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 atomic_clear_int(&lp->lwp_thread->td_mpflags,
390 switch(lp->lwp_rqtype) {
391 case RTP_PRIO_NORMAL:
392 tsqbits = dd->queuebits;
393 spin_unlock(&dd->spin);
395 lp->lwp_rrcount = usched_dfly_rrinterval;
397 lp->lwp_rqindex = bsrl(tsqbits);
400 spin_unlock(&dd->spin);
403 lwkt_deschedule(lp->lwp_thread);
404 dfly_setrunqueue_dd(dd, lp);
407 dd = &dfly_pcpu[gd->gd_cpuid];
412 * Can we steal the current designated user thread?
414 * If we do the other thread will stall when it tries to
415 * return to userland, possibly rescheduling elsewhere.
417 * It is important to do a masked test to avoid the edge
418 * case where two near-equal-priority threads are constantly
419 * interrupting each other.
421 * In the exact match case another thread has already gained
422 * uschedcp and lowered its priority, if we steal it the
423 * other thread will stay stuck on the LWKT runq and not
424 * push to another cpu. So don't steal on equal-priority even
425 * though it might appear to be more beneficial due to not
426 * having to switch back to the other thread's context.
428 * usched_dfly_fast_resched requires that two threads be
429 * significantly far apart in priority in order to interrupt.
431 * If better but not sufficiently far apart, the current
432 * uschedcp will be interrupted at the next scheduler clock.
435 (dd->upri & ~PPQMASK) >
436 (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
438 dd->upri = lp->lwp_priority;
439 KKASSERT(lp->lwp_qcpu == dd->cpuid);
440 spin_unlock(&dd->spin);
444 * We are not the current lwp, figure out the best cpu
445 * to run on (our current cpu will be given significant
446 * weight). Loop on cpu change.
448 if ((usched_dfly_features & 0x02) &&
449 force_resched == 0 &&
450 (rdd = dfly_choose_best_queue(lp)) != dd) {
451 dfly_changeqcpu_locked(lp, dd, rdd);
452 spin_unlock(&dd->spin);
453 lwkt_deschedule(lp->lwp_thread);
454 dfly_setrunqueue_dd(rdd, lp);
457 dd = &dfly_pcpu[gd->gd_cpuid];
462 * We cannot become the current lwp, place the lp on the
463 * run-queue of this or another cpu and deschedule ourselves.
465 * When we are reactivated we will have another chance.
467 * Reload after a switch or setrunqueue/switch possibly
468 * moved us to another cpu.
470 spin_unlock(&dd->spin);
471 lwkt_deschedule(lp->lwp_thread);
472 dfly_setrunqueue_dd(dd, lp);
475 dd = &dfly_pcpu[gd->gd_cpuid];
479 * Make sure upri is synchronized, then yield to LWKT threads as
480 * needed before returning. This could result in another reschedule.
485 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
489 * DFLY_RELEASE_CURPROC
491 * This routine detaches the current thread from the userland scheduler,
492 * usually because the thread needs to run or block in the kernel (at
493 * kernel priority) for a while.
495 * This routine is also responsible for selecting a new thread to
496 * make the current thread.
498 * NOTE: This implementation differs from the dummy example in that
499 * dfly_select_curproc() is able to select the current process, whereas
500 * dummy_select_curproc() is not able to select the current process.
501 * This means we have to NULL out uschedcp.
503 * Additionally, note that we may already be on a run queue if releasing
504 * via the lwkt_switch() in dfly_setrunqueue().
507 dfly_release_curproc(struct lwp *lp)
509 globaldata_t gd = mycpu;
510 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
513 * Make sure td_wakefromcpu is defaulted. This will be overwritten
516 if (dd->uschedcp == lp) {
517 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
518 spin_lock(&dd->spin);
519 if (dd->uschedcp == lp) {
520 dd->uschedcp = NULL; /* don't let lp be selected */
521 dd->upri = PRIBASE_NULL;
522 atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
523 spin_unlock(&dd->spin);
524 dfly_select_curproc(gd);
526 spin_unlock(&dd->spin);
532 * DFLY_SELECT_CURPROC
534 * Select a new current process for this cpu and clear any pending user
535 * reschedule request. The cpu currently has no current process.
537 * This routine is also responsible for equal-priority round-robining,
538 * typically triggered from dfly_schedulerclock(). In our dummy example
539 * all the 'user' threads are LWKT scheduled all at once and we just
540 * call lwkt_switch().
542 * The calling process is not on the queue and cannot be selected.
546 dfly_select_curproc(globaldata_t gd)
548 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
550 int cpuid = gd->gd_cpuid;
554 spin_lock(&dd->spin);
555 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
558 atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
559 dd->upri = nlp->lwp_priority;
562 dd->rrcount = 0; /* reset round robin */
564 spin_unlock(&dd->spin);
565 lwkt_acquire(nlp->lwp_thread);
566 lwkt_schedule(nlp->lwp_thread);
568 spin_unlock(&dd->spin);
574 * Place the specified lwp on the user scheduler's run queue. This routine
575 * must be called with the thread descheduled. The lwp must be runnable.
576 * It must not be possible for anyone else to explicitly schedule this thread.
578 * The thread may be the current thread as a special case.
581 dfly_setrunqueue(struct lwp *lp)
587 * First validate the process LWKT state.
589 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
590 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
591 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
592 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
593 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
596 * NOTE: dd/rdd do not necessarily represent the current cpu.
597 * Instead they may represent the cpu the thread was last
598 * scheduled on or inherited by its parent.
600 dd = &dfly_pcpu[lp->lwp_qcpu];
604 * This process is not supposed to be scheduled anywhere or assigned
605 * as the current process anywhere. Assert the condition.
607 KKASSERT(rdd->uschedcp != lp);
610 * Ok, we have to setrunqueue some target cpu and request a reschedule
613 * We have to choose the best target cpu. It might not be the current
614 * target even if the current cpu has no running user thread (for
615 * example, because the current cpu might be a hyperthread and its
616 * sibling has a thread assigned).
618 * If we just forked it is most optimal to run the child on the same
619 * cpu just in case the parent decides to wait for it (thus getting
620 * off that cpu). As long as there is nothing else runnable on the
621 * cpu, that is. If we did this unconditionally a parent forking
622 * multiple children before waiting (e.g. make -j N) leaves other
623 * cpus idle that could be working.
625 if (lp->lwp_forked) {
627 if (usched_dfly_features & 0x20)
628 rdd = dfly_choose_best_queue(lp);
629 else if (usched_dfly_features & 0x40)
630 rdd = &dfly_pcpu[lp->lwp_qcpu];
631 else if (usched_dfly_features & 0x80)
632 rdd = dfly_choose_queue_simple(rdd, lp);
633 else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
634 rdd = dfly_choose_best_queue(lp);
636 rdd = &dfly_pcpu[lp->lwp_qcpu];
638 rdd = dfly_choose_best_queue(lp);
639 /* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
641 if (lp->lwp_qcpu != rdd->cpuid) {
642 spin_lock(&dd->spin);
643 dfly_changeqcpu_locked(lp, dd, rdd);
644 spin_unlock(&dd->spin);
646 dfly_setrunqueue_dd(rdd, lp);
650 * Change qcpu to rdd->cpuid. The dd the lp is CURRENTLY on must be
651 * spin-locked on-call. rdd does not have to be.
654 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
656 if (lp->lwp_qcpu != rdd->cpuid) {
657 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
658 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
659 atomic_add_int(&dd->uload, -lp->lwp_uload);
660 atomic_add_int(&dd->ucount, -1);
661 atomic_add_int(&dfly_ucount, -1);
663 lp->lwp_qcpu = rdd->cpuid;
668 * Place lp on rdd's runqueue. Nothing is locked on call. This function
669 * also performs all necessary ancillary notification actions.
672 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
677 * We might be moving the lp to another cpu's run queue, and once
678 * on the runqueue (even if it is our cpu's), another cpu can rip
681 * TDF_MIGRATING might already be set if this is part of a
682 * remrunqueue+setrunqueue sequence.
684 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
685 lwkt_giveaway(lp->lwp_thread);
687 rgd = globaldata_find(rdd->cpuid);
690 * We lose control of the lp the moment we release the spinlock
691 * after having placed it on the queue. i.e. another cpu could pick
692 * it up, or it could exit, or its priority could be further
693 * adjusted, or something like that.
695 * WARNING! rdd can point to a foreign cpu!
697 spin_lock(&rdd->spin);
698 dfly_setrunqueue_locked(rdd, lp);
701 * Potentially interrupt the currently-running thread
703 if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
705 * Currently running thread is better or same, do not
708 spin_unlock(&rdd->spin);
709 } else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
710 usched_dfly_fast_resched) {
712 * Currently running thread is not better, but not so bad
713 * that we need to interrupt it. Let it run for one more
717 rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
718 rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
720 spin_unlock(&rdd->spin);
721 } else if (rgd == mycpu) {
723 * We should interrupt the currently running thread, which
724 * is on the current cpu.
726 spin_unlock(&rdd->spin);
727 if (rdd->uschedcp == NULL) {
728 wakeup_mycpu(&rdd->helper_thread); /* XXX */
735 * We should interrupt the currently running thread, which
736 * is on a different cpu.
738 spin_unlock(&rdd->spin);
739 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
744 * This routine is called from a systimer IPI. It MUST be MP-safe and
745 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
750 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
752 globaldata_t gd = mycpu;
753 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
756 * Spinlocks also hold a critical section so there should not be
759 KKASSERT(gd->gd_spinlocks == 0);
765 * Do we need to round-robin? We round-robin 10 times a second.
766 * This should only occur for cpu-bound batch processes.
768 if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
769 lp->lwp_thread->td_wakefromcpu = -1;
774 * Adjust estcpu upward using a real time equivalent calculation,
775 * and recalculate lp's priority.
777 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
778 dfly_resetpriority(lp);
781 * Rebalance two cpus every 8 ticks, pulling the worst thread
782 * from the worst cpu's queue into a rotating cpu number.
784 * This mechanic is needed because the push algorithms can
785 * steady-state in an non-optimal configuration. We need to mix it
786 * up a little, even if it means breaking up a paired thread, so
787 * the push algorithms can rebalance the degenerate conditions.
788 * This portion of the algorithm exists to ensure stability at the
789 * selected weightings.
791 * Because we might be breaking up optimal conditions we do not want
792 * to execute this too quickly, hence we only rebalance approximately
793 * ~7-8 times per second. The push's, on the otherhand, are capable
794 * moving threads to other cpus at a much higher rate.
796 * We choose the most heavily loaded thread from the worst queue
797 * in order to ensure that multiple heavy-weight threads on the same
798 * queue get broken up, and also because these threads are the most
799 * likely to be able to remain in place. Hopefully then any pairings,
800 * if applicable, migrate to where these threads are.
802 if ((usched_dfly_features & 0x04) &&
803 ((u_int)sched_ticks & 7) == 0 &&
804 (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
811 rdd = dfly_choose_worst_queue(dd);
813 spin_lock(&dd->spin);
814 if (spin_trylock(&rdd->spin)) {
815 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
816 spin_unlock(&rdd->spin);
818 spin_unlock(&dd->spin);
820 spin_unlock(&dd->spin);
826 /* dd->spin held if nlp != NULL */
829 * Either schedule it or add it to our queue.
832 (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
833 atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
834 dd->upri = nlp->lwp_priority;
837 dd->rrcount = 0; /* reset round robin */
839 spin_unlock(&dd->spin);
840 lwkt_acquire(nlp->lwp_thread);
841 lwkt_schedule(nlp->lwp_thread);
843 dfly_setrunqueue_locked(dd, nlp);
844 spin_unlock(&dd->spin);
850 * Called from acquire and from kern_synch's one-second timer (one of the
851 * callout helper threads) with a critical section held.
853 * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
854 * overall system load.
856 * Note that no recalculation occurs for a process which sleeps and wakes
857 * up in the same tick. That is, a system doing thousands of context
858 * switches per second will still only do serious estcpu calculations
859 * ESTCPUFREQ times per second.
863 dfly_recalculate_estcpu(struct lwp *lp)
865 globaldata_t gd = mycpu;
873 * We have to subtract periodic to get the last schedclock
874 * timeout time, otherwise we would get the upcoming timeout.
875 * Keep in mind that a process can migrate between cpus and
876 * while the scheduler clock should be very close, boundary
877 * conditions could lead to a small negative delta.
879 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
881 if (lp->lwp_slptime > 1) {
883 * Too much time has passed, do a coarse correction.
885 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
886 dfly_resetpriority(lp);
887 lp->lwp_cpbase = cpbase;
890 } else if (lp->lwp_cpbase != cpbase) {
892 * Adjust estcpu if we are in a different tick. Don't waste
893 * time if we are in the same tick.
895 * First calculate the number of ticks in the measurement
896 * interval. The ttlticks calculation can wind up 0 due to
897 * a bug in the handling of lwp_slptime (as yet not found),
898 * so make sure we do not get a divide by 0 panic.
900 ttlticks = (cpbase - lp->lwp_cpbase) /
901 gd->gd_schedclock.periodic;
902 if ((ssysclock_t)ttlticks < 0) {
904 lp->lwp_cpbase = cpbase;
908 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
911 * Calculate the percentage of one cpu being used then
912 * compensate for any system load in excess of ncpus.
914 * For example, if we have 8 cores and 16 running cpu-bound
915 * processes then all things being equal each process will
916 * get 50% of one cpu. We need to pump this value back
917 * up to 100% so the estcpu calculation properly adjusts
918 * the process's dynamic priority.
920 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
922 estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
923 ucount = dfly_ucount;
924 if (ucount > ncpus) {
925 estcpu += estcpu * (ucount - ncpus) / ncpus;
928 if (usched_dfly_debug == lp->lwp_proc->p_pid) {
929 kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
930 lp->lwp_proc->p_pid, lp,
931 estcpu, lp->lwp_estcpu,
932 lp->lwp_cpticks, ttlticks);
936 * Adjust lp->lwp_esetcpu. The decay factor determines how
937 * quickly lwp_estcpu collapses to its realtime calculation.
938 * A slower collapse gives us a more accurate number over
939 * the long term but can create problems with bursty threads
940 * or threads which become cpu hogs.
942 * To solve this problem, newly started lwps and lwps which
943 * are restarting after having been asleep for a while are
944 * given a much, much faster decay in order to quickly
945 * detect whether they become cpu-bound.
947 * NOTE: p_nice is accounted for in dfly_resetpriority(),
948 * and not here, but we must still ensure that a
949 * cpu-bound nice -20 process does not completely
950 * override a cpu-bound nice +20 process.
952 * NOTE: We must use ESTCPULIM() here to deal with any
955 decay_factor = usched_dfly_decay;
956 if (decay_factor < 1)
958 if (decay_factor > 1024)
961 if (lp->lwp_estfast < usched_dfly_decay) {
963 lp->lwp_estcpu = ESTCPULIM(
964 (lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
965 (lp->lwp_estfast + 1));
967 lp->lwp_estcpu = ESTCPULIM(
968 (lp->lwp_estcpu * decay_factor + estcpu) /
972 if (usched_dfly_debug == lp->lwp_proc->p_pid)
973 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
974 dfly_resetpriority(lp);
975 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
981 * Compute the priority of a process when running in user mode.
982 * Arrange to reschedule if the resulting priority is better
983 * than that of the current process.
985 * This routine may be called with any process.
987 * This routine is called by fork1() for initial setup with the process
988 * of the run queue, and also may be called normally with the process on or
992 dfly_resetpriority(struct lwp *lp)
1005 * Lock the scheduler (lp) belongs to. This can be on a different
1006 * cpu. Handle races. This loop breaks out with the appropriate
1010 rcpu = lp->lwp_qcpu;
1012 rdd = &dfly_pcpu[rcpu];
1013 spin_lock(&rdd->spin);
1014 if (rcpu == lp->lwp_qcpu)
1016 spin_unlock(&rdd->spin);
1020 * Calculate the new priority and queue type
1022 newrqtype = lp->lwp_rtprio.type;
1025 case RTP_PRIO_REALTIME:
1027 newpriority = PRIBASE_REALTIME +
1028 (lp->lwp_rtprio.prio & PRIMASK);
1030 case RTP_PRIO_NORMAL:
1034 estcpu = lp->lwp_estcpu;
1037 * p_nice piece Adds (0-40) * 2 0-80
1038 * estcpu Adds 16384 * 4 / 512 0-128
1040 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1041 newpriority += estcpu * PPQ / ESTCPUPPQ;
1042 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1043 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1044 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1047 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1049 case RTP_PRIO_THREAD:
1050 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1053 panic("Bad RTP_PRIO %d", newrqtype);
1058 * The LWKT scheduler doesn't dive usched structures, give it a hint
1059 * on the relative priority of user threads running in the kernel.
1060 * The LWKT scheduler will always ensure that a user thread running
1061 * in the kernel will get cpu some time, regardless of its upri,
1062 * but can decide not to instantly switch from one kernel or user
1063 * mode user thread to a kernel-mode user thread when it has a less
1064 * desireable user priority.
1066 * td_upri has normal sense (higher values are more desireable), so
1069 lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1072 * The newpriority incorporates the queue type so do a simple masked
1073 * check to determine if the process has moved to another queue. If
1074 * it has, and it is currently on a run queue, then move it.
1076 * Since uload is ~PPQMASK masked, no modifications are necessary if
1077 * we end up in the same run queue.
1079 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1080 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1081 dfly_remrunqueue_locked(rdd, lp);
1082 lp->lwp_priority = newpriority;
1083 lp->lwp_rqtype = newrqtype;
1084 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1085 dfly_setrunqueue_locked(rdd, lp);
1088 lp->lwp_priority = newpriority;
1089 lp->lwp_rqtype = newrqtype;
1090 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1095 * In the same PPQ, uload cannot change.
1097 lp->lwp_priority = newpriority;
1103 * Adjust effective load.
1105 * Calculate load then scale up or down geometrically based on p_nice.
1106 * Processes niced up (positive) are less important, and processes
1107 * niced downard (negative) are more important. The higher the uload,
1108 * the more important the thread.
1110 /* 0-511, 0-100% cpu */
1111 delta_uload = lp->lwp_estcpu / NQS;
1112 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1115 delta_uload -= lp->lwp_uload;
1116 lp->lwp_uload += delta_uload;
1117 if (lp->lwp_mpflags & LWP_MP_ULOAD)
1118 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1121 * Determine if we need to reschedule the target cpu. This only
1122 * occurs if the LWP is already on a scheduler queue, which means
1123 * that idle cpu notification has already occured. At most we
1124 * need only issue a need_user_resched() on the appropriate cpu.
1126 * The LWP may be owned by a CPU different from the current one,
1127 * in which case dd->uschedcp may be modified without an MP lock
1128 * or a spinlock held. The worst that happens is that the code
1129 * below causes a spurious need_user_resched() on the target CPU
1130 * and dd->pri to be wrong for a short period of time, both of
1131 * which are harmless.
1133 * If checkpri is 0 we are adjusting the priority of the current
1134 * process, possibly higher (less desireable), so ignore the upri
1135 * check which will fail in that case.
1138 if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1140 (rdd->upri & ~PRIMASK) >
1141 (lp->lwp_priority & ~PRIMASK))) {
1142 if (rcpu == mycpu->gd_cpuid) {
1143 spin_unlock(&rdd->spin);
1144 need_user_resched();
1146 spin_unlock(&rdd->spin);
1147 lwkt_send_ipiq(globaldata_find(rcpu),
1148 dfly_need_user_resched_remote,
1152 spin_unlock(&rdd->spin);
1155 spin_unlock(&rdd->spin);
1162 dfly_yield(struct lwp *lp)
1164 if (lp->lwp_qcpu != mycpu->gd_cpuid)
1166 KKASSERT(lp == curthread->td_lwp);
1169 * Don't set need_user_resched() or mess with rrcount or anything.
1170 * the TDF flag will override everything as long as we release.
1172 atomic_set_int(&lp->lwp_thread->td_mpflags, TDF_MP_DIDYIELD);
1173 dfly_release_curproc(lp);
1177 * Thread was forcefully migrated to another cpu. Normally forced migrations
1178 * are used for iterations and the kernel returns to the original cpu before
1179 * returning and this is not needed. However, if the kernel migrates a
1180 * thread to another cpu and wants to leave it there, it has to call this
1183 * Note that the lwkt_migratecpu() function also released the thread, so
1184 * we don't have to worry about that.
1188 dfly_changedcpu(struct lwp *lp)
1190 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1191 dfly_pcpu_t rdd = &dfly_pcpu[mycpu->gd_cpuid];
1194 spin_lock(&dd->spin);
1195 dfly_changeqcpu_locked(lp, dd, rdd);
1196 spin_unlock(&dd->spin);
1201 * Called from fork1() when a new child process is being created.
1203 * Give the child process an initial estcpu that is more batch then
1204 * its parent and dock the parent for the fork (but do not
1205 * reschedule the parent).
1209 * XXX lwp should be "spawning" instead of "forking"
1212 dfly_forking(struct lwp *plp, struct lwp *lp)
1215 * Put the child 4 queue slots (out of 32) higher than the parent
1216 * (less desireable than the parent).
1218 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1220 lp->lwp_estfast = 0;
1223 * Dock the parent a cost for the fork, protecting us from fork
1224 * bombs. If the parent is forking quickly make the child more
1227 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1231 * Called when a lwp is being removed from this scheduler, typically
1232 * during lwp_exit(). We have to clean out any ULOAD accounting before
1233 * we can let the lp go. The dd->spin lock is not needed for uload
1236 * Scheduler dequeueing has already occurred, no further action in that
1240 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1242 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1244 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1245 atomic_clear_int(&lp->lwp_mpflags, 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);
1253 * This function cannot block in any way, but spinlocks are ok.
1255 * Update the uload based on the state of the thread (whether it is going
1256 * to sleep or running again). The uload is meant to be a longer-term
1257 * load and not an instantanious load.
1260 dfly_uload_update(struct lwp *lp)
1262 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1264 if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1265 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1266 spin_lock(&dd->spin);
1267 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1268 atomic_set_int(&lp->lwp_mpflags,
1270 atomic_add_int(&dd->uload, lp->lwp_uload);
1271 atomic_add_int(&dd->ucount, 1);
1272 atomic_add_int(&dfly_ucount, 1);
1274 spin_unlock(&dd->spin);
1276 } else if (lp->lwp_slptime > 0) {
1277 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1278 spin_lock(&dd->spin);
1279 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1280 atomic_clear_int(&lp->lwp_mpflags,
1282 atomic_add_int(&dd->uload, -lp->lwp_uload);
1283 atomic_add_int(&dd->ucount, -1);
1284 atomic_add_int(&dfly_ucount, -1);
1286 spin_unlock(&dd->spin);
1292 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1293 * it selects a user process and returns it. If chklp is non-NULL and chklp
1294 * has a better or equal priority then the process that would otherwise be
1295 * chosen, NULL is returned.
1297 * Until we fix the RUNQ code the chklp test has to be strict or we may
1298 * bounce between processes trying to acquire the current process designation.
1300 * Must be called with rdd->spin locked. The spinlock is left intact through
1301 * the entire routine. dd->spin does not have to be locked.
1303 * If worst is non-zero this function finds the worst thread instead of the
1304 * best thread (used by the schedulerclock-based rover).
1308 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1309 struct lwp *chklp, int worst)
1319 rtqbits = rdd->rtqueuebits;
1320 tsqbits = rdd->queuebits;
1321 idqbits = rdd->idqueuebits;
1325 pri = bsrl(idqbits);
1326 q = &rdd->idqueues[pri];
1327 which = &rdd->idqueuebits;
1328 } else if (tsqbits) {
1329 pri = bsrl(tsqbits);
1330 q = &rdd->queues[pri];
1331 which = &rdd->queuebits;
1332 } else if (rtqbits) {
1333 pri = bsrl(rtqbits);
1334 q = &rdd->rtqueues[pri];
1335 which = &rdd->rtqueuebits;
1339 lp = TAILQ_LAST(q, rq);
1342 pri = bsfl(rtqbits);
1343 q = &rdd->rtqueues[pri];
1344 which = &rdd->rtqueuebits;
1345 } else if (tsqbits) {
1346 pri = bsfl(tsqbits);
1347 q = &rdd->queues[pri];
1348 which = &rdd->queuebits;
1349 } else if (idqbits) {
1350 pri = bsfl(idqbits);
1351 q = &rdd->idqueues[pri];
1352 which = &rdd->idqueuebits;
1356 lp = TAILQ_FIRST(q);
1358 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1361 * If the passed lwp <chklp> is reasonably close to the selected
1362 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1364 * Note that we must error on the side of <chklp> to avoid bouncing
1365 * between threads in the acquire code.
1368 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1372 KTR_COND_LOG(usched_chooseproc,
1373 lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1374 lp->lwp_proc->p_pid,
1375 lp->lwp_thread->td_gd->gd_cpuid,
1378 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1379 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1380 TAILQ_REMOVE(q, lp, lwp_procq);
1383 *which &= ~(1 << pri);
1386 * If we are choosing a process from rdd with the intent to
1387 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1391 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1392 atomic_add_int(&rdd->uload, -lp->lwp_uload);
1393 atomic_add_int(&rdd->ucount, -1);
1394 atomic_add_int(&dfly_ucount, -1);
1396 lp->lwp_qcpu = dd->cpuid;
1397 atomic_add_int(&dd->uload, lp->lwp_uload);
1398 atomic_add_int(&dd->ucount, 1);
1399 atomic_add_int(&dfly_ucount, 1);
1400 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1406 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1408 * Choose a cpu node to schedule lp on, hopefully nearby its current
1411 * We give the current node a modest advantage for obvious reasons.
1413 * We also give the node the thread was woken up FROM a slight advantage
1414 * in order to try to schedule paired threads which synchronize/block waiting
1415 * for each other fairly close to each other. Similarly in a network setting
1416 * this feature will also attempt to place a user process near the kernel
1417 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the
1418 * algorithm as it heuristically groups synchronizing processes for locality
1419 * of reference in multi-socket systems.
1421 * We check against running processes and give a big advantage if there
1424 * The caller will normally dfly_setrunqueue() lp on the returned queue.
1426 * When the topology is known choose a cpu whos group has, in aggregate,
1427 * has the lowest weighted load.
1431 dfly_choose_best_queue(struct lwp *lp)
1438 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1448 * When the topology is unknown choose a random cpu that is hopefully
1451 if (dd->cpunode == NULL)
1452 return (dfly_choose_queue_simple(dd, lp));
1457 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1458 wakemask = dfly_pcpu[wakecpu].cpumask;
1463 * When the topology is known choose a cpu whos group has, in
1464 * aggregate, has the lowest weighted load.
1466 cpup = root_cpu_node;
1471 * Degenerate case super-root
1473 if (cpup->child_node && cpup->child_no == 1) {
1474 cpup = cpup->child_node;
1481 if (cpup->child_node == NULL) {
1482 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1487 lowest_load = 0x7FFFFFFF;
1489 for (n = 0; n < cpup->child_no; ++n) {
1491 * Accumulate load information for all cpus
1492 * which are members of this node.
1494 cpun = &cpup->child_node[n];
1495 mask = cpun->members & usched_global_cpumask &
1496 smp_active_mask & lp->lwp_cpumask;
1504 cpuid = BSFCPUMASK(mask);
1505 rdd = &dfly_pcpu[cpuid];
1507 load += rdd->ucount * usched_dfly_weight3;
1509 if (rdd->uschedcp == NULL &&
1510 rdd->runqcount == 0 &&
1511 globaldata_find(cpuid)->gd_tdrunqcount == 0
1513 load -= usched_dfly_weight4;
1516 else if (rdd->upri > lp->lwp_priority + PPQ) {
1517 load -= usched_dfly_weight4 / 2;
1520 mask &= ~CPUMASK(cpuid);
1525 * Compensate if the lp is already accounted for in
1526 * the aggregate uload for this mask set. We want
1527 * to calculate the loads as if lp were not present,
1528 * otherwise the calculation is bogus.
1530 if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1531 (dd->cpumask & cpun->members)) {
1532 load -= lp->lwp_uload;
1533 load -= usched_dfly_weight3;
1539 * Advantage the cpu group (lp) is already on.
1541 if (cpun->members & dd->cpumask)
1542 load -= usched_dfly_weight1;
1545 * Advantage the cpu group we want to pair (lp) to,
1546 * but don't let it go to the exact same cpu as
1547 * the wakecpu target.
1549 * We do this by checking whether cpun is a
1550 * terminal node or not. All cpun's at the same
1551 * level will either all be terminal or all not
1554 * If it is and we match we disadvantage the load.
1555 * If it is and we don't match we advantage the load.
1557 * Also note that we are effectively disadvantaging
1558 * all-but-one by the same amount, so it won't effect
1559 * the weight1 factor for the all-but-one nodes.
1561 if (cpun->members & wakemask) {
1562 if (cpun->child_node != NULL) {
1564 load -= usched_dfly_weight2;
1566 if (usched_dfly_features & 0x10)
1567 load += usched_dfly_weight2;
1569 load -= usched_dfly_weight2;
1574 * Calculate the best load
1576 if (cpub == NULL || lowest_load > load ||
1577 (lowest_load == load &&
1578 (cpun->members & dd->cpumask))
1586 if (usched_dfly_chooser)
1587 kprintf("lp %02d->%02d %s\n",
1588 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1593 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1595 * Choose the worst queue close to dd's cpu node with a non-empty runq
1596 * that is NOT dd. Also require that the moving of the highest-load thread
1597 * from rdd to dd does not cause the uload's to cross each other.
1599 * This is used by the thread chooser when the current cpu's queues are
1600 * empty to steal a thread from another cpu's queue. We want to offload
1601 * the most heavily-loaded queue.
1605 dfly_choose_worst_queue(dfly_pcpu_t dd)
1623 * When the topology is unknown choose a random cpu that is hopefully
1626 if (dd->cpunode == NULL) {
1631 * When the topology is known choose a cpu whos group has, in
1632 * aggregate, has the lowest weighted load.
1634 cpup = root_cpu_node;
1638 * Degenerate case super-root
1640 if (cpup->child_node && cpup->child_no == 1) {
1641 cpup = cpup->child_node;
1648 if (cpup->child_node == NULL) {
1649 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1656 for (n = 0; n < cpup->child_no; ++n) {
1658 * Accumulate load information for all cpus
1659 * which are members of this node.
1661 cpun = &cpup->child_node[n];
1662 mask = cpun->members & usched_global_cpumask &
1670 cpuid = BSFCPUMASK(mask);
1671 rdd = &dfly_pcpu[cpuid];
1673 load += rdd->ucount * usched_dfly_weight3;
1674 if (rdd->uschedcp == NULL &&
1675 rdd->runqcount == 0 &&
1676 globaldata_find(cpuid)->gd_tdrunqcount == 0
1678 load -= usched_dfly_weight4;
1681 else if (rdd->upri > dd->upri + PPQ) {
1682 load -= usched_dfly_weight4 / 2;
1685 mask &= ~CPUMASK(cpuid);
1691 * Prefer candidates which are somewhat closer to
1694 if (dd->cpumask & cpun->members)
1695 load += usched_dfly_weight1;
1698 * The best candidate is the one with the worst
1701 if (cpub == NULL || highest_load < load) {
1702 highest_load = load;
1710 * We never return our own node (dd), and only return a remote
1711 * node if it's load is significantly worse than ours (i.e. where
1712 * stealing a thread would be considered reasonable).
1714 * This also helps us avoid breaking paired threads apart which
1715 * can have disastrous effects on performance.
1722 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1724 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1726 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1729 if (rdd->uload - hpri < dd->uload + hpri)
1737 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1745 * Fallback to the original heuristic, select random cpu,
1746 * first checking cpus not currently running a user thread.
1749 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1750 mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1751 smp_active_mask & usched_global_cpumask;
1754 tmpmask = ~(CPUMASK(cpuid) - 1);
1756 cpuid = BSFCPUMASK(mask & tmpmask);
1758 cpuid = BSFCPUMASK(mask);
1759 rdd = &dfly_pcpu[cpuid];
1761 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1763 mask &= ~CPUMASK(cpuid);
1767 * Then cpus which might have a currently running lp
1769 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1770 mask = dfly_curprocmask & dfly_rdyprocmask &
1771 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1774 tmpmask = ~(CPUMASK(cpuid) - 1);
1776 cpuid = BSFCPUMASK(mask & tmpmask);
1778 cpuid = BSFCPUMASK(mask);
1779 rdd = &dfly_pcpu[cpuid];
1781 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1783 mask &= ~CPUMASK(cpuid);
1787 * If we cannot find a suitable cpu we reload from dfly_scancpu
1788 * and round-robin. Other cpus will pickup as they release their
1789 * current lwps or become ready.
1791 * Avoid a degenerate system lockup case if usched_global_cpumask
1792 * is set to 0 or otherwise does not cover lwp_cpumask.
1794 * We only kick the target helper thread in this case, we do not
1795 * set the user resched flag because
1797 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1798 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1800 rdd = &dfly_pcpu[cpuid];
1807 dfly_need_user_resched_remote(void *dummy)
1809 globaldata_t gd = mycpu;
1810 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
1813 * Flag reschedule needed
1815 need_user_resched();
1818 * If no user thread is currently running we need to kick the helper
1819 * on our cpu to recover. Otherwise the cpu will never schedule
1822 * We cannot schedule the process ourselves because this is an
1823 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1825 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1827 if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1828 atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1829 wakeup_mycpu(&dd->helper_thread);
1834 * dfly_remrunqueue_locked() removes a given process from the run queue
1835 * that it is on, clearing the queue busy bit if it becomes empty.
1837 * Note that user process scheduler is different from the LWKT schedule.
1838 * The user process scheduler only manages user processes but it uses LWKT
1839 * underneath, and a user process operating in the kernel will often be
1840 * 'released' from our management.
1842 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes
1843 * to sleep or the lwp is moved to a different runq.
1846 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1852 KKASSERT(rdd->runqcount >= 0);
1854 pri = lp->lwp_rqindex;
1856 switch(lp->lwp_rqtype) {
1857 case RTP_PRIO_NORMAL:
1858 q = &rdd->queues[pri];
1859 which = &rdd->queuebits;
1861 case RTP_PRIO_REALTIME:
1863 q = &rdd->rtqueues[pri];
1864 which = &rdd->rtqueuebits;
1867 q = &rdd->idqueues[pri];
1868 which = &rdd->idqueuebits;
1871 panic("remrunqueue: invalid rtprio type");
1874 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1875 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1876 TAILQ_REMOVE(q, lp, lwp_procq);
1878 if (TAILQ_EMPTY(q)) {
1879 KASSERT((*which & (1 << pri)) != 0,
1880 ("remrunqueue: remove from empty queue"));
1881 *which &= ~(1 << pri);
1886 * dfly_setrunqueue_locked()
1888 * Add a process whos rqtype and rqindex had previously been calculated
1889 * onto the appropriate run queue. Determine if the addition requires
1890 * a reschedule on a cpu and return the cpuid or -1.
1892 * NOTE: Lower priorities are better priorities.
1894 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1895 * sum of the rough lwp_priority for all running and runnable
1896 * processes. Lower priority processes (higher lwp_priority
1897 * values) actually DO count as more load, not less, because
1898 * these are the programs which require the most care with
1899 * regards to cpu selection.
1902 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1908 KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1910 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1911 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1912 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1913 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1914 atomic_add_int(&dfly_ucount, 1);
1917 pri = lp->lwp_rqindex;
1919 switch(lp->lwp_rqtype) {
1920 case RTP_PRIO_NORMAL:
1921 q = &rdd->queues[pri];
1922 which = &rdd->queuebits;
1924 case RTP_PRIO_REALTIME:
1926 q = &rdd->rtqueues[pri];
1927 which = &rdd->rtqueuebits;
1930 q = &rdd->idqueues[pri];
1931 which = &rdd->idqueuebits;
1934 panic("remrunqueue: invalid rtprio type");
1939 * Place us on the selected queue. Determine if we should be
1940 * placed at the head of the queue or at the end.
1942 * We are placed at the tail if our round-robin count has expired,
1943 * or is about to expire and the system thinks its a good place to
1944 * round-robin, or there is already a next thread on the queue
1945 * (it might be trying to pick up where it left off and we don't
1946 * want to interfere).
1948 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1949 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1952 if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1953 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1954 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1957 atomic_clear_int(&lp->lwp_thread->td_mpflags,
1958 TDF_MP_BATCH_DEMARC);
1959 lp->lwp_rrcount = 0;
1960 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1963 lp->lwp_rrcount = 0;
1964 TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1970 * For SMP systems a user scheduler helper thread is created for each
1971 * cpu and is used to allow one cpu to wakeup another for the purposes of
1972 * scheduling userland threads from setrunqueue().
1974 * UP systems do not need the helper since there is only one cpu.
1976 * We can't use the idle thread for this because we might block.
1977 * Additionally, doing things this way allows us to HLT idle cpus
1981 dfly_helper_thread(void *dummy)
1991 cpuid = gd->gd_cpuid; /* doesn't change */
1992 mask = gd->gd_cpumask; /* doesn't change */
1993 dd = &dfly_pcpu[cpuid];
1996 * Since we only want to be woken up only when no user processes
1997 * are scheduled on a cpu, run at an ultra low priority.
1999 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
2001 tsleep(&dd->helper_thread, 0, "schslp", 0);
2005 * We use the LWKT deschedule-interlock trick to avoid racing
2006 * dfly_rdyprocmask. This means we cannot block through to the
2007 * manual lwkt_switch() call we make below.
2010 tsleep_interlock(&dd->helper_thread, 0);
2012 spin_lock(&dd->spin);
2014 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2015 clear_user_resched(); /* This satisfied the reschedule request */
2017 dd->rrcount = 0; /* Reset the round-robin counter */
2020 if (dd->runqcount || dd->uschedcp != NULL) {
2022 * Threads are available. A thread may or may not be
2023 * currently scheduled. Get the best thread already queued
2026 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
2028 atomic_set_cpumask(&dfly_curprocmask, mask);
2029 dd->upri = nlp->lwp_priority;
2032 dd->rrcount = 0; /* reset round robin */
2034 spin_unlock(&dd->spin);
2035 lwkt_acquire(nlp->lwp_thread);
2036 lwkt_schedule(nlp->lwp_thread);
2039 * This situation should not occur because we had
2040 * at least one thread available.
2042 spin_unlock(&dd->spin);
2044 } else if (usched_dfly_features & 0x01) {
2046 * This cpu is devoid of runnable threads, steal a thread
2047 * from another cpu. Since we're stealing, might as well
2048 * load balance at the same time.
2050 * We choose the highest-loaded thread from the worst queue.
2052 * NOTE! This function only returns a non-NULL rdd when
2053 * another cpu's queue is obviously overloaded. We
2054 * do not want to perform the type of rebalancing
2055 * the schedclock does here because it would result
2056 * in insane process pulling when 'steady' state is
2057 * partially unbalanced (e.g. 6 runnables and only
2060 rdd = dfly_choose_worst_queue(dd);
2061 if (rdd && spin_trylock(&rdd->spin)) {
2062 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2063 spin_unlock(&rdd->spin);
2068 atomic_set_cpumask(&dfly_curprocmask, mask);
2069 dd->upri = nlp->lwp_priority;
2072 dd->rrcount = 0; /* reset round robin */
2074 spin_unlock(&dd->spin);
2075 lwkt_acquire(nlp->lwp_thread);
2076 lwkt_schedule(nlp->lwp_thread);
2079 * Leave the thread on our run queue. Another
2080 * scheduler will try to pull it later.
2082 spin_unlock(&dd->spin);
2086 * devoid of runnable threads and not allowed to steal
2089 spin_unlock(&dd->spin);
2093 * We're descheduled unless someone scheduled us. Switch away.
2094 * Exiting the critical section will cause splz() to be called
2095 * for us if interrupts and such are pending.
2098 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2104 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2108 new_val = usched_dfly_stick_to_level;
2110 error = sysctl_handle_int(oidp, &new_val, 0, req);
2111 if (error != 0 || req->newptr == NULL)
2113 if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2115 usched_dfly_stick_to_level = new_val;
2121 * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2122 * Note that curprocmask bit 0 has already been cleared by rqinit() and
2123 * we should not mess with it further.
2126 usched_dfly_cpu_init(void)
2131 int smt_not_supported = 0;
2132 int cache_coherent_not_supported = 0;
2135 kprintf("Start scheduler helpers on cpus:\n");
2137 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2138 usched_dfly_sysctl_tree =
2139 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2140 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2141 "usched_dfly", CTLFLAG_RD, 0, "");
2143 for (i = 0; i < ncpus; ++i) {
2144 dfly_pcpu_t dd = &dfly_pcpu[i];
2145 cpumask_t mask = CPUMASK(i);
2147 if ((mask & smp_active_mask) == 0)
2150 spin_init(&dd->spin);
2151 dd->cpunode = get_cpu_node_by_cpuid(i);
2153 dd->cpumask = CPUMASK(i);
2154 for (j = 0; j < NQS; j++) {
2155 TAILQ_INIT(&dd->queues[j]);
2156 TAILQ_INIT(&dd->rtqueues[j]);
2157 TAILQ_INIT(&dd->idqueues[j]);
2159 atomic_clear_cpumask(&dfly_curprocmask, 1);
2161 if (dd->cpunode == NULL) {
2162 smt_not_supported = 1;
2163 cache_coherent_not_supported = 1;
2165 kprintf ("\tcpu%d - WARNING: No CPU NODE "
2166 "found for cpu\n", i);
2168 switch (dd->cpunode->type) {
2171 kprintf ("\tcpu%d - HyperThreading "
2172 "available. Core siblings: ",
2176 smt_not_supported = 1;
2179 kprintf ("\tcpu%d - No HT available, "
2180 "multi-core/physical "
2181 "cpu. Physical siblings: ",
2185 smt_not_supported = 1;
2188 kprintf ("\tcpu%d - No HT available, "
2189 "single-core/physical cpu. "
2190 "Package Siblings: ",
2194 /* Let's go for safe defaults here */
2195 smt_not_supported = 1;
2196 cache_coherent_not_supported = 1;
2198 kprintf ("\tcpu%d - Unknown cpunode->"
2199 "type=%u. Siblings: ",
2201 (u_int)dd->cpunode->type);
2206 if (dd->cpunode->parent_node != NULL) {
2207 CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2208 kprintf("cpu%d ", cpuid);
2211 kprintf(" no siblings\n");
2216 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2217 0, i, "usched %d", i);
2220 * Allow user scheduling on the target cpu. cpu #0 has already
2221 * been enabled in rqinit().
2224 atomic_clear_cpumask(&dfly_curprocmask, mask);
2225 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2226 dd->upri = PRIBASE_NULL;
2230 /* usched_dfly sysctl configurable parameters */
2232 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2233 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2234 OID_AUTO, "rrinterval", CTLFLAG_RW,
2235 &usched_dfly_rrinterval, 0, "");
2236 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2237 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2238 OID_AUTO, "decay", CTLFLAG_RW,
2239 &usched_dfly_decay, 0, "Extra decay when not running");
2241 /* Add enable/disable option for SMT scheduling if supported */
2242 if (smt_not_supported) {
2243 usched_dfly_smt = 0;
2244 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2245 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2246 OID_AUTO, "smt", CTLFLAG_RD,
2247 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2249 usched_dfly_smt = 1;
2250 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2251 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2252 OID_AUTO, "smt", CTLFLAG_RW,
2253 &usched_dfly_smt, 0, "Enable SMT scheduling");
2257 * Add enable/disable option for cache coherent scheduling
2260 if (cache_coherent_not_supported) {
2261 usched_dfly_cache_coherent = 0;
2262 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2263 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2264 OID_AUTO, "cache_coherent", CTLFLAG_RD,
2266 "Cache coherence NOT SUPPORTED");
2268 usched_dfly_cache_coherent = 1;
2269 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2270 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2271 OID_AUTO, "cache_coherent", CTLFLAG_RW,
2272 &usched_dfly_cache_coherent, 0,
2273 "Enable/Disable cache coherent scheduling");
2275 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2276 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2277 OID_AUTO, "weight1", CTLFLAG_RW,
2278 &usched_dfly_weight1, 200,
2279 "Weight selection for current cpu");
2281 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2282 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2283 OID_AUTO, "weight2", CTLFLAG_RW,
2284 &usched_dfly_weight2, 180,
2285 "Weight selection for wakefrom cpu");
2287 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2288 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2289 OID_AUTO, "weight3", CTLFLAG_RW,
2290 &usched_dfly_weight3, 40,
2291 "Weight selection for num threads on queue");
2293 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2294 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2295 OID_AUTO, "weight4", CTLFLAG_RW,
2296 &usched_dfly_weight4, 160,
2297 "Availability of other idle cpus");
2299 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2300 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2301 OID_AUTO, "fast_resched", CTLFLAG_RW,
2302 &usched_dfly_fast_resched, 0,
2303 "Availability of other idle cpus");
2305 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2306 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2307 OID_AUTO, "features", CTLFLAG_RW,
2308 &usched_dfly_features, 0x8F,
2309 "Allow pulls into empty queues");
2311 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2312 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2313 OID_AUTO, "swmask", CTLFLAG_RW,
2314 &usched_dfly_swmask, ~PPQMASK,
2315 "Queue mask to force thread switch");
2318 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2319 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2320 OID_AUTO, "stick_to_level",
2321 CTLTYPE_INT | CTLFLAG_RW,
2322 NULL, sizeof usched_dfly_stick_to_level,
2323 sysctl_usched_dfly_stick_to_level, "I",
2324 "Stick a process to this level. See sysctl"
2325 "paremter hw.cpu_topology.level_description");
2329 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2330 usched_dfly_cpu_init, NULL)