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;
105 struct thread helper_thread;
107 struct thread helper_thread_UNUSED; /* field unused */
113 struct lwp *uschedcp;
114 struct rq queues[NQS];
115 struct rq rtqueues[NQS];
116 struct rq idqueues[NQS];
118 u_int32_t rtqueuebits;
119 u_int32_t idqueuebits;
128 typedef struct usched_dfly_pcpu *dfly_pcpu_t;
130 static void dfly_acquire_curproc(struct lwp *lp);
131 static void dfly_release_curproc(struct lwp *lp);
132 static void dfly_select_curproc(globaldata_t gd);
133 static void dfly_setrunqueue(struct lwp *lp);
134 static void dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp);
135 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period,
137 static void dfly_recalculate_estcpu(struct lwp *lp);
138 static void dfly_resetpriority(struct lwp *lp);
139 static void dfly_forking(struct lwp *plp, struct lwp *lp);
140 static void dfly_exiting(struct lwp *lp, struct proc *);
141 static void dfly_uload_update(struct lwp *lp);
142 static void dfly_yield(struct lwp *lp);
144 static void dfly_changeqcpu_locked(struct lwp *lp,
145 dfly_pcpu_t dd, dfly_pcpu_t rdd);
146 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
147 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
148 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
152 static void dfly_need_user_resched_remote(void *dummy);
154 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
155 struct lwp *chklp, int worst);
156 static void dfly_remrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
157 static void dfly_setrunqueue_locked(dfly_pcpu_t dd, struct lwp *lp);
159 struct usched usched_dfly = {
161 "dfly", "Original DragonFly Scheduler",
162 NULL, /* default registration */
163 NULL, /* default deregistration */
164 dfly_acquire_curproc,
165 dfly_release_curproc,
168 dfly_recalculate_estcpu,
173 NULL, /* setcpumask not supported */
178 * We have NQS (32) run queues per scheduling class. For the normal
179 * class, there are 128 priorities scaled onto these 32 queues. New
180 * processes are added to the last entry in each queue, and processes
181 * are selected for running by taking them from the head and maintaining
182 * a simple FIFO arrangement. Realtime and Idle priority processes have
183 * and explicit 0-31 priority which maps directly onto their class queue
184 * index. When a queue has something in it, the corresponding bit is
185 * set in the queuebits variable, allowing a single read to determine
186 * the state of all 32 queues and then a ffs() to find the first busy
189 static cpumask_t dfly_curprocmask = -1; /* currently running a user process */
190 static cpumask_t dfly_rdyprocmask; /* ready to accept a user process */
192 static volatile int dfly_scancpu;
194 static volatile int dfly_ucount; /* total running on whole system */
195 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
196 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
197 static struct sysctl_oid *usched_dfly_sysctl_tree;
199 /* Debug info exposed through debug.* sysctl */
201 static int usched_dfly_debug = -1;
202 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
203 &usched_dfly_debug, 0,
204 "Print debug information for this pid");
206 static int usched_dfly_pid_debug = -1;
207 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
208 &usched_dfly_pid_debug, 0,
209 "Print KTR debug information for this pid");
211 static int usched_dfly_chooser = 0;
212 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
213 &usched_dfly_chooser, 0,
214 "Print KTR debug information for this pid");
217 * Tunning usched_dfly - configurable through kern.usched_dfly.
219 * weight1 - Tries to keep threads on their current cpu. If you
220 * make this value too large the scheduler will not be
221 * able to load-balance large loads.
223 * weight2 - If non-zero, detects thread pairs undergoing synchronous
224 * communications and tries to move them closer together.
225 * Behavior is adjusted by bit 4 of features (0x10).
227 * WARNING! Weight2 is a ridiculously sensitive parameter,
228 * a small value is recommended.
230 * weight3 - Weighting based on the number of recently runnable threads
231 * on the userland scheduling queue (ignoring their loads).
232 * A nominal value here prevents high-priority (low-load)
233 * threads from accumulating on one cpu core when other
234 * cores are available.
236 * This value should be left fairly small relative to weight1
239 * weight4 - Weighting based on other cpu queues being available
240 * or running processes with higher lwp_priority's.
242 * This allows a thread to migrate to another nearby cpu if it
243 * is unable to run on the current cpu based on the other cpu
244 * being idle or running a lower priority (higher lwp_priority)
245 * thread. This value should be large enough to override weight1
247 * features - These flags can be set or cleared to enable or disable various
250 * 0x01 Enable idle-cpu pulling (default)
251 * 0x02 Enable proactive pushing (default)
252 * 0x04 Enable rebalancing rover (default)
253 * 0x08 Enable more proactive pushing (default)
254 * 0x10 (flip weight2 limit on same cpu) (default)
255 * 0x20 choose best cpu for forked process
256 * 0x40 choose current cpu for forked process
257 * 0x80 choose random cpu for forked process (default)
260 static int usched_dfly_smt = 0;
261 static int usched_dfly_cache_coherent = 0;
262 static int usched_dfly_weight1 = 200; /* keep thread on current cpu */
263 static int usched_dfly_weight2 = 180; /* synchronous peer's current cpu */
264 static int usched_dfly_weight3 = 40; /* number of threads on queue */
265 static int usched_dfly_weight4 = 160; /* availability of idle cores */
266 static int usched_dfly_features = 0x8F; /* allow pulls */
268 static int usched_dfly_fast_resched = 0;/* delta priority / resched */
269 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */
270 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
271 static int usched_dfly_decay = 8;
273 /* KTR debug printings */
275 KTR_INFO_MASTER(usched);
277 #if !defined(KTR_USCHED_DFLY)
278 #define KTR_USCHED_DFLY KTR_ALL
281 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
282 "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
283 pid_t pid, int old_cpuid, int curr);
286 * This function is called when the kernel intends to return to userland.
287 * It is responsible for making the thread the current designated userland
288 * thread for this cpu, blocking if necessary.
290 * The kernel will not depress our LWKT priority until after we return,
291 * in case we have to shove over to another cpu.
293 * We must determine our thread's disposition before we switch away. This
294 * is very sensitive code.
296 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
297 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will
298 * occur, this function is called only under very controlled circumstances.
301 dfly_acquire_curproc(struct lwp *lp)
312 * Make sure we aren't sitting on a tsleep queue.
315 crit_enter_quick(td);
316 if (td->td_flags & TDF_TSLEEPQ)
318 dfly_recalculate_estcpu(lp);
321 dd = &dfly_pcpu[gd->gd_cpuid];
324 * Process any pending interrupts/ipi's, then handle reschedule
325 * requests. dfly_release_curproc() will try to assign a new
326 * uschedcp that isn't us and otherwise NULL it out.
329 if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
330 lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
334 if (user_resched_wanted()) {
335 if (dd->uschedcp == lp)
337 clear_user_resched();
338 dfly_release_curproc(lp);
342 * Loop until we are the current user thread.
