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_rqindex lwp_usdata.dfly.rqindex
94 #define lwp_estcpu lwp_usdata.dfly.estcpu
95 #define lwp_batch lwp_usdata.dfly.batch
96 #define lwp_rqtype lwp_usdata.dfly.rqtype
97 #define lwp_qcpu lwp_usdata.dfly.qcpu
99 struct usched_dfly_pcpu {
100 struct spinlock spin;
101 struct thread helper_thread;
105 struct lwp *uschedcp;
106 struct rq queues[NQS];
107 struct rq rtqueues[NQS];
108 struct rq idqueues[NQS];
110 u_int32_t rtqueuebits;
111 u_int32_t idqueuebits;
120 typedef struct usched_dfly_pcpu *dfly_pcpu_t;
122 static void dfly_acquire_curproc(struct lwp *lp);
123 static void dfly_release_curproc(struct lwp *lp);
124 static void dfly_select_curproc(globaldata_t gd);
125 static void dfly_setrunqueue(struct lwp *lp);
126 static void dfly_schedulerclock(struct lwp *lp, sysclock_t period,
128 static void dfly_recalculate_estcpu(struct lwp *lp);
129 static void dfly_resetpriority(struct lwp *lp);
130 static void dfly_forking(struct lwp *plp, struct lwp *lp);
131 static void dfly_exiting(struct lwp *lp, struct proc *);
132 static void dfly_uload_update(struct lwp *lp);
133 static void dfly_yield(struct lwp *lp);
135 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
136 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
137 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
141 static void dfly_need_user_resched_remote(void *dummy);
143 static struct lwp *dfly_chooseproc_locked(dfly_pcpu_t dd, struct lwp *chklp,
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 */
181 static volatile int dfly_scancpu;
182 /*static struct spinlock dfly_spin = SPINLOCK_INITIALIZER(dfly_spin);*/
184 static struct usched_dfly_pcpu dfly_pcpu[MAXCPU];
185 static struct sysctl_ctx_list usched_dfly_sysctl_ctx;
186 static struct sysctl_oid *usched_dfly_sysctl_tree;
188 /* Debug info exposed through debug.* sysctl */
190 static int usched_dfly_debug = -1;
191 SYSCTL_INT(_debug, OID_AUTO, dfly_scdebug, CTLFLAG_RW,
192 &usched_dfly_debug, 0,
193 "Print debug information for this pid");
195 static int usched_dfly_pid_debug = -1;
196 SYSCTL_INT(_debug, OID_AUTO, dfly_pid_debug, CTLFLAG_RW,
197 &usched_dfly_pid_debug, 0,
198 "Print KTR debug information for this pid");
200 static int usched_dfly_chooser = 0;
201 SYSCTL_INT(_debug, OID_AUTO, dfly_chooser, CTLFLAG_RW,
202 &usched_dfly_chooser, 0,
203 "Print KTR debug information for this pid");
205 /* Tunning usched_dfly - configurable through kern.usched_dfly.* */
207 static int usched_dfly_smt = 0;
208 static int usched_dfly_cache_coherent = 0;
209 static int usched_dfly_weight1 = 10;
210 static int usched_dfly_weight2 = 5;
211 static int usched_dfly_stick_to_level = 0;
213 static int usched_dfly_rrinterval = (ESTCPUFREQ + 9) / 10;
214 static int usched_dfly_decay = 8;
215 static int usched_dfly_batch_time = 10;
217 /* KTR debug printings */
219 KTR_INFO_MASTER(usched);
221 #if !defined(KTR_USCHED_DFLY)
222 #define KTR_USCHED_DFLY KTR_ALL
226 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_acquire_curproc_urw, 0,
227 "USCHED_DFLY(dfly_acquire_curproc in user_reseched_wanted "
228 "after release: pid %d, cpuid %d, curr_cpuid %d)",
229 pid_t pid, int cpuid, int curr);
230 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_acquire_curproc_before_loop, 0,
231 "USCHED_DFLY(dfly_acquire_curproc before loop: pid %d, cpuid %d, "
233 pid_t pid, int cpuid, int curr);
234 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_acquire_curproc_not, 0,
235 "USCHED_DFLY(dfly_acquire_curproc couldn't acquire after "
236 "dfly_setrunqueue: pid %d, cpuid %d, curr_lp pid %d, curr_cpuid %d)",
237 pid_t pid, int cpuid, pid_t curr_pid, int curr_cpuid);
238 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_acquire_curproc_switch, 