFLAG_RW,
2294 &usched_dfly_weight4, 160,
2295 "Availability of other idle cpus");
2297 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2298 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2299 OID_AUTO, "fast_resched", CTLFLAG_RW,
2300 &usched_dfly_fast_resched, 0,
2301 "Availability of other idle cpus");
2303 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2304 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2305 OID_AUTO, "features", CTLFLAG_RW,
2306 &usched_dfly_features, 0x8F,
2307 "Allow pulls into empty queues");
2309 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2310 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2311 OID_AUTO, "swmask", CTLFLAG_RW,
2312 &usched_dfly_swmask, ~PPQMASK,
2313 "Queue mask to force thread switch");
2316 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2317 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2318 OID_AUTO, "stick_to_level",
2319 CTLTYPE_INT | CTLFLAG_RW,
2320 NULL, sizeof usched_dfly_stick_to_level,
2321 sysctl_usched_dfly_stick_to_level, "I",
2322 "Stick a process to this level. See sysctl"
2323 "paremter hw.cpu_topology.level_description");
2327 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2328 usched_dfly_cpu_init, NULL)