344 * NOTE: dd spinlock not held at top of loop.
346 if (dd->uschedcp == lp)
349 while (dd->uschedcp != lp) {
352 spin_lock(&dd->spin);
355 * We are not or are no longer the current lwp and a forced
356 * reschedule was requested. Figure out the best cpu to
357 * run on (our current cpu will be given significant weight).
359 * (if a reschedule was not requested we want to move this
360 * step after the uschedcp tests).
364 (usched_dfly_features & 0x08) &&
365 (rdd = dfly_choose_best_queue(lp)) != dd) {
366 dfly_changeqcpu_locked(lp, dd, rdd);
367 spin_unlock(&dd->spin);
368 lwkt_deschedule(lp->lwp_thread);
369 dfly_setrunqueue_dd(rdd, lp);
372 dd = &dfly_pcpu[gd->gd_cpuid];
378 * Either no reschedule was requested or the best queue was
379 * dd, and no current process has been selected. We can
380 * trivially become the current lwp on the current cpu.
382 if (dd->uschedcp == NULL) {
383 atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
385 dd->upri = lp->lwp_priority;
386 KKASSERT(lp->lwp_qcpu == dd->cpuid);
387 spin_unlock(&dd->spin);
392 * Can we steal the current designated user thread?
394 * If we do the other thread will stall when it tries to
395 * return to userland, possibly rescheduling elsewhere.
397 * It is important to do a masked test to avoid the edge
398 * case where two near-equal-priority threads are constantly
399 * interrupting each other.
401 * In the exact match case another thread has already gained
402 * uschedcp and lowered its priority, if we steal it the
403 * other thread will stay stuck on the LWKT runq and not
404 * push to another cpu. So don't steal on equal-priority even
405 * though it might appear to be more beneficial due to not
406 * having to switch back to the other thread's context.
408 * usched_dfly_fast_resched requires that two threads be
409 * significantly far apart in priority in order to interrupt.
411 * If better but not sufficiently far apart, the current
412 * uschedcp will be interrupted at the next scheduler clock.
415 (dd->upri & ~PPQMASK) >
416 (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
418 dd->upri = lp->lwp_priority;
419 KKASSERT(lp->lwp_qcpu == dd->cpuid);
420 spin_unlock(&dd->spin);
425 * We are not the current lwp, figure out the best cpu
426 * to run on (our current cpu will be given significant
427 * weight). Loop on cpu change.
429 if ((usched_dfly_features & 0x02) &&
430 force_resched == 0 &&
431 (rdd = dfly_choose_best_queue(lp)) != dd) {
432 dfly_changeqcpu_locked(lp, dd, rdd);
433 spin_unlock(&dd->spin);
434 lwkt_deschedule(lp->lwp_thread);
435 dfly_setrunqueue_dd(rdd, lp);
438 dd = &dfly_pcpu[gd->gd_cpuid];
444 * We cannot become the current lwp, place the lp on the
445 * run-queue of this or another cpu and deschedule ourselves.
447 * When we are reactivated we will have another chance.
449 * Reload after a switch or setrunqueue/switch possibly
450 * moved us to another cpu.
452 spin_unlock(&dd->spin);
453 lwkt_deschedule(lp->lwp_thread);
454 dfly_setrunqueue_dd(dd, lp);
457 dd = &dfly_pcpu[gd->gd_cpuid];
461 * Make sure upri is synchronized, then yield to LWKT threads as
462 * needed before returning. This could result in another reschedule.
467 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
471 * DFLY_RELEASE_CURPROC
473 * This routine detaches the current thread from the userland scheduler,
474 * usually because the thread needs to run or block in the kernel (at
475 * kernel priority) for a while.
477 * This routine is also responsible for selecting a new thread to
478 * make the current thread.
480 * NOTE: This implementation differs from the dummy example in that
481 * dfly_select_curproc() is able to select the current process, whereas
482 * dummy_select_curproc() is not able to select the current process.
483 * This means we have to NULL out uschedcp.
485 * Additionally, note that we may already be on a run queue if releasing
486 * via the lwkt_switch() in dfly_setrunqueue().
489 dfly_release_curproc(struct lwp *lp)
491 globaldata_t gd = mycpu;
492 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
495 * Make sure td_wakefromcpu is defaulted. This will be overwritten
498 if (dd->uschedcp == lp) {
499 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
500 spin_lock(&dd->spin);
501 if (dd->uschedcp == lp) {
502 dd->uschedcp = NULL; /* don't let lp be selected */
503 dd->upri = PRIBASE_NULL;
504 atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
505 spin_unlock(&dd->spin);
506 dfly_select_curproc(gd);
508 spin_unlock(&dd->spin);
514 * DFLY_SELECT_CURPROC
516 * Select a new current process for this cpu and clear any pending user
517 * reschedule request. The cpu currently has no current process.
519 * This routine is also responsible for equal-priority round-robining,
520 * typically triggered from dfly_schedulerclock(). In our dummy example
521 * all the 'user' threads are LWKT scheduled all at once and we just
522 * call lwkt_switch().
524 * The calling process is not on the queue and cannot be selected.
528 dfly_select_curproc(globaldata_t gd)
530 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
532 int cpuid = gd->gd_cpuid;
536 spin_lock(&dd->spin);
537 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
540 atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
541 dd->upri = nlp->lwp_priority;
544 dd->rrcount = 0; /* reset round robin */
546 spin_unlock(&dd->spin);
548 lwkt_acquire(nlp->lwp_thread);
550 lwkt_schedule(nlp->lwp_thread);
552 spin_unlock(&dd->spin);
558 * Place the specified lwp on the user scheduler's run queue. This routine
559 * must be called with the thread descheduled. The lwp must be runnable.
560 * It must not be possible for anyone else to explicitly schedule this thread.
562 * The thread may be the current thread as a special case.
565 dfly_setrunqueue(struct lwp *lp)
571 * First validate the process LWKT state.
573 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
574 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
575 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
576 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
577 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
580 * NOTE: dd/rdd do not necessarily represent the current cpu.
581 * Instead they may represent the cpu the thread was last
582 * scheduled on or inherited by its parent.
584 dd = &dfly_pcpu[lp->lwp_qcpu];
588 * This process is not supposed to be scheduled anywhere or assigned
589 * as the current process anywhere. Assert the condition.
591 KKASSERT(rdd->uschedcp != lp);
595 * If we are not SMP we do not have a scheduler helper to kick
596 * and must directly activate the process if none are scheduled.
598 * This is really only an issue when bootstrapping init since
599 * the caller in all other cases will be a user process, and
600 * even if released (rdd->uschedcp == NULL), that process will
601 * kickstart the scheduler when it returns to user mode from
604 * NOTE: On SMP we can't just set some other cpu's uschedcp.
606 if (rdd->uschedcp == NULL) {
607 spin_lock(&rdd->spin);
608 if (rdd->uschedcp == NULL) {
609 atomic_set_cpumask(&dfly_curprocmask, 1);
611 rdd->upri = lp->lwp_priority;
612 spin_unlock(&rdd->spin);
613 lwkt_schedule(lp->lwp_thread);
616 spin_unlock(&rdd->spin);
622 * Ok, we have to setrunqueue some target cpu and request a reschedule
625 * We have to choose the best target cpu. It might not be the current
626 * target even if the current cpu has no running user thread (for
627 * example, because the current cpu might be a hyperthread and its
628 * sibling has a thread assigned).
630 * If we just forked it is most optimal to run the child on the same
631 * cpu just in case the parent decides to wait for it (thus getting
632 * off that cpu). As long as there is nothing else runnable on the
633 * cpu, that is. If we did this unconditionally a parent forking
634 * multiple children before waiting (e.g. make -j N) leaves other
635 * cpus idle that could be working.
637 if (lp->lwp_forked) {
639 if (usched_dfly_features & 0x20)
640 rdd = dfly_choose_best_queue(lp);
641 else if (usched_dfly_features & 0x40)
642 rdd = &dfly_pcpu[lp->lwp_qcpu];
643 else if (usched_dfly_features & 0x80)
644 rdd = dfly_choose_queue_simple(rdd, lp);
645 else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
646 rdd = dfly_choose_best_queue(lp);
648 rdd = &dfly_pcpu[lp->lwp_qcpu];
650 rdd = dfly_choose_best_queue(lp);
651 /* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
653 if (lp->lwp_qcpu != rdd->cpuid) {
654 spin_lock(&dd->spin);
655 dfly_changeqcpu_locked(lp, dd, rdd);
656 spin_unlock(&dd->spin);
659 dfly_setrunqueue_dd(rdd, lp);
665 * Change qcpu to rdd->cpuid. The dd the lp is CURRENTLY on must be
666 * spin-locked on-call. rdd does not have to be.
669 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
671 if (lp->lwp_qcpu != rdd->cpuid) {
672 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
673 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
674 atomic_add_int(&dd->uload, -lp->lwp_uload);
675 atomic_add_int(&dd->ucount, -1);
676 atomic_add_int(&dfly_ucount, -1);
678 lp->lwp_qcpu = rdd->cpuid;
685 * Place lp on rdd's runqueue. Nothing is locked on call. This function
686 * also performs all necessary ancillary notification actions.
689 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
695 * We might be moving the lp to another cpu's run queue, and once
696 * on the runqueue (even if it is our cpu's), another cpu can rip
699 * TDF_MIGRATING might already be set if this is part of a
700 * remrunqueue+setrunqueue sequence.
702 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
703 lwkt_giveaway(lp->lwp_thread);
705 rgd = globaldata_find(rdd->cpuid);
708 * We lose control of the lp the moment we release the spinlock
709 * after having placed it on the queue. i.e. another cpu could pick
710 * it up, or it could exit, or its priority could be further
711 * adjusted, or something like that.
713 * WARNING! rdd can point to a foreign cpu!
715 spin_lock(&rdd->spin);
716 dfly_setrunqueue_locked(rdd, lp);
719 * Potentially interrupt the currently-running thread
721 if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
723 * Currently running thread is better or same, do not
726 spin_unlock(&rdd->spin);
727 } else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
728 usched_dfly_fast_resched) {
730 * Currently running thread is not better, but not so bad
731 * that we need to interrupt it. Let it run for one more
735 rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
736 rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
738 spin_unlock(&rdd->spin);
739 } else if (rgd == mycpu) {
741 * We should interrupt the currently running thread, which
742 * is on the current cpu.
744 spin_unlock(&rdd->spin);
745 if (rdd->uschedcp == NULL) {
746 wakeup_mycpu(&rdd->helper_thread); /* XXX */
753 * We should interrupt the currently running thread, which
754 * is on a different cpu.
756 spin_unlock(&rdd->spin);
757 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
761 * Request a reschedule if appropriate.
763 spin_lock(&rdd->spin);
764 dfly_setrunqueue_locked(rdd, lp);
765 spin_unlock(&rdd->spin);
766 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
773 * This routine is called from a systimer IPI. It MUST be MP-safe and
774 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
779 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
781 globaldata_t gd = mycpu;
783 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
787 * Spinlocks also hold a critical section so there should not be
790 KKASSERT(gd->gd_spinlocks == 0);
796 * Do we need to round-robin? We round-robin 10 times a second.
797 * This should only occur for cpu-bound batch processes.
799 if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
800 lp->lwp_thread->td_wakefromcpu = -1;
805 * Adjust estcpu upward using a real time equivalent calculation,
806 * and recalculate lp's priority.
808 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
809 dfly_resetpriority(lp);
812 * Rebalance two cpus every 8 ticks, pulling the worst thread
813 * from the worst cpu's queue into a rotating cpu number.
815 * This mechanic is needed because the push algorithms can
816 * steady-state in an non-optimal configuration. We need to mix it
817 * up a little, even if it means breaking up a paired thread, so
818 * the push algorithms can rebalance the degenerate conditions.
819 * This portion of the algorithm exists to ensure stability at the
820 * selected weightings.
822 * Because we might be breaking up optimal conditions we do not want
823 * to execute this too quickly, hence we only rebalance approximately
824 * ~7-8 times per second. The push's, on the otherhand, are capable
825 * moving threads to other cpus at a much higher rate.
827 * We choose the most heavily loaded thread from the worst queue
828 * in order to ensure that multiple heavy-weight threads on the same
829 * queue get broken up, and also because these threads are the most
830 * likely to be able to remain in place. Hopefully then any pairings,
831 * if applicable, migrate to where these threads are.
834 if ((usched_dfly_features & 0x04) &&
835 ((u_int)sched_ticks & 7) == 0 &&
836 (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
843 rdd = dfly_choose_worst_queue(dd);
845 spin_lock(&dd->spin);
846 if (spin_trylock(&rdd->spin)) {
847 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
848 spin_unlock(&rdd->spin);
850 spin_unlock(&dd->spin);
852 spin_unlock(&dd->spin);
858 /* dd->spin held if nlp != NULL */
861 * Either schedule it or add it to our queue.
864 (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
865 atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
866 dd->upri = nlp->lwp_priority;
869 dd->rrcount = 0; /* reset round robin */
871 spin_unlock(&dd->spin);
872 lwkt_acquire(nlp->lwp_thread);
873 lwkt_schedule(nlp->lwp_thread);
875 dfly_setrunqueue_locked(dd, nlp);
876 spin_unlock(&dd->spin);
883 * Called from acquire and from kern_synch's one-second timer (one of the
884 * callout helper threads) with a critical section held.
886 * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
887 * overall system load.
889 * Note that no recalculation occurs for a process which sleeps and wakes
890 * up in the same tick. That is, a system doing thousands of context
891 * switches per second will still only do serious estcpu calculations
892 * ESTCPUFREQ times per second.
896 dfly_recalculate_estcpu(struct lwp *lp)
898 globaldata_t gd = mycpu;
906 * We have to subtract periodic to get the last schedclock
907 * timeout time, otherwise we would get the upcoming timeout.
908 * Keep in mind that a process can migrate between cpus and
909 * while the scheduler clock should be very close, boundary
910 * conditions could lead to a small negative delta.
912 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
914 if (lp->lwp_slptime > 1) {
916 * Too much time has passed, do a coarse correction.
918 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
919 dfly_resetpriority(lp);
920 lp->lwp_cpbase = cpbase;
923 } else if (lp->lwp_cpbase != cpbase) {
925 * Adjust estcpu if we are in a different tick. Don't waste
926 * time if we are in the same tick.
928 * First calculate the number of ticks in the measurement
929 * interval. The ttlticks calculation can wind up 0 due to
930 * a bug in the handling of lwp_slptime (as yet not found),
931 * so make sure we do not get a divide by 0 panic.
933 ttlticks = (cpbase - lp->lwp_cpbase) /
934 gd->gd_schedclock.periodic;
935 if ((ssysclock_t)ttlticks < 0) {
937 lp->lwp_cpbase = cpbase;
941 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
944 * Calculate the percentage of one cpu being used then
945 * compensate for any system load in excess of ncpus.
947 * For example, if we have 8 cores and 16 running cpu-bound
948 * processes then all things being equal each process will
949 * get 50% of one cpu. We need to pump this value back
950 * up to 100% so the estcpu calculation properly adjusts
951 * the process's dynamic priority.
953 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
955 estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
956 ucount = dfly_ucount;
957 if (ucount > ncpus) {
958 estcpu += estcpu * (ucount - ncpus) / ncpus;
961 if (usched_dfly_debug == lp->lwp_proc->p_pid) {
962 kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
963 lp->lwp_proc->p_pid, lp,
964 estcpu, lp->lwp_estcpu,
965 lp->lwp_cpticks, ttlticks);
969 * Adjust lp->lwp_esetcpu. The decay factor determines how
970 * quickly lwp_estcpu collapses to its realtime calculation.
971 * A slower collapse gives us a more accurate number over
972 * the long term but can create problems with bursty threads
973 * or threads which become cpu hogs.
975 * To solve this problem, newly started lwps and lwps which
976 * are restarting after having been asleep for a while are
977 * given a much, much faster decay in order to quickly
978 * detect whether they become cpu-bound.
980 * NOTE: p_nice is accounted for in dfly_resetpriority(),
981 * and not here, but we must still ensure that a
982 * cpu-bound nice -20 process does not completely
983 * override a cpu-bound nice +20 process.
985 * NOTE: We must use ESTCPULIM() here to deal with any
988 decay_factor = usched_dfly_decay;
989 if (decay_factor < 1)
991 if (decay_factor > 1024)
994 if (lp->lwp_estfast < usched_dfly_decay) {
996 lp->lwp_estcpu = ESTCPULIM(
997 (lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
998 (lp->lwp_estfast + 1));
1000 lp->lwp_estcpu = ESTCPULIM(
1001 (lp->lwp_estcpu * decay_factor + estcpu) /
1002 (decay_factor + 1));
1005 if (usched_dfly_debug == lp->lwp_proc->p_pid)
1006 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
1007 dfly_resetpriority(lp);
1008 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
1009 lp->lwp_cpticks = 0;
1014 * Compute the priority of a process when running in user mode.
1015 * Arrange to reschedule if the resulting priority is better
1016 * than that of the current process.
1018 * This routine may be called with any process.
1020 * This routine is called by fork1() for initial setup with the process
1021 * of the run queue, and also may be called normally with the process on or
1022 * off the run queue.
1025 dfly_resetpriority(struct lwp *lp)
1038 * Lock the scheduler (lp) belongs to. This can be on a different
1039 * cpu. Handle races. This loop breaks out with the appropriate
1043 rcpu = lp->lwp_qcpu;
1045 rdd = &dfly_pcpu[rcpu];
1046 spin_lock(&rdd->spin);
1047 if (rcpu == lp->lwp_qcpu)
1049 spin_unlock(&rdd->spin);
1053 * Calculate the new priority and queue type
1055 newrqtype = lp->lwp_rtprio.type;
1058 case RTP_PRIO_REALTIME:
1060 newpriority = PRIBASE_REALTIME +
1061 (lp->lwp_rtprio.prio & PRIMASK);
1063 case RTP_PRIO_NORMAL:
1067 estcpu = lp->lwp_estcpu;
1070 * p_nice piece Adds (0-40) * 2 0-80
1071 * estcpu Adds 16384 * 4 / 512 0-128
1073 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1074 newpriority += estcpu * PPQ / ESTCPUPPQ;
1075 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1076 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1077 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1080 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1082 case RTP_PRIO_THREAD:
1083 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1086 panic("Bad RTP_PRIO %d", newrqtype);
1091 * The LWKT scheduler doesn't dive usched structures, give it a hint
1092 * on the relative priority of user threads running in the kernel.
1093 * The LWKT scheduler will always ensure that a user thread running
1094 * in the kernel will get cpu some time, regardless of its upri,
1095 * but can decide not to instantly switch from one kernel or user
1096 * mode user thread to a kernel-mode user thread when it has a less
1097 * desireable user priority.
1099 * td_upri has normal sense (higher values are more desireable), so
1102 lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1105 * The newpriority incorporates the queue type so do a simple masked
1106 * check to determine if the process has moved to another queue. If
1107 * it has, and it is currently on a run queue, then move it.
1109 * Since uload is ~PPQMASK masked, no modifications are necessary if
1110 * we end up in the same run queue.
1112 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1113 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1114 dfly_remrunqueue_locked(rdd, lp);
1115 lp->lwp_priority = newpriority;
1116 lp->lwp_rqtype = newrqtype;
1117 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1118 dfly_setrunqueue_locked(rdd, lp);
1121 lp->lwp_priority = newpriority;
1122 lp->lwp_rqtype = newrqtype;
1123 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1128 * In the same PPQ, uload cannot change.
1130 lp->lwp_priority = newpriority;
1136 * Adjust effective load.
1138 * Calculate load then scale up or down geometrically based on p_nice.
1139 * Processes niced up (positive) are less important, and processes
1140 * niced downard (negative) are more important. The higher the uload,
1141 * the more important the thread.
1143 /* 0-511, 0-100% cpu */
1144 delta_uload = lp->lwp_estcpu / NQS;
1145 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1148 delta_uload -= lp->lwp_uload;
1149 lp->lwp_uload += delta_uload;
1150 if (lp->lwp_mpflags & LWP_MP_ULOAD)
1151 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1154 * Determine if we need to reschedule the target cpu. This only
1155 * occurs if the LWP is already on a scheduler queue, which means
1156 * that idle cpu notification has already occured. At most we
1157 * need only issue a need_user_resched() on the appropriate cpu.
1159 * The LWP may be owned by a CPU different from the current one,
1160 * in which case dd->uschedcp may be modified without an MP lock
1161 * or a spinlock held. The worst that happens is that the code
1162 * below causes a spurious need_user_resched() on the target CPU
1163 * and dd->pri to be wrong for a short period of time, both of
1164 * which are harmless.
1166 * If checkpri is 0 we are adjusting the priority of the current
1167 * process, possibly higher (less desireable), so ignore the upri
1168 * check which will fail in that case.
1171 if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1173 (rdd->upri & ~PRIMASK) >
1174 (lp->lwp_priority & ~PRIMASK))) {
1176 if (rcpu == mycpu->gd_cpuid) {
1177 spin_unlock(&rdd->spin);
1178 need_user_resched();
1180 spin_unlock(&rdd->spin);
1181 lwkt_send_ipiq(globaldata_find(rcpu),
1182 dfly_need_user_resched_remote,
1186 spin_unlock(&rdd->spin);
1187 need_user_resched();
1190 spin_unlock(&rdd->spin);
1193 spin_unlock(&rdd->spin);
1200 dfly_yield(struct lwp *lp)
1203 /* FUTURE (or something similar) */
1204 switch(lp->lwp_rqtype) {
1205 case RTP_PRIO_NORMAL:
1206 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
1212 need_user_resched();
1216 * Called from fork1() when a new child process is being created.
1218 * Give the child process an initial estcpu that is more batch then
1219 * its parent and dock the parent for the fork (but do not
1220 * reschedule the parent).
1224 * XXX lwp should be "spawning" instead of "forking"
1227 dfly_forking(struct lwp *plp, struct lwp *lp)
1230 * Put the child 4 queue slots (out of 32) higher than the parent
1231 * (less desireable than the parent).
1233 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1235 lp->lwp_estfast = 0;
1238 * Dock the parent a cost for the fork, protecting us from fork
1239 * bombs. If the parent is forking quickly make the child more
1242 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1246 * Called when a lwp is being removed from this scheduler, typically
1247 * during lwp_exit(). We have to clean out any ULOAD accounting before
1248 * we can let the lp go. The dd->spin lock is not needed for uload
1251 * Scheduler dequeueing has already occurred, no further action in that
1255 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1257 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1259 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1260 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1261 atomic_add_int(&dd->uload, -lp->lwp_uload);
1262 atomic_add_int(&dd->ucount, -1);
1263 atomic_add_int(&dfly_ucount, -1);
1268 * This function cannot block in any way, but spinlocks are ok.
1270 * Update the uload based on the state of the thread (whether it is going
1271 * to sleep or running again). The uload is meant to be a longer-term
1272 * load and not an instantanious load.
1275 dfly_uload_update(struct lwp *lp)
1277 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1279 if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1280 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1281 spin_lock(&dd->spin);
1282 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1283 atomic_set_int(&lp->lwp_mpflags,
1285 atomic_add_int(&dd->uload, lp->lwp_uload);
1286 atomic_add_int(&dd->ucount, 1);
1287 atomic_add_int(&dfly_ucount, 1);
1289 spin_unlock(&dd->spin);
1291 } else if (lp->lwp_slptime > 0) {
1292 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1293 spin_lock(&dd->spin);
1294 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1295 atomic_clear_int(&lp->lwp_mpflags,
1297 atomic_add_int(&dd->uload, -lp->lwp_uload);
1298 atomic_add_int(&dd->ucount, -1);
1299 atomic_add_int(&dfly_ucount, -1);
1301 spin_unlock(&dd->spin);
1307 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1308 * it selects a user process and returns it. If chklp is non-NULL and chklp
1309 * has a better or equal priority then the process that would otherwise be
1310 * chosen, NULL is returned.
1312 * Until we fix the RUNQ code the chklp test has to be strict or we may
1313 * bounce between processes trying to acquire the current process designation.
1315 * Must be called with rdd->spin locked. The spinlock is left intact through
1316 * the entire routine. dd->spin does not have to be locked.
1318 * If worst is non-zero this function finds the worst thread instead of the
1319 * best thread (used by the schedulerclock-based rover).
1323 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1324 struct lwp *chklp, int worst)
1328 u_int32_t *which, *which2;
1334 rtqbits = rdd->rtqueuebits;
1335 tsqbits = rdd->queuebits;
1336 idqbits = rdd->idqueuebits;
1340 pri = bsrl(idqbits);
1341 q = &rdd->idqueues[pri];
1342 which = &rdd->idqueuebits;
1344 } else if (tsqbits) {
1345 pri = bsrl(tsqbits);
1346 q = &rdd->queues[pri];
1347 which = &rdd->queuebits;
1349 } else if (rtqbits) {
1350 pri = bsrl(rtqbits);
1351 q = &rdd->rtqueues[pri];
1352 which = &rdd->rtqueuebits;
1357 lp = TAILQ_LAST(q, rq);
1360 pri = bsfl(rtqbits);
1361 q = &rdd->rtqueues[pri];
1362 which = &rdd->rtqueuebits;
1364 } else if (tsqbits) {
1365 pri = bsfl(tsqbits);
1366 q = &rdd->queues[pri];
1367 which = &rdd->queuebits;
1369 } else if (idqbits) {
1370 pri = bsfl(idqbits);
1371 q = &rdd->idqueues[pri];
1372 which = &rdd->idqueuebits;
1377 lp = TAILQ_FIRST(q);
1379 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1382 * If the passed lwp <chklp> is reasonably close to the selected
1383 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1385 * Note that we must error on the side of <chklp> to avoid bouncing
1386 * between threads in the acquire code.
1389 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1393 KTR_COND_LOG(usched_chooseproc,
1394 lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1395 lp->lwp_proc->p_pid,
1396 lp->lwp_thread->td_gd->gd_cpuid,
1399 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1400 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1401 TAILQ_REMOVE(q, lp, lwp_procq);
1404 *which &= ~(1 << pri);
1407 * If we are choosing a process from rdd with the intent to
1408 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1412 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1413 atomic_add_int(&rdd->uload, -lp->lwp_uload);
1414 atomic_add_int(&rdd->ucount, -1);
1415 atomic_add_int(&dfly_ucount, -1);
1417 lp->lwp_qcpu = dd->cpuid;
1418 atomic_add_int(&dd->uload, lp->lwp_uload);
1419 atomic_add_int(&dd->ucount, 1);
1420 atomic_add_int(&dfly_ucount, 1);
1421 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1429 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1431 * Choose a cpu node to schedule lp on, hopefully nearby its current
1434 * We give the current node a modest advantage for obvious reasons.
1436 * We also give the node the thread was woken up FROM a slight advantage
1437 * in order to try to schedule paired threads which synchronize/block waiting
1438 * for each other fairly close to each other. Similarly in a network setting
1439 * this feature will also attempt to place a user process near the kernel
1440 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the
1441 * algorithm as it heuristically groups synchronizing processes for locality
1442 * of reference in multi-socket systems.
1444 * We check against running processes and give a big advantage if there
1447 * The caller will normally dfly_setrunqueue() lp on the returned queue.
1449 * When the topology is known choose a cpu whos group has, in aggregate,
1450 * has the lowest weighted load.
1454 dfly_choose_best_queue(struct lwp *lp)
1461 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1471 * When the topology is unknown choose a random cpu that is hopefully
1474 if (dd->cpunode == NULL)
1475 return (dfly_choose_queue_simple(dd, lp));
1480 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1481 wakemask = dfly_pcpu[wakecpu].cpumask;
1486 * When the topology is known choose a cpu whos group has, in
1487 * aggregate, has the lowest weighted load.
1489 cpup = root_cpu_node;
1494 * Degenerate case super-root
1496 if (cpup->child_node && cpup->child_no == 1) {
1497 cpup = cpup->child_node;
1504 if (cpup->child_node == NULL) {
1505 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1510 lowest_load = 0x7FFFFFFF;
1512 for (n = 0; n < cpup->child_no; ++n) {
1514 * Accumulate load information for all cpus
1515 * which are members of this node.
1517 cpun = &cpup->child_node[n];
1518 mask = cpun->members & usched_global_cpumask &
1519 smp_active_mask & lp->lwp_cpumask;
1527 cpuid = BSFCPUMASK(mask);
1528 rdd = &dfly_pcpu[cpuid];
1530 load += rdd->ucount * usched_dfly_weight3;
1532 if (rdd->uschedcp == NULL &&
1533 rdd->runqcount == 0 &&
1534 globaldata_find(cpuid)->gd_tdrunqcount == 0
1536 load -= usched_dfly_weight4;
1539 else if (rdd->upri > lp->lwp_priority + PPQ) {
1540 load -= usched_dfly_weight4 / 2;
1543 mask &= ~CPUMASK(cpuid);
1548 * Compensate if the lp is already accounted for in
1549 * the aggregate uload for this mask set. We want
1550 * to calculate the loads as if lp were not present,
1551 * otherwise the calculation is bogus.
1553 if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1554 (dd->cpumask & cpun->members)) {
1555 load -= lp->lwp_uload;
1556 load -= usched_dfly_weight3;
1562 * Advantage the cpu group (lp) is already on.
1564 if (cpun->members & dd->cpumask)
1565 load -= usched_dfly_weight1;
1568 * Advantage the cpu group we want to pair (lp) to,
1569 * but don't let it go to the exact same cpu as
1570 * the wakecpu target.
1572 * We do this by checking whether cpun is a
1573 * terminal node or not. All cpun's at the same
1574 * level will either all be terminal or all not
1577 * If it is and we match we disadvantage the load.
1578 * If it is and we don't match we advantage the load.
1580 * Also note that we are effectively disadvantaging
1581 * all-but-one by the same amount, so it won't effect
1582 * the weight1 factor for the all-but-one nodes.
1584 if (cpun->members & wakemask) {
1585 if (cpun->child_node != NULL) {
1587 load -= usched_dfly_weight2;
1589 if (usched_dfly_features & 0x10)
1590 load += usched_dfly_weight2;
1592 load -= usched_dfly_weight2;
1597 * Calculate the best load
1599 if (cpub == NULL || lowest_load > load ||
1600 (lowest_load == load &&
1601 (cpun->members & dd->cpumask))
1609 if (usched_dfly_chooser)
1610 kprintf("lp %02d->%02d %s\n",
1611 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1616 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1618 * Choose the worst queue close to dd's cpu node with a non-empty runq
1619 * that is NOT dd. Also require that the moving of the highest-load thread
1620 * from rdd to dd does not cause the uload's to cross each other.
1622 * This is used by the thread chooser when the current cpu's queues are
1623 * empty to steal a thread from another cpu's queue. We want to offload
1624 * the most heavily-loaded queue.
1628 dfly_choose_worst_queue(dfly_pcpu_t dd)
1646 * When the topology is unknown choose a random cpu that is hopefully
1649 if (dd->cpunode == NULL) {
1654 * When the topology is known choose a cpu whos group has, in
1655 * aggregate, has the lowest weighted load.
1657 cpup = root_cpu_node;
1661 * Degenerate case super-root
1663 if (cpup->child_node && cpup->child_no == 1) {
1664 cpup = cpup->child_node;
1671 if (cpup->child_node == NULL) {
1672 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1679 for (n = 0; n < cpup->child_no; ++n) {
1681 * Accumulate load information for all cpus
1682 * which are members of this node.
1684 cpun = &cpup->child_node[n];
1685 mask = cpun->members & usched_global_cpumask &
1693 cpuid = BSFCPUMASK(mask);
1694 rdd = &dfly_pcpu[cpuid];
1696 load += rdd->ucount * usched_dfly_weight3;
1697 if (rdd->uschedcp == NULL &&
1698 rdd->runqcount == 0 &&
1699 globaldata_find(cpuid)->gd_tdrunqcount == 0
1701 load -= usched_dfly_weight4;
1704 else if (rdd->upri > dd->upri + PPQ) {
1705 load -= usched_dfly_weight4 / 2;
1708 mask &= ~CPUMASK(cpuid);
1714 * Prefer candidates which are somewhat closer to
1717 if (dd->cpumask & cpun->members)
1718 load += usched_dfly_weight1;
1721 * The best candidate is the one with the worst
1724 if (cpub == NULL || highest_load < load) {
1725 highest_load = load;
1733 * We never return our own node (dd), and only return a remote
1734 * node if it's load is significantly worse than ours (i.e. where
1735 * stealing a thread would be considered reasonable).
1737 * This also helps us avoid breaking paired threads apart which
1738 * can have disastrous effects on performance.
1745 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1747 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1749 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1752 if (rdd->uload - hpri < dd->uload + hpri)
1760 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1768 * Fallback to the original heuristic, select random cpu,
1769 * first checking cpus not currently running a user thread.
1772 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1773 mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1774 smp_active_mask & usched_global_cpumask;
1777 tmpmask = ~(CPUMASK(cpuid) - 1);
1779 cpuid = BSFCPUMASK(mask & tmpmask);
1781 cpuid = BSFCPUMASK(mask);
1782 rdd = &dfly_pcpu[cpuid];
1784 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1786 mask &= ~CPUMASK(cpuid);
1790 * Then cpus which might have a currently running lp
1792 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1793 mask = dfly_curprocmask & dfly_rdyprocmask &
1794 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1797 tmpmask = ~(CPUMASK(cpuid) - 1);
1799 cpuid = BSFCPUMASK(mask & tmpmask);
1801 cpuid = BSFCPUMASK(mask);
1802 rdd = &dfly_pcpu[cpuid];
1804 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1806 mask &= ~CPUMASK(cpuid);
1810 * If we cannot find a suitable cpu we reload from dfly_scancpu
1811 * and round-robin. Other cpus will pickup as they release their
1812 * current lwps or become ready.
1814 * Avoid a degenerate system lockup case if usched_global_cpumask
1815 * is set to 0 or otherwise does not cover lwp_cpumask.
1817 * We only kick the target helper thread in this case, we do not
1818 * set the user resched flag because
1820 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1821 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1823 rdd = &dfly_pcpu[cpuid];
1830 dfly_need_user_resched_remote(void *dummy)
1832 globaldata_t gd = mycpu;
1833 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
1836 * Flag reschedule needed
1838 need_user_resched();
1841 * If no user thread is currently running we need to kick the helper
1842 * on our cpu to recover. Otherwise the cpu will never schedule
1845 * We cannot schedule the process ourselves because this is an
1846 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1848 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1850 if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1851 atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1852 wakeup_mycpu(&dd->helper_thread);
1859 * dfly_remrunqueue_locked() removes a given process from the run queue
1860 * that it is on, clearing the queue busy bit if it becomes empty.
1862 * Note that user process scheduler is different from the LWKT schedule.
1863 * The user process scheduler only manages user processes but it uses LWKT
1864 * underneath, and a user process operating in the kernel will often be
1865 * 'released' from our management.
1867 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes
1868 * to sleep or the lwp is moved to a different runq.
1871 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1877 KKASSERT(rdd->runqcount >= 0);
1879 pri = lp->lwp_rqindex;
1881 switch(lp->lwp_rqtype) {
1882 case RTP_PRIO_NORMAL:
1883 q = &rdd->queues[pri];
1884 which = &rdd->queuebits;
1886 case RTP_PRIO_REALTIME:
1888 q = &rdd->rtqueues[pri];
1889 which = &rdd->rtqueuebits;
1892 q = &rdd->idqueues[pri];
1893 which = &rdd->idqueuebits;
1896 panic("remrunqueue: invalid rtprio type");
1899 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1900 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1901 TAILQ_REMOVE(q, lp, lwp_procq);
1903 if (TAILQ_EMPTY(q)) {
1904 KASSERT((*which & (1 << pri)) != 0,
1905 ("remrunqueue: remove from empty queue"));
1906 *which &= ~(1 << pri);
1911 * dfly_setrunqueue_locked()
1913 * Add a process whos rqtype and rqindex had previously been calculated
1914 * onto the appropriate run queue. Determine if the addition requires
1915 * a reschedule on a cpu and return the cpuid or -1.
1917 * NOTE: Lower priorities are better priorities.
1919 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1920 * sum of the rough lwp_priority for all running and runnable
1921 * processes. Lower priority processes (higher lwp_priority
1922 * values) actually DO count as more load, not less, because
1923 * these are the programs which require the most care with
1924 * regards to cpu selection.
1927 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1933 KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1935 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1936 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1937 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1938 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1939 atomic_add_int(&dfly_ucount, 1);
1942 pri = lp->lwp_rqindex;
1944 switch(lp->lwp_rqtype) {
1945 case RTP_PRIO_NORMAL:
1946 q = &rdd->queues[pri];
1947 which = &rdd->queuebits;
1949 case RTP_PRIO_REALTIME:
1951 q = &rdd->rtqueues[pri];
1952 which = &rdd->rtqueuebits;
1955 q = &rdd->idqueues[pri];
1956 which = &rdd->idqueuebits;
1959 panic("remrunqueue: invalid rtprio type");
1964 * Place us on the selected queue. Determine if we should be
1965 * placed at the head of the queue or at the end.
1967 * We are placed at the tail if our round-robin count has expired,
1968 * or is about to expire and the system thinks its a good place to
1969 * round-robin, or there is already a next thread on the queue
1970 * (it might be trying to pick up where it left off and we don't
1971 * want to interfere).
1973 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1974 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1977 if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1978 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1979 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1982 atomic_clear_int(&lp->lwp_thread->td_mpflags,
1983 TDF_MP_BATCH_DEMARC);
1984 lp->lwp_rrcount = 0;
1985 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1988 lp->lwp_rrcount = 0;
1989 TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1997 * For SMP systems a user scheduler helper thread is created for each
1998 * cpu and is used to allow one cpu to wakeup another for the purposes of
1999 * scheduling userland threads from setrunqueue().
2001 * UP systems do not need the helper since there is only one cpu.
2003 * We can't use the idle thread for this because we might block.
2004 * Additionally, doing things this way allows us to HLT idle cpus
2008 dfly_helper_thread(void *dummy)
2018 cpuid = gd->gd_cpuid; /* doesn't change */
2019 mask = gd->gd_cpumask; /* doesn't change */
2020 dd = &dfly_pcpu[cpuid];
2023 * Since we only want to be woken up only when no user processes
2024 * are scheduled on a cpu, run at an ultra low priority.
2026 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
2028 tsleep(&dd->helper_thread, 0, "schslp", 0);
2032 * We use the LWKT deschedule-interlock trick to avoid racing
2033 * dfly_rdyprocmask. This means we cannot block through to the
2034 * manual lwkt_switch() call we make below.
2037 tsleep_interlock(&dd->helper_thread, 0);
2039 spin_lock(&dd->spin);
2041 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2042 clear_user_resched(); /* This satisfied the reschedule request */
2044 dd->rrcount = 0; /* Reset the round-robin counter */
2047 if (dd->runqcount || dd->uschedcp != NULL) {
2049 * Threads are available. A thread may or may not be
2050 * currently scheduled. Get the best thread already queued
2053 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
2055 atomic_set_cpumask(&dfly_curprocmask, mask);
2056 dd->upri = nlp->lwp_priority;
2059 dd->rrcount = 0; /* reset round robin */
2061 spin_unlock(&dd->spin);
2062 lwkt_acquire(nlp->lwp_thread);
2063 lwkt_schedule(nlp->lwp_thread);
2066 * This situation should not occur because we had
2067 * at least one thread available.
2069 spin_unlock(&dd->spin);
2071 } else if (usched_dfly_features & 0x01) {
2073 * This cpu is devoid of runnable threads, steal a thread
2074 * from another cpu. Since we're stealing, might as well
2075 * load balance at the same time.
2077 * We choose the highest-loaded thread from the worst queue.
2079 * NOTE! This function only returns a non-NULL rdd when
2080 * another cpu's queue is obviously overloaded. We
2081 * do not want to perform the type of rebalancing
2082 * the schedclock does here because it would result
2083 * in insane process pulling when 'steady' state is
2084 * partially unbalanced (e.g. 6 runnables and only
2087 rdd = dfly_choose_worst_queue(dd);
2088 if (rdd && spin_trylock(&rdd->spin)) {
2089 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2090 spin_unlock(&rdd->spin);
2095 atomic_set_cpumask(&dfly_curprocmask, mask);
2096 dd->upri = nlp->lwp_priority;
2099 dd->rrcount = 0; /* reset round robin */
2101 spin_unlock(&dd->spin);
2102 lwkt_acquire(nlp->lwp_thread);
2103 lwkt_schedule(nlp->lwp_thread);
2106 * Leave the thread on our run queue. Another
2107 * scheduler will try to pull it later.
2109 spin_unlock(&dd->spin);
2113 * devoid of runnable threads and not allowed to steal
2116 spin_unlock(&dd->spin);
2120 * We're descheduled unless someone scheduled us. Switch away.
2121 * Exiting the critical section will cause splz() to be called
2122 * for us if interrupts and such are pending.
2125 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2131 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2135 new_val = usched_dfly_stick_to_level;
2137 error = sysctl_handle_int(oidp, &new_val, 0, req);
2138 if (error != 0 || req->newptr == NULL)
2140 if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2142 usched_dfly_stick_to_level = new_val;
2150 * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2151 * Note that curprocmask bit 0 has already been cleared by rqinit() and
2152 * we should not mess with it further.
2155 usched_dfly_cpu_init(void)
2161 int smt_not_supported = 0;
2162 int cache_coherent_not_supported = 0;
2166 kprintf("Start scheduler helpers on cpus:\n");
2168 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2169 usched_dfly_sysctl_tree =
2170 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2171 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2172 "usched_dfly", CTLFLAG_RD, 0, "");
2174 for (i = 0; i < ncpus; ++i) {
2175 dfly_pcpu_t dd = &dfly_pcpu[i];
2176 cpumask_t mask = CPUMASK(i);
2178 if ((mask & smp_active_mask) == 0)
2181 spin_init(&dd->spin);
2183 dd->cpunode = get_cpu_node_by_cpuid(i);
2186 dd->cpumask = CPUMASK(i);
2187 for (j = 0; j < NQS; j++) {
2188 TAILQ_INIT(&dd->queues[j]);
2189 TAILQ_INIT(&dd->rtqueues[j]);
2190 TAILQ_INIT(&dd->idqueues[j]);
2192 atomic_clear_cpumask(&dfly_curprocmask, 1);
2195 if (dd->cpunode == NULL) {
2196 smt_not_supported = 1;
2197 cache_coherent_not_supported = 1;
2199 kprintf ("\tcpu%d - WARNING: No CPU NODE "
2200 "found for cpu\n", i);
2202 switch (dd->cpunode->type) {
2205 kprintf ("\tcpu%d - HyperThreading "
2206 "available. Core siblings: ",
2210 smt_not_supported = 1;
2213 kprintf ("\tcpu%d - No HT available, "
2214 "multi-core/physical "
2215 "cpu. Physical siblings: ",
2219 smt_not_supported = 1;
2222 kprintf ("\tcpu%d - No HT available, "
2223 "single-core/physical cpu. "
2224 "Package Siblings: ",
2228 /* Let's go for safe defaults here */
2229 smt_not_supported = 1;
2230 cache_coherent_not_supported = 1;
2232 kprintf ("\tcpu%d - Unknown cpunode->"
2233 "type=%u. Siblings: ",
2235 (u_int)dd->cpunode->type);
2240 if (dd->cpunode->parent_node != NULL) {
2241 CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2242 kprintf("cpu%d ", cpuid);
2245 kprintf(" no siblings\n");
2250 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2251 0, i, "usched %d", i);
2255 * Allow user scheduling on the target cpu. cpu #0 has already
2256 * been enabled in rqinit().
2259 atomic_clear_cpumask(&dfly_curprocmask, mask);
2260 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2261 dd->upri = PRIBASE_NULL;
2265 /* usched_dfly sysctl configurable parameters */
2267 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2268 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2269 OID_AUTO, "rrinterval", CTLFLAG_RW,
2270 &usched_dfly_rrinterval, 0, "");
2271 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2272 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2273 OID_AUTO, "decay", CTLFLAG_RW,
2274 &usched_dfly_decay, 0, "Extra decay when not running");
2277 /* Add enable/disable option for SMT scheduling if supported */
2278 if (smt_not_supported) {
2279 usched_dfly_smt = 0;
2280 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2281 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2282 OID_AUTO, "smt", CTLFLAG_RD,
2283 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2285 usched_dfly_smt = 1;
2286 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2287 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2288 OID_AUTO, "smt", CTLFLAG_RW,
2289 &usched_dfly_smt, 0, "Enable SMT scheduling");
2293 * Add enable/disable option for cache coherent scheduling
2296 if (cache_coherent_not_supported) {
2297 usched_dfly_cache_coherent = 0;
2298 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2299 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2300 OID_AUTO, "cache_coherent", CTLFLAG_RD,
2302 "Cache coherence NOT SUPPORTED");
2304 usched_dfly_cache_coherent = 1;
2305 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2306 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2307 OID_AUTO, "cache_coherent", CTLFLAG_RW,
2308 &usched_dfly_cache_coherent, 0,
2309 "Enable/Disable cache coherent scheduling");
2311 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2312 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2313 OID_AUTO, "weight1", CTLFLAG_RW,
2314 &usched_dfly_weight1, 200,
2315 "Weight selection for current cpu");
2317 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2318 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2319 OID_AUTO, "weight2", CTLFLAG_RW,
2320 &usched_dfly_weight2, 180,
2321 "Weight selection for wakefrom cpu");
2323 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2324 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2325 OID_AUTO, "weight3", CTLFLAG_RW,
2326 &usched_dfly_weight3, 40,
2327 "Weight selection for num threads on queue");
2329 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2330 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2331 OID_AUTO, "weight4", CTLFLAG_RW,
2332 &usched_dfly_weight4, 160,
2333 "Availability of other idle cpus");
2335 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2336 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2337 OID_AUTO, "fast_resched", CTLFLAG_RW,
2338 &usched_dfly_fast_resched, 0,
2339 "Availability of other idle cpus");
2341 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2342 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2343 OID_AUTO, "features", CTLFLAG_RW,
2344 &usched_dfly_features, 0x8F,
2345 "Allow pulls into empty queues");
2347 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2348 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2349 OID_AUTO, "swmask", CTLFLAG_RW,
2350 &usched_dfly_swmask, ~PPQMASK,
2351 "Queue mask to force thread switch");
2354 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2355 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2356 OID_AUTO, "stick_to_level",
2357 CTLTYPE_INT | CTLFLAG_RW,
2358 NULL, sizeof usched_dfly_stick_to_level,
2359 sysctl_usched_dfly_stick_to_level, "I",
2360 "Stick a process to this level. See sysctl"
2361 "paremter hw.cpu_topology.level_description");
2366 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2367 usched_dfly_cpu_init, NULL)