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);
148 struct usched usched_dfly = {
150 "dfly", "Original DragonFly Scheduler",
151 NULL, /* default registration */
152 NULL, /* default deregistration */
153 dfly_acquire_curproc,
154 dfly_release_curproc,
157 dfly_recalculate_estcpu,
162 NULL, /* setcpumask not supported */
167 * We have NQS (32) run queues per scheduling class. For the normal
168 * class, there are 128 priorities scaled onto these 32 queues. New
169 * processes are added to the last entry in each queue, and processes
170 * are selected for running by taking them from the head and maintaining
171 * a simple FIFO arrangement. Realtime and Idle priority processes have
172 * and explicit 0-31 priority which maps directly onto their class queue
173 * index. When a queue has something in it, the corresponding bit is
174 * set in the queuebits variable, allowing a single read to determine
175 * the state of all 32 queues and then a ffs() to find the first busy
178 static cpumask_t dfly_curprocmask = -1; /* currently running a user process */
179 static cpumask_t dfly_rdyprocmask; /* ready to accept a user process */
180 static volatile int dfly_scancpu;
181 static volatile int dfly_ucount; /* total running on whole system */
182 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
183 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
184 static struct sysctl_oid *usched_dfly_sysctl_tree;
186 /* Debug info exposed through debug.* sysctl */
188 static int usched_dfly_debug = -1;
189 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
190 &usched_dfly_debug, 0,
191 "Print debug information for this pid");
193 static int usched_dfly_pid_debug = -1;
194 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
195 &usched_dfly_pid_debug, 0,
196 "Print KTR debug information for this pid");
198 static int usched_dfly_chooser = 0;
199 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
200 &usched_dfly_chooser, 0,
201 "Print KTR debug information for this pid");
204 * Tunning usched_dfly - configurable through kern.usched_dfly.
206 * weight1 - Tries to keep threads on their current cpu. If you
207 * make this value too large the scheduler will not be
208 * able to load-balance large loads.
210 * weight2 - If non-zero, detects thread pairs undergoing synchronous
211 * communications and tries to move them closer together.
212 * Behavior is adjusted by bit 4 of features (0x10).
214 * WARNING! Weight2 is a ridiculously sensitive parameter,
215 * a small value is recommended.
217 * weight3 - Weighting based on the number of recently runnable threads
218 * on the userland scheduling queue (ignoring their loads).
219 * A nominal value here prevents high-priority (low-load)
220 * threads from accumulating on one cpu core when other
221 * cores are available.
223 * This value should be left fairly small relative to weight1
226 * weight4 - Weighting based on other cpu queues being available
227 * or running processes with higher lwp_priority's.
229 * This allows a thread to migrate to another nearby cpu if it
230 * is unable to run on the current cpu based on the other cpu
231 * being idle or running a lower priority (higher lwp_priority)
232 * thread. This value should be large enough to override weight1
234 * features - These flags can be set or cleared to enable or disable various
237 * 0x01 Enable idle-cpu pulling (default)
238 * 0x02 Enable proactive pushing (default)
239 * 0x04 Enable rebalancing rover (default)
240 * 0x08 Enable more proactive pushing (default)
241 * 0x10 (flip weight2 limit on same cpu) (default)
242 * 0x20 choose best cpu for forked process
243 * 0x40 choose current cpu for forked process
244 * 0x80 choose random cpu for forked process (default)
246 static int usched_dfly_smt = 0;
247 static int usched_dfly_cache_coherent = 0;
248 static int usched_dfly_weight1 = 200; /* keep thread on current cpu */
249 static int usched_dfly_weight2 = 180; /* synchronous peer's current cpu */
250 static int usched_dfly_weight3 = 40; /* number of threads on queue */
251 static int usched_dfly_weight4 = 160; /* availability of idle cores */
252 static int usched_dfly_features = 0x8F; /* allow pulls */
253 static int usched_dfly_fast_resched = 0;/* delta priority / resched */
254 static int usched_dfly_swmask = ~PPQMASK; /* allow pulls */
255 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
256 static int usched_dfly_decay = 8;
258 /* KTR debug printings */
260 KTR_INFO_MASTER(usched);
262 #if !defined(KTR_USCHED_DFLY)
263 #define KTR_USCHED_DFLY KTR_ALL
266 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
267 "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
268 pid_t pid, int old_cpuid, int curr);
271 * This function is called when the kernel intends to return to userland.
272 * It is responsible for making the thread the current designated userland
273 * thread for this cpu, blocking if necessary.
275 * The kernel will not depress our LWKT priority until after we return,
276 * in case we have to shove over to another cpu.
278 * We must determine our thread's disposition before we switch away. This
279 * is very sensitive code.
281 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
282 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will
283 * occur, this function is called only under very controlled circumstances.
286 dfly_acquire_curproc(struct lwp *lp)
295 * Make sure we aren't sitting on a tsleep queue.
298 crit_enter_quick(td);
299 if (td->td_flags & TDF_TSLEEPQ)
301 dfly_recalculate_estcpu(lp);
304 dd = &dfly_pcpu[gd->gd_cpuid];
307 * Process any pending interrupts/ipi's, then handle reschedule
308 * requests. dfly_release_curproc() will try to assign a new
309 * uschedcp that isn't us and otherwise NULL it out.
312 if ((td->td_mpflags & TDF_MP_BATCH_DEMARC) &&
313 lp->lwp_rrcount >= usched_dfly_rrinterval / 2) {
317 if (user_resched_wanted()) {
318 if (dd->uschedcp == lp)
320 clear_user_resched();
321 dfly_release_curproc(lp);
325 * Loop until we are the current user thread.
327 * NOTE: dd spinlock not held at top of loop.
329 if (dd->uschedcp == lp)
332 while (dd->uschedcp != lp) {
335 spin_lock(&dd->spin);
338 * We are not or are no longer the current lwp and a forced
339 * reschedule was requested. Figure out the best cpu to
340 * run on (our current cpu will be given significant weight).
342 * (if a reschedule was not requested we want to move this
343 * step after the uschedcp tests).
346 (usched_dfly_features & 0x08) &&
347 (rdd = dfly_choose_best_queue(lp)) != dd) {
348 dfly_changeqcpu_locked(lp, dd, rdd);
349 spin_unlock(&dd->spin);
350 lwkt_deschedule(lp->lwp_thread);
351 dfly_setrunqueue_dd(rdd, lp);
354 dd = &dfly_pcpu[gd->gd_cpuid];
359 * Either no reschedule was requested or the best queue was
360 * dd, and no current process has been selected. We can
361 * trivially become the current lwp on the current cpu.
363 if (dd->uschedcp == NULL) {
364 atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
366 dd->upri = lp->lwp_priority;
367 KKASSERT(lp->lwp_qcpu == dd->cpuid);
368 spin_unlock(&dd->spin);
373 * Can we steal the current designated user thread?
375 * If we do the other thread will stall when it tries to
376 * return to userland, possibly rescheduling elsewhere.
378 * It is important to do a masked test to avoid the edge
379 * case where two near-equal-priority threads are constantly
380 * interrupting each other.
382 * In the exact match case another thread has already gained
383 * uschedcp and lowered its priority, if we steal it the
384 * other thread will stay stuck on the LWKT runq and not
385 * push to another cpu. So don't steal on equal-priority even
386 * though it might appear to be more beneficial due to not
387 * having to switch back to the other thread's context.
389 * usched_dfly_fast_resched requires that two threads be
390 * significantly far apart in priority in order to interrupt.
392 * If better but not sufficiently far apart, the current
393 * uschedcp will be interrupted at the next scheduler clock.
396 (dd->upri & ~PPQMASK) >
397 (lp->lwp_priority & ~PPQMASK) + usched_dfly_fast_resched) {
399 dd->upri = lp->lwp_priority;
400 KKASSERT(lp->lwp_qcpu == dd->cpuid);
401 spin_unlock(&dd->spin);
405 * We are not the current lwp, figure out the best cpu
406 * to run on (our current cpu will be given significant
407 * weight). Loop on cpu change.
409 if ((usched_dfly_features & 0x02) &&
410 force_resched == 0 &&
411 (rdd = dfly_choose_best_queue(lp)) != dd) {
412 dfly_changeqcpu_locked(lp, dd, rdd);
413 spin_unlock(&dd->spin);
414 lwkt_deschedule(lp->lwp_thread);
415 dfly_setrunqueue_dd(rdd, lp);
418 dd = &dfly_pcpu[gd->gd_cpuid];
423 * We cannot become the current lwp, place the lp on the
424 * run-queue of this or another cpu and deschedule ourselves.
426 * When we are reactivated we will have another chance.
428 * Reload after a switch or setrunqueue/switch possibly
429 * moved us to another cpu.
431 spin_unlock(&dd->spin);
432 lwkt_deschedule(lp->lwp_thread);
433 dfly_setrunqueue_dd(dd, lp);
436 dd = &dfly_pcpu[gd->gd_cpuid];
440 * Make sure upri is synchronized, then yield to LWKT threads as
441 * needed before returning. This could result in another reschedule.
446 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
450 * DFLY_RELEASE_CURPROC
452 * This routine detaches the current thread from the userland scheduler,
453 * usually because the thread needs to run or block in the kernel (at
454 * kernel priority) for a while.
456 * This routine is also responsible for selecting a new thread to
457 * make the current thread.
459 * NOTE: This implementation differs from the dummy example in that
460 * dfly_select_curproc() is able to select the current process, whereas
461 * dummy_select_curproc() is not able to select the current process.
462 * This means we have to NULL out uschedcp.
464 * Additionally, note that we may already be on a run queue if releasing
465 * via the lwkt_switch() in dfly_setrunqueue().
468 dfly_release_curproc(struct lwp *lp)
470 globaldata_t gd = mycpu;
471 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
474 * Make sure td_wakefromcpu is defaulted. This will be overwritten
477 if (dd->uschedcp == lp) {
478 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
479 spin_lock(&dd->spin);
480 if (dd->uschedcp == lp) {
481 dd->uschedcp = NULL; /* don't let lp be selected */
482 dd->upri = PRIBASE_NULL;
483 atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
484 spin_unlock(&dd->spin);
485 dfly_select_curproc(gd);
487 spin_unlock(&dd->spin);
493 * DFLY_SELECT_CURPROC
495 * Select a new current process for this cpu and clear any pending user
496 * reschedule request. The cpu currently has no current process.
498 * This routine is also responsible for equal-priority round-robining,
499 * typically triggered from dfly_schedulerclock(). In our dummy example
500 * all the 'user' threads are LWKT scheduled all at once and we just
501 * call lwkt_switch().
503 * The calling process is not on the queue and cannot be selected.
507 dfly_select_curproc(globaldata_t gd)
509 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
511 int cpuid = gd->gd_cpuid;
515 spin_lock(&dd->spin);
516 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
519 atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
520 dd->upri = nlp->lwp_priority;
523 dd->rrcount = 0; /* reset round robin */
525 spin_unlock(&dd->spin);
526 lwkt_acquire(nlp->lwp_thread);
527 lwkt_schedule(nlp->lwp_thread);
529 spin_unlock(&dd->spin);
535 * Place the specified lwp on the user scheduler's run queue. This routine
536 * must be called with the thread descheduled. The lwp must be runnable.
537 * It must not be possible for anyone else to explicitly schedule this thread.
539 * The thread may be the current thread as a special case.
542 dfly_setrunqueue(struct lwp *lp)
548 * First validate the process LWKT state.
550 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
551 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
552 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
553 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
554 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
557 * NOTE: dd/rdd do not necessarily represent the current cpu.
558 * Instead they may represent the cpu the thread was last
559 * scheduled on or inherited by its parent.
561 dd = &dfly_pcpu[lp->lwp_qcpu];
565 * This process is not supposed to be scheduled anywhere or assigned
566 * as the current process anywhere. Assert the condition.
568 KKASSERT(rdd->uschedcp != lp);
571 * Ok, we have to setrunqueue some target cpu and request a reschedule
574 * We have to choose the best target cpu. It might not be the current
575 * target even if the current cpu has no running user thread (for
576 * example, because the current cpu might be a hyperthread and its
577 * sibling has a thread assigned).
579 * If we just forked it is most optimal to run the child on the same
580 * cpu just in case the parent decides to wait for it (thus getting
581 * off that cpu). As long as there is nothing else runnable on the
582 * cpu, that is. If we did this unconditionally a parent forking
583 * multiple children before waiting (e.g. make -j N) leaves other
584 * cpus idle that could be working.
586 if (lp->lwp_forked) {
588 if (usched_dfly_features & 0x20)
589 rdd = dfly_choose_best_queue(lp);
590 else if (usched_dfly_features & 0x40)
591 rdd = &dfly_pcpu[lp->lwp_qcpu];
592 else if (usched_dfly_features & 0x80)
593 rdd = dfly_choose_queue_simple(rdd, lp);
594 else if (dfly_pcpu[lp->lwp_qcpu].runqcount)
595 rdd = dfly_choose_best_queue(lp);
597 rdd = &dfly_pcpu[lp->lwp_qcpu];
599 rdd = dfly_choose_best_queue(lp);
600 /* rdd = &dfly_pcpu[lp->lwp_qcpu]; */
602 if (lp->lwp_qcpu != rdd->cpuid) {
603 spin_lock(&dd->spin);
604 dfly_changeqcpu_locked(lp, dd, rdd);
605 spin_unlock(&dd->spin);
607 dfly_setrunqueue_dd(rdd, lp);
611 * Change qcpu to rdd->cpuid. The dd the lp is CURRENTLY on must be
612 * spin-locked on-call. rdd does not have to be.
615 dfly_changeqcpu_locked(struct lwp *lp, dfly_pcpu_t dd, dfly_pcpu_t rdd)
617 if (lp->lwp_qcpu != rdd->cpuid) {
618 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
619 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
620 atomic_add_int(&dd->uload, -lp->lwp_uload);
621 atomic_add_int(&dd->ucount, -1);
622 atomic_add_int(&dfly_ucount, -1);
624 lp->lwp_qcpu = rdd->cpuid;
629 * Place lp on rdd's runqueue. Nothing is locked on call. This function
630 * also performs all necessary ancillary notification actions.
633 dfly_setrunqueue_dd(dfly_pcpu_t rdd, struct lwp *lp)
638 * We might be moving the lp to another cpu's run queue, and once
639 * on the runqueue (even if it is our cpu's), another cpu can rip
642 * TDF_MIGRATING might already be set if this is part of a
643 * remrunqueue+setrunqueue sequence.
645 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
646 lwkt_giveaway(lp->lwp_thread);
648 rgd = globaldata_find(rdd->cpuid);
651 * We lose control of the lp the moment we release the spinlock
652 * after having placed it on the queue. i.e. another cpu could pick
653 * it up, or it could exit, or its priority could be further
654 * adjusted, or something like that.
656 * WARNING! rdd can point to a foreign cpu!
658 spin_lock(&rdd->spin);
659 dfly_setrunqueue_locked(rdd, lp);
662 * Potentially interrupt the currently-running thread
664 if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK)) {
666 * Currently running thread is better or same, do not
669 spin_unlock(&rdd->spin);
670 } else if ((rdd->upri & ~PPQMASK) <= (lp->lwp_priority & ~PPQMASK) +
671 usched_dfly_fast_resched) {
673 * Currently running thread is not better, but not so bad
674 * that we need to interrupt it. Let it run for one more
678 rdd->uschedcp->lwp_rrcount < usched_dfly_rrinterval) {
679 rdd->uschedcp->lwp_rrcount = usched_dfly_rrinterval - 1;
681 spin_unlock(&rdd->spin);
682 } else if (rgd == mycpu) {
684 * We should interrupt the currently running thread, which
685 * is on the current cpu.
687 spin_unlock(&rdd->spin);
688 if (rdd->uschedcp == NULL) {
689 wakeup_mycpu(&rdd->helper_thread); /* XXX */
696 * We should interrupt the currently running thread, which
697 * is on a different cpu.
699 spin_unlock(&rdd->spin);
700 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote, NULL);
705 * This routine is called from a systimer IPI. It MUST be MP-safe and
706 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
711 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
713 globaldata_t gd = mycpu;
714 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
717 * Spinlocks also hold a critical section so there should not be
720 KKASSERT(gd->gd_spinlocks == 0);
726 * Do we need to round-robin? We round-robin 10 times a second.
727 * This should only occur for cpu-bound batch processes.
729 if (++lp->lwp_rrcount >= usched_dfly_rrinterval) {
730 lp->lwp_thread->td_wakefromcpu = -1;
735 * Adjust estcpu upward using a real time equivalent calculation,
736 * and recalculate lp's priority.
738 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
739 dfly_resetpriority(lp);
742 * Rebalance two cpus every 8 ticks, pulling the worst thread
743 * from the worst cpu's queue into a rotating cpu number.
745 * This mechanic is needed because the push algorithms can
746 * steady-state in an non-optimal configuration. We need to mix it
747 * up a little, even if it means breaking up a paired thread, so
748 * the push algorithms can rebalance the degenerate conditions.
749 * This portion of the algorithm exists to ensure stability at the
750 * selected weightings.
752 * Because we might be breaking up optimal conditions we do not want
753 * to execute this too quickly, hence we only rebalance approximately
754 * ~7-8 times per second. The push's, on the otherhand, are capable
755 * moving threads to other cpus at a much higher rate.
757 * We choose the most heavily loaded thread from the worst queue
758 * in order to ensure that multiple heavy-weight threads on the same
759 * queue get broken up, and also because these threads are the most
760 * likely to be able to remain in place. Hopefully then any pairings,
761 * if applicable, migrate to where these threads are.
763 if ((usched_dfly_features & 0x04) &&
764 ((u_int)sched_ticks & 7) == 0 &&
765 (u_int)sched_ticks / 8 % ncpus == gd->gd_cpuid) {
772 rdd = dfly_choose_worst_queue(dd);
774 spin_lock(&dd->spin);
775 if (spin_trylock(&rdd->spin)) {
776 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
777 spin_unlock(&rdd->spin);
779 spin_unlock(&dd->spin);
781 spin_unlock(&dd->spin);
787 /* dd->spin held if nlp != NULL */
790 * Either schedule it or add it to our queue.
793 (nlp->lwp_priority & ~PPQMASK) < (dd->upri & ~PPQMASK)) {
794 atomic_set_cpumask(&dfly_curprocmask, dd->cpumask);
795 dd->upri = nlp->lwp_priority;
798 dd->rrcount = 0; /* reset round robin */
800 spin_unlock(&dd->spin);
801 lwkt_acquire(nlp->lwp_thread);
802 lwkt_schedule(nlp->lwp_thread);
804 dfly_setrunqueue_locked(dd, nlp);
805 spin_unlock(&dd->spin);
811 * Called from acquire and from kern_synch's one-second timer (one of the
812 * callout helper threads) with a critical section held.
814 * Adjust p_estcpu based on our single-cpu load, p_nice, and compensate for
815 * overall system load.
817 * Note that no recalculation occurs for a process which sleeps and wakes
818 * up in the same tick. That is, a system doing thousands of context
819 * switches per second will still only do serious estcpu calculations
820 * ESTCPUFREQ times per second.
824 dfly_recalculate_estcpu(struct lwp *lp)
826 globaldata_t gd = mycpu;
834 * We have to subtract periodic to get the last schedclock
835 * timeout time, otherwise we would get the upcoming timeout.
836 * Keep in mind that a process can migrate between cpus and
837 * while the scheduler clock should be very close, boundary
838 * conditions could lead to a small negative delta.
840 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
842 if (lp->lwp_slptime > 1) {
844 * Too much time has passed, do a coarse correction.
846 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
847 dfly_resetpriority(lp);
848 lp->lwp_cpbase = cpbase;
851 } else if (lp->lwp_cpbase != cpbase) {
853 * Adjust estcpu if we are in a different tick. Don't waste
854 * time if we are in the same tick.
856 * First calculate the number of ticks in the measurement
857 * interval. The ttlticks calculation can wind up 0 due to
858 * a bug in the handling of lwp_slptime (as yet not found),
859 * so make sure we do not get a divide by 0 panic.
861 ttlticks = (cpbase - lp->lwp_cpbase) /
862 gd->gd_schedclock.periodic;
863 if ((ssysclock_t)ttlticks < 0) {
865 lp->lwp_cpbase = cpbase;
869 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
872 * Calculate the percentage of one cpu being used then
873 * compensate for any system load in excess of ncpus.
875 * For example, if we have 8 cores and 16 running cpu-bound
876 * processes then all things being equal each process will
877 * get 50% of one cpu. We need to pump this value back
878 * up to 100% so the estcpu calculation properly adjusts
879 * the process's dynamic priority.
881 * estcpu is scaled by ESTCPUMAX, pctcpu is scaled by FSCALE.
883 estcpu = (lp->lwp_pctcpu * ESTCPUMAX) >> FSHIFT;
884 ucount = dfly_ucount;
885 if (ucount > ncpus) {
886 estcpu += estcpu * (ucount - ncpus) / ncpus;
889 if (usched_dfly_debug == lp->lwp_proc->p_pid) {
890 kprintf("pid %d lwp %p estcpu %3d %3d cp %d/%d",
891 lp->lwp_proc->p_pid, lp,
892 estcpu, lp->lwp_estcpu,
893 lp->lwp_cpticks, ttlticks);
897 * Adjust lp->lwp_esetcpu. The decay factor determines how
898 * quickly lwp_estcpu collapses to its realtime calculation.
899 * A slower collapse gives us a more accurate number over
900 * the long term but can create problems with bursty threads
901 * or threads which become cpu hogs.
903 * To solve this problem, newly started lwps and lwps which
904 * are restarting after having been asleep for a while are
905 * given a much, much faster decay in order to quickly
906 * detect whether they become cpu-bound.
908 * NOTE: p_nice is accounted for in dfly_resetpriority(),
909 * and not here, but we must still ensure that a
910 * cpu-bound nice -20 process does not completely
911 * override a cpu-bound nice +20 process.
913 * NOTE: We must use ESTCPULIM() here to deal with any
916 decay_factor = usched_dfly_decay;
917 if (decay_factor < 1)
919 if (decay_factor > 1024)
922 if (lp->lwp_estfast < usched_dfly_decay) {
924 lp->lwp_estcpu = ESTCPULIM(
925 (lp->lwp_estcpu * lp->lwp_estfast + estcpu) /
926 (lp->lwp_estfast + 1));
928 lp->lwp_estcpu = ESTCPULIM(
929 (lp->lwp_estcpu * decay_factor + estcpu) /
933 if (usched_dfly_debug == lp->lwp_proc->p_pid)
934 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
935 dfly_resetpriority(lp);
936 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
942 * Compute the priority of a process when running in user mode.
943 * Arrange to reschedule if the resulting priority is better
944 * than that of the current process.
946 * This routine may be called with any process.
948 * This routine is called by fork1() for initial setup with the process
949 * of the run queue, and also may be called normally with the process on or
953 dfly_resetpriority(struct lwp *lp)
966 * Lock the scheduler (lp) belongs to. This can be on a different
967 * cpu. Handle races. This loop breaks out with the appropriate
973 rdd = &dfly_pcpu[rcpu];
974 spin_lock(&rdd->spin);
975 if (rcpu == lp->lwp_qcpu)
977 spin_unlock(&rdd->spin);
981 * Calculate the new priority and queue type
983 newrqtype = lp->lwp_rtprio.type;
986 case RTP_PRIO_REALTIME:
988 newpriority = PRIBASE_REALTIME +
989 (lp->lwp_rtprio.prio & PRIMASK);
991 case RTP_PRIO_NORMAL:
995 estcpu = lp->lwp_estcpu;
998 * p_nice piece Adds (0-40) * 2 0-80
999 * estcpu Adds 16384 * 4 / 512 0-128
1001 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
1002 newpriority += estcpu * PPQ / ESTCPUPPQ;
1003 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
1004 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
1005 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
1008 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
1010 case RTP_PRIO_THREAD:
1011 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
1014 panic("Bad RTP_PRIO %d", newrqtype);
1019 * The LWKT scheduler doesn't dive usched structures, give it a hint
1020 * on the relative priority of user threads running in the kernel.
1021 * The LWKT scheduler will always ensure that a user thread running
1022 * in the kernel will get cpu some time, regardless of its upri,
1023 * but can decide not to instantly switch from one kernel or user
1024 * mode user thread to a kernel-mode user thread when it has a less
1025 * desireable user priority.
1027 * td_upri has normal sense (higher values are more desireable), so
1030 lp->lwp_thread->td_upri = -(newpriority & usched_dfly_swmask);
1033 * The newpriority incorporates the queue type so do a simple masked
1034 * check to determine if the process has moved to another queue. If
1035 * it has, and it is currently on a run queue, then move it.
1037 * Since uload is ~PPQMASK masked, no modifications are necessary if
1038 * we end up in the same run queue.
1040 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
1041 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
1042 dfly_remrunqueue_locked(rdd, lp);
1043 lp->lwp_priority = newpriority;
1044 lp->lwp_rqtype = newrqtype;
1045 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1046 dfly_setrunqueue_locked(rdd, lp);
1049 lp->lwp_priority = newpriority;
1050 lp->lwp_rqtype = newrqtype;
1051 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
1056 * In the same PPQ, uload cannot change.
1058 lp->lwp_priority = newpriority;
1064 * Adjust effective load.
1066 * Calculate load then scale up or down geometrically based on p_nice.
1067 * Processes niced up (positive) are less important, and processes
1068 * niced downard (negative) are more important. The higher the uload,
1069 * the more important the thread.
1071 /* 0-511, 0-100% cpu */
1072 delta_uload = lp->lwp_estcpu / NQS;
1073 delta_uload -= delta_uload * lp->lwp_proc->p_nice / (PRIO_MAX + 1);
1076 delta_uload -= lp->lwp_uload;
1077 lp->lwp_uload += delta_uload;
1078 if (lp->lwp_mpflags & LWP_MP_ULOAD)
1079 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, delta_uload);
1082 * Determine if we need to reschedule the target cpu. This only
1083 * occurs if the LWP is already on a scheduler queue, which means
1084 * that idle cpu notification has already occured. At most we
1085 * need only issue a need_user_resched() on the appropriate cpu.
1087 * The LWP may be owned by a CPU different from the current one,
1088 * in which case dd->uschedcp may be modified without an MP lock
1089 * or a spinlock held. The worst that happens is that the code
1090 * below causes a spurious need_user_resched() on the target CPU
1091 * and dd->pri to be wrong for a short period of time, both of
1092 * which are harmless.
1094 * If checkpri is 0 we are adjusting the priority of the current
1095 * process, possibly higher (less desireable), so ignore the upri
1096 * check which will fail in that case.
1099 if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
1101 (rdd->upri & ~PRIMASK) >
1102 (lp->lwp_priority & ~PRIMASK))) {
1103 if (rcpu == mycpu->gd_cpuid) {
1104 spin_unlock(&rdd->spin);
1105 need_user_resched();
1107 spin_unlock(&rdd->spin);
1108 lwkt_send_ipiq(globaldata_find(rcpu),
1109 dfly_need_user_resched_remote,
1113 spin_unlock(&rdd->spin);
1116 spin_unlock(&rdd->spin);
1123 dfly_yield(struct lwp *lp)
1126 /* FUTURE (or something similar) */
1127 switch(lp->lwp_rqtype) {
1128 case RTP_PRIO_NORMAL:
1129 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
1135 need_user_resched();
1139 * Called from fork1() when a new child process is being created.
1141 * Give the child process an initial estcpu that is more batch then
1142 * its parent and dock the parent for the fork (but do not
1143 * reschedule the parent).
1147 * XXX lwp should be "spawning" instead of "forking"
1150 dfly_forking(struct lwp *plp, struct lwp *lp)
1153 * Put the child 4 queue slots (out of 32) higher than the parent
1154 * (less desireable than the parent).
1156 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1158 lp->lwp_estfast = 0;
1161 * Dock the parent a cost for the fork, protecting us from fork
1162 * bombs. If the parent is forking quickly make the child more
1165 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1169 * Called when a lwp is being removed from this scheduler, typically
1170 * during lwp_exit(). We have to clean out any ULOAD accounting before
1171 * we can let the lp go. The dd->spin lock is not needed for uload
1174 * Scheduler dequeueing has already occurred, no further action in that
1178 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1180 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1182 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1183 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1184 atomic_add_int(&dd->uload, -lp->lwp_uload);
1185 atomic_add_int(&dd->ucount, -1);
1186 atomic_add_int(&dfly_ucount, -1);
1191 * This function cannot block in any way, but spinlocks are ok.
1193 * Update the uload based on the state of the thread (whether it is going
1194 * to sleep or running again). The uload is meant to be a longer-term
1195 * load and not an instantanious load.
1198 dfly_uload_update(struct lwp *lp)
1200 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1202 if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1203 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1204 spin_lock(&dd->spin);
1205 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1206 atomic_set_int(&lp->lwp_mpflags,
1208 atomic_add_int(&dd->uload, lp->lwp_uload);
1209 atomic_add_int(&dd->ucount, 1);
1210 atomic_add_int(&dfly_ucount, 1);
1212 spin_unlock(&dd->spin);
1214 } else if (lp->lwp_slptime > 0) {
1215 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1216 spin_lock(&dd->spin);
1217 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1218 atomic_clear_int(&lp->lwp_mpflags,
1220 atomic_add_int(&dd->uload, -lp->lwp_uload);
1221 atomic_add_int(&dd->ucount, -1);
1222 atomic_add_int(&dfly_ucount, -1);
1224 spin_unlock(&dd->spin);
1230 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1231 * it selects a user process and returns it. If chklp is non-NULL and chklp
1232 * has a better or equal priority then the process that would otherwise be
1233 * chosen, NULL is returned.
1235 * Until we fix the RUNQ code the chklp test has to be strict or we may
1236 * bounce between processes trying to acquire the current process designation.
1238 * Must be called with rdd->spin locked. The spinlock is left intact through
1239 * the entire routine. dd->spin does not have to be locked.
1241 * If worst is non-zero this function finds the worst thread instead of the
1242 * best thread (used by the schedulerclock-based rover).
1246 dfly_chooseproc_locked(dfly_pcpu_t rdd, dfly_pcpu_t dd,
1247 struct lwp *chklp, int worst)
1257 rtqbits = rdd->rtqueuebits;
1258 tsqbits = rdd->queuebits;
1259 idqbits = rdd->idqueuebits;
1263 pri = bsrl(idqbits);
1264 q = &rdd->idqueues[pri];
1265 which = &rdd->idqueuebits;
1266 } else if (tsqbits) {
1267 pri = bsrl(tsqbits);
1268 q = &rdd->queues[pri];
1269 which = &rdd->queuebits;
1270 } else if (rtqbits) {
1271 pri = bsrl(rtqbits);
1272 q = &rdd->rtqueues[pri];
1273 which = &rdd->rtqueuebits;
1277 lp = TAILQ_LAST(q, rq);
1280 pri = bsfl(rtqbits);
1281 q = &rdd->rtqueues[pri];
1282 which = &rdd->rtqueuebits;
1283 } else if (tsqbits) {
1284 pri = bsfl(tsqbits);
1285 q = &rdd->queues[pri];
1286 which = &rdd->queuebits;
1287 } else if (idqbits) {
1288 pri = bsfl(idqbits);
1289 q = &rdd->idqueues[pri];
1290 which = &rdd->idqueuebits;
1294 lp = TAILQ_FIRST(q);
1296 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1299 * If the passed lwp <chklp> is reasonably close to the selected
1300 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1302 * Note that we must error on the side of <chklp> to avoid bouncing
1303 * between threads in the acquire code.
1306 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1310 KTR_COND_LOG(usched_chooseproc,
1311 lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1312 lp->lwp_proc->p_pid,
1313 lp->lwp_thread->td_gd->gd_cpuid,
1316 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1317 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1318 TAILQ_REMOVE(q, lp, lwp_procq);
1321 *which &= ~(1 << pri);
1324 * If we are choosing a process from rdd with the intent to
1325 * move it to dd, lwp_qcpu must be adjusted while rdd's spinlock
1329 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1330 atomic_add_int(&rdd->uload, -lp->lwp_uload);
1331 atomic_add_int(&rdd->ucount, -1);
1332 atomic_add_int(&dfly_ucount, -1);
1334 lp->lwp_qcpu = dd->cpuid;
1335 atomic_add_int(&dd->uload, lp->lwp_uload);
1336 atomic_add_int(&dd->ucount, 1);
1337 atomic_add_int(&dfly_ucount, 1);
1338 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1344 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1346 * Choose a cpu node to schedule lp on, hopefully nearby its current
1349 * We give the current node a modest advantage for obvious reasons.
1351 * We also give the node the thread was woken up FROM a slight advantage
1352 * in order to try to schedule paired threads which synchronize/block waiting
1353 * for each other fairly close to each other. Similarly in a network setting
1354 * this feature will also attempt to place a user process near the kernel
1355 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the
1356 * algorithm as it heuristically groups synchronizing processes for locality
1357 * of reference in multi-socket systems.
1359 * We check against running processes and give a big advantage if there
1362 * The caller will normally dfly_setrunqueue() lp on the returned queue.
1364 * When the topology is known choose a cpu whos group has, in aggregate,
1365 * has the lowest weighted load.
1369 dfly_choose_best_queue(struct lwp *lp)
1376 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1386 * When the topology is unknown choose a random cpu that is hopefully
1389 if (dd->cpunode == NULL)
1390 return (dfly_choose_queue_simple(dd, lp));
1395 if ((wakecpu = lp->lwp_thread->td_wakefromcpu) >= 0)
1396 wakemask = dfly_pcpu[wakecpu].cpumask;
1401 * When the topology is known choose a cpu whos group has, in
1402 * aggregate, has the lowest weighted load.
1404 cpup = root_cpu_node;
1409 * Degenerate case super-root
1411 if (cpup->child_node && cpup->child_no == 1) {
1412 cpup = cpup->child_node;
1419 if (cpup->child_node == NULL) {
1420 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1425 lowest_load = 0x7FFFFFFF;
1427 for (n = 0; n < cpup->child_no; ++n) {
1429 * Accumulate load information for all cpus
1430 * which are members of this node.
1432 cpun = &cpup->child_node[n];
1433 mask = cpun->members & usched_global_cpumask &
1434 smp_active_mask & lp->lwp_cpumask;
1442 cpuid = BSFCPUMASK(mask);
1443 rdd = &dfly_pcpu[cpuid];
1445 load += rdd->ucount * usched_dfly_weight3;
1447 if (rdd->uschedcp == NULL &&
1448 rdd->runqcount == 0 &&
1449 globaldata_find(cpuid)->gd_tdrunqcount == 0
1451 load -= usched_dfly_weight4;
1454 else if (rdd->upri > lp->lwp_priority + PPQ) {
1455 load -= usched_dfly_weight4 / 2;
1458 mask &= ~CPUMASK(cpuid);
1463 * Compensate if the lp is already accounted for in
1464 * the aggregate uload for this mask set. We want
1465 * to calculate the loads as if lp were not present,
1466 * otherwise the calculation is bogus.
1468 if ((lp->lwp_mpflags & LWP_MP_ULOAD) &&
1469 (dd->cpumask & cpun->members)) {
1470 load -= lp->lwp_uload;
1471 load -= usched_dfly_weight3;
1477 * Advantage the cpu group (lp) is already on.
1479 if (cpun->members & dd->cpumask)
1480 load -= usched_dfly_weight1;
1483 * Advantage the cpu group we want to pair (lp) to,
1484 * but don't let it go to the exact same cpu as
1485 * the wakecpu target.
1487 * We do this by checking whether cpun is a
1488 * terminal node or not. All cpun's at the same
1489 * level will either all be terminal or all not
1492 * If it is and we match we disadvantage the load.
1493 * If it is and we don't match we advantage the load.
1495 * Also note that we are effectively disadvantaging
1496 * all-but-one by the same amount, so it won't effect
1497 * the weight1 factor for the all-but-one nodes.
1499 if (cpun->members & wakemask) {
1500 if (cpun->child_node != NULL) {
1502 load -= usched_dfly_weight2;
1504 if (usched_dfly_features & 0x10)
1505 load += usched_dfly_weight2;
1507 load -= usched_dfly_weight2;
1512 * Calculate the best load
1514 if (cpub == NULL || lowest_load > load ||
1515 (lowest_load == load &&
1516 (cpun->members & dd->cpumask))
1524 if (usched_dfly_chooser)
1525 kprintf("lp %02d->%02d %s\n",
1526 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1531 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1533 * Choose the worst queue close to dd's cpu node with a non-empty runq
1534 * that is NOT dd. Also require that the moving of the highest-load thread
1535 * from rdd to dd does not cause the uload's to cross each other.
1537 * This is used by the thread chooser when the current cpu's queues are
1538 * empty to steal a thread from another cpu's queue. We want to offload
1539 * the most heavily-loaded queue.
1543 dfly_choose_worst_queue(dfly_pcpu_t dd)
1561 * When the topology is unknown choose a random cpu that is hopefully
1564 if (dd->cpunode == NULL) {
1569 * When the topology is known choose a cpu whos group has, in
1570 * aggregate, has the lowest weighted load.
1572 cpup = root_cpu_node;
1576 * Degenerate case super-root
1578 if (cpup->child_node && cpup->child_no == 1) {
1579 cpup = cpup->child_node;
1586 if (cpup->child_node == NULL) {
1587 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1594 for (n = 0; n < cpup->child_no; ++n) {
1596 * Accumulate load information for all cpus
1597 * which are members of this node.
1599 cpun = &cpup->child_node[n];
1600 mask = cpun->members & usched_global_cpumask &
1608 cpuid = BSFCPUMASK(mask);
1609 rdd = &dfly_pcpu[cpuid];
1611 load += rdd->ucount * usched_dfly_weight3;
1612 if (rdd->uschedcp == NULL &&
1613 rdd->runqcount == 0 &&
1614 globaldata_find(cpuid)->gd_tdrunqcount == 0
1616 load -= usched_dfly_weight4;
1619 else if (rdd->upri > dd->upri + PPQ) {
1620 load -= usched_dfly_weight4 / 2;
1623 mask &= ~CPUMASK(cpuid);
1629 * Prefer candidates which are somewhat closer to
1632 if (dd->cpumask & cpun->members)
1633 load += usched_dfly_weight1;
1636 * The best candidate is the one with the worst
1639 if (cpub == NULL || highest_load < load) {
1640 highest_load = load;
1648 * We never return our own node (dd), and only return a remote
1649 * node if it's load is significantly worse than ours (i.e. where
1650 * stealing a thread would be considered reasonable).
1652 * This also helps us avoid breaking paired threads apart which
1653 * can have disastrous effects on performance.
1660 if (rdd->rtqueuebits && hpri < (pri = bsrl(rdd->rtqueuebits)))
1662 if (rdd->queuebits && hpri < (pri = bsrl(rdd->queuebits)))
1664 if (rdd->idqueuebits && hpri < (pri = bsrl(rdd->idqueuebits)))
1667 if (rdd->uload - hpri < dd->uload + hpri)
1675 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1683 * Fallback to the original heuristic, select random cpu,
1684 * first checking cpus not currently running a user thread.
1687 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1688 mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1689 smp_active_mask & usched_global_cpumask;
1692 tmpmask = ~(CPUMASK(cpuid) - 1);
1694 cpuid = BSFCPUMASK(mask & tmpmask);
1696 cpuid = BSFCPUMASK(mask);
1697 rdd = &dfly_pcpu[cpuid];
1699 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1701 mask &= ~CPUMASK(cpuid);
1705 * Then cpus which might have a currently running lp
1707 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1708 mask = dfly_curprocmask & dfly_rdyprocmask &
1709 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1712 tmpmask = ~(CPUMASK(cpuid) - 1);
1714 cpuid = BSFCPUMASK(mask & tmpmask);
1716 cpuid = BSFCPUMASK(mask);
1717 rdd = &dfly_pcpu[cpuid];
1719 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1721 mask &= ~CPUMASK(cpuid);
1725 * If we cannot find a suitable cpu we reload from dfly_scancpu
1726 * and round-robin. Other cpus will pickup as they release their
1727 * current lwps or become ready.
1729 * Avoid a degenerate system lockup case if usched_global_cpumask
1730 * is set to 0 or otherwise does not cover lwp_cpumask.
1732 * We only kick the target helper thread in this case, we do not
1733 * set the user resched flag because
1735 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1736 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1738 rdd = &dfly_pcpu[cpuid];
1745 dfly_need_user_resched_remote(void *dummy)
1747 globaldata_t gd = mycpu;
1748 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
1751 * Flag reschedule needed
1753 need_user_resched();
1756 * If no user thread is currently running we need to kick the helper
1757 * on our cpu to recover. Otherwise the cpu will never schedule
1760 * We cannot schedule the process ourselves because this is an
1761 * IPI callback and we cannot acquire spinlocks in an IPI callback.
1763 * Call wakeup_mycpu to avoid sending IPIs to other CPUs
1765 if (dd->uschedcp == NULL && (dfly_rdyprocmask & gd->gd_cpumask)) {
1766 atomic_clear_cpumask(&dfly_rdyprocmask, gd->gd_cpumask);
1767 wakeup_mycpu(&dd->helper_thread);
1772 * dfly_remrunqueue_locked() removes a given process from the run queue
1773 * that it is on, clearing the queue busy bit if it becomes empty.
1775 * Note that user process scheduler is different from the LWKT schedule.
1776 * The user process scheduler only manages user processes but it uses LWKT
1777 * underneath, and a user process operating in the kernel will often be
1778 * 'released' from our management.
1780 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes
1781 * to sleep or the lwp is moved to a different runq.
1784 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1790 KKASSERT(rdd->runqcount >= 0);
1792 pri = lp->lwp_rqindex;
1794 switch(lp->lwp_rqtype) {
1795 case RTP_PRIO_NORMAL:
1796 q = &rdd->queues[pri];
1797 which = &rdd->queuebits;
1799 case RTP_PRIO_REALTIME:
1801 q = &rdd->rtqueues[pri];
1802 which = &rdd->rtqueuebits;
1805 q = &rdd->idqueues[pri];
1806 which = &rdd->idqueuebits;
1809 panic("remrunqueue: invalid rtprio type");
1812 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1813 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1814 TAILQ_REMOVE(q, lp, lwp_procq);
1816 if (TAILQ_EMPTY(q)) {
1817 KASSERT((*which & (1 << pri)) != 0,
1818 ("remrunqueue: remove from empty queue"));
1819 *which &= ~(1 << pri);
1824 * dfly_setrunqueue_locked()
1826 * Add a process whos rqtype and rqindex had previously been calculated
1827 * onto the appropriate run queue. Determine if the addition requires
1828 * a reschedule on a cpu and return the cpuid or -1.
1830 * NOTE: Lower priorities are better priorities.
1832 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1833 * sum of the rough lwp_priority for all running and runnable
1834 * processes. Lower priority processes (higher lwp_priority
1835 * values) actually DO count as more load, not less, because
1836 * these are the programs which require the most care with
1837 * regards to cpu selection.
1840 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1846 KKASSERT(lp->lwp_qcpu == rdd->cpuid);
1848 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1849 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1850 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload, lp->lwp_uload);
1851 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].ucount, 1);
1852 atomic_add_int(&dfly_ucount, 1);
1855 pri = lp->lwp_rqindex;
1857 switch(lp->lwp_rqtype) {
1858 case RTP_PRIO_NORMAL:
1859 q = &rdd->queues[pri];
1860 which = &rdd->queuebits;
1862 case RTP_PRIO_REALTIME:
1864 q = &rdd->rtqueues[pri];
1865 which = &rdd->rtqueuebits;
1868 q = &rdd->idqueues[pri];
1869 which = &rdd->idqueuebits;
1872 panic("remrunqueue: invalid rtprio type");
1877 * Place us on the selected queue. Determine if we should be
1878 * placed at the head of the queue or at the end.
1880 * We are placed at the tail if our round-robin count has expired,
1881 * or is about to expire and the system thinks its a good place to
1882 * round-robin, or there is already a next thread on the queue
1883 * (it might be trying to pick up where it left off and we don't
1884 * want to interfere).
1886 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1887 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1890 if (lp->lwp_rrcount >= usched_dfly_rrinterval ||
1891 (lp->lwp_rrcount >= usched_dfly_rrinterval / 2 &&
1892 (lp->lwp_thread->td_mpflags & TDF_MP_BATCH_DEMARC)) ||
1895 atomic_clear_int(&lp->lwp_thread->td_mpflags,
1896 TDF_MP_BATCH_DEMARC);
1897 lp->lwp_rrcount = 0;
1898 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1901 lp->lwp_rrcount = 0;
1902 TAILQ_INSERT_HEAD(q, lp, lwp_procq);
1908 * For SMP systems a user scheduler helper thread is created for each
1909 * cpu and is used to allow one cpu to wakeup another for the purposes of
1910 * scheduling userland threads from setrunqueue().
1912 * UP systems do not need the helper since there is only one cpu.
1914 * We can't use the idle thread for this because we might block.
1915 * Additionally, doing things this way allows us to HLT idle cpus
1919 dfly_helper_thread(void *dummy)
1929 cpuid = gd->gd_cpuid; /* doesn't change */
1930 mask = gd->gd_cpumask; /* doesn't change */
1931 dd = &dfly_pcpu[cpuid];
1934 * Since we only want to be woken up only when no user processes
1935 * are scheduled on a cpu, run at an ultra low priority.
1937 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1939 tsleep(&dd->helper_thread, 0, "schslp", 0);
1943 * We use the LWKT deschedule-interlock trick to avoid racing
1944 * dfly_rdyprocmask. This means we cannot block through to the
1945 * manual lwkt_switch() call we make below.
1948 tsleep_interlock(&dd->helper_thread, 0);
1950 spin_lock(&dd->spin);
1952 atomic_set_cpumask(&dfly_rdyprocmask, mask);
1953 clear_user_resched(); /* This satisfied the reschedule request */
1955 dd->rrcount = 0; /* Reset the round-robin counter */
1958 if (dd->runqcount || dd->uschedcp != NULL) {
1960 * Threads are available. A thread may or may not be
1961 * currently scheduled. Get the best thread already queued
1964 nlp = dfly_chooseproc_locked(dd, dd, dd->uschedcp, 0);
1966 atomic_set_cpumask(&dfly_curprocmask, mask);
1967 dd->upri = nlp->lwp_priority;
1970 dd->rrcount = 0; /* reset round robin */
1972 spin_unlock(&dd->spin);
1973 lwkt_acquire(nlp->lwp_thread);
1974 lwkt_schedule(nlp->lwp_thread);
1977 * This situation should not occur because we had
1978 * at least one thread available.
1980 spin_unlock(&dd->spin);
1982 } else if (usched_dfly_features & 0x01) {
1984 * This cpu is devoid of runnable threads, steal a thread
1985 * from another cpu. Since we're stealing, might as well
1986 * load balance at the same time.
1988 * We choose the highest-loaded thread from the worst queue.
1990 * NOTE! This function only returns a non-NULL rdd when
1991 * another cpu's queue is obviously overloaded. We
1992 * do not want to perform the type of rebalancing
1993 * the schedclock does here because it would result
1994 * in insane process pulling when 'steady' state is
1995 * partially unbalanced (e.g. 6 runnables and only
1998 rdd = dfly_choose_worst_queue(dd);
1999 if (rdd && spin_trylock(&rdd->spin)) {
2000 nlp = dfly_chooseproc_locked(rdd, dd, NULL, 1);
2001 spin_unlock(&rdd->spin);
2006 atomic_set_cpumask(&dfly_curprocmask, mask);
2007 dd->upri = nlp->lwp_priority;
2010 dd->rrcount = 0; /* reset round robin */
2012 spin_unlock(&dd->spin);
2013 lwkt_acquire(nlp->lwp_thread);
2014 lwkt_schedule(nlp->lwp_thread);
2017 * Leave the thread on our run queue. Another
2018 * scheduler will try to pull it later.
2020 spin_unlock(&dd->spin);
2024 * devoid of runnable threads and not allowed to steal
2027 spin_unlock(&dd->spin);
2031 * We're descheduled unless someone scheduled us. Switch away.
2032 * Exiting the critical section will cause splz() to be called
2033 * for us if interrupts and such are pending.
2036 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
2042 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
2046 new_val = usched_dfly_stick_to_level;
2048 error = sysctl_handle_int(oidp, &new_val, 0, req);
2049 if (error != 0 || req->newptr == NULL)
2051 if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
2053 usched_dfly_stick_to_level = new_val;
2059 * Setup the queues and scheduler helpers (scheduler helpers are SMP only).
2060 * Note that curprocmask bit 0 has already been cleared by rqinit() and
2061 * we should not mess with it further.
2064 usched_dfly_cpu_init(void)
2069 int smt_not_supported = 0;
2070 int cache_coherent_not_supported = 0;
2073 kprintf("Start scheduler helpers on cpus:\n");
2075 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
2076 usched_dfly_sysctl_tree =
2077 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
2078 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
2079 "usched_dfly", CTLFLAG_RD, 0, "");
2081 for (i = 0; i < ncpus; ++i) {
2082 dfly_pcpu_t dd = &dfly_pcpu[i];
2083 cpumask_t mask = CPUMASK(i);
2085 if ((mask & smp_active_mask) == 0)
2088 spin_init(&dd->spin);
2089 dd->cpunode = get_cpu_node_by_cpuid(i);
2091 dd->cpumask = CPUMASK(i);
2092 for (j = 0; j < NQS; j++) {
2093 TAILQ_INIT(&dd->queues[j]);
2094 TAILQ_INIT(&dd->rtqueues[j]);
2095 TAILQ_INIT(&dd->idqueues[j]);
2097 atomic_clear_cpumask(&dfly_curprocmask, 1);
2099 if (dd->cpunode == NULL) {
2100 smt_not_supported = 1;
2101 cache_coherent_not_supported = 1;
2103 kprintf ("\tcpu%d - WARNING: No CPU NODE "
2104 "found for cpu\n", i);
2106 switch (dd->cpunode->type) {
2109 kprintf ("\tcpu%d - HyperThreading "
2110 "available. Core siblings: ",
2114 smt_not_supported = 1;
2117 kprintf ("\tcpu%d - No HT available, "
2118 "multi-core/physical "
2119 "cpu. Physical siblings: ",
2123 smt_not_supported = 1;
2126 kprintf ("\tcpu%d - No HT available, "
2127 "single-core/physical cpu. "
2128 "Package Siblings: ",
2132 /* Let's go for safe defaults here */
2133 smt_not_supported = 1;
2134 cache_coherent_not_supported = 1;
2136 kprintf ("\tcpu%d - Unknown cpunode->"
2137 "type=%u. Siblings: ",
2139 (u_int)dd->cpunode->type);
2144 if (dd->cpunode->parent_node != NULL) {
2145 CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
2146 kprintf("cpu%d ", cpuid);
2149 kprintf(" no siblings\n");
2154 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
2155 0, i, "usched %d", i);
2158 * Allow user scheduling on the target cpu. cpu #0 has already
2159 * been enabled in rqinit().
2162 atomic_clear_cpumask(&dfly_curprocmask, mask);
2163 atomic_set_cpumask(&dfly_rdyprocmask, mask);
2164 dd->upri = PRIBASE_NULL;
2168 /* usched_dfly sysctl configurable parameters */
2170 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2171 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2172 OID_AUTO, "rrinterval", CTLFLAG_RW,
2173 &usched_dfly_rrinterval, 0, "");
2174 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2175 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2176 OID_AUTO, "decay", CTLFLAG_RW,
2177 &usched_dfly_decay, 0, "Extra decay when not running");
2179 /* Add enable/disable option for SMT scheduling if supported */
2180 if (smt_not_supported) {
2181 usched_dfly_smt = 0;
2182 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2183 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2184 OID_AUTO, "smt", CTLFLAG_RD,
2185 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
2187 usched_dfly_smt = 1;
2188 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2189 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2190 OID_AUTO, "smt", CTLFLAG_RW,
2191 &usched_dfly_smt, 0, "Enable SMT scheduling");
2195 * Add enable/disable option for cache coherent scheduling
2198 if (cache_coherent_not_supported) {
2199 usched_dfly_cache_coherent = 0;
2200 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
2201 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2202 OID_AUTO, "cache_coherent", CTLFLAG_RD,
2204 "Cache coherence NOT SUPPORTED");
2206 usched_dfly_cache_coherent = 1;
2207 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2208 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2209 OID_AUTO, "cache_coherent", CTLFLAG_RW,
2210 &usched_dfly_cache_coherent, 0,
2211 "Enable/Disable cache coherent scheduling");
2213 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2214 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2215 OID_AUTO, "weight1", CTLFLAG_RW,
2216 &usched_dfly_weight1, 200,
2217 "Weight selection for current cpu");
2219 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2220 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2221 OID_AUTO, "weight2", CTLFLAG_RW,
2222 &usched_dfly_weight2, 180,
2223 "Weight selection for wakefrom cpu");
2225 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2226 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2227 OID_AUTO, "weight3", CTLFLAG_RW,
2228 &usched_dfly_weight3, 40,
2229 "Weight selection for num threads on queue");
2231 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2232 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2233 OID_AUTO, "weight4", CTLFLAG_RW,
2234 &usched_dfly_weight4, 160,
2235 "Availability of other idle cpus");
2237 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2238 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2239 OID_AUTO, "fast_resched", CTLFLAG_RW,
2240 &usched_dfly_fast_resched, 0,
2241 "Availability of other idle cpus");
2243 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2244 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2245 OID_AUTO, "features", CTLFLAG_RW,
2246 &usched_dfly_features, 0x8F,
2247 "Allow pulls into empty queues");
2249 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
2250 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2251 OID_AUTO, "swmask", CTLFLAG_RW,
2252 &usched_dfly_swmask, ~PPQMASK,
2253 "Queue mask to force thread switch");
2256 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
2257 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
2258 OID_AUTO, "stick_to_level",
2259 CTLTYPE_INT | CTLFLAG_RW,
2260 NULL, sizeof usched_dfly_stick_to_level,
2261 sysctl_usched_dfly_stick_to_level, "I",
2262 "Stick a process to this level. See sysctl"
2263 "paremter hw.cpu_topology.level_description");
2267 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
2268 usched_dfly_cpu_init, NULL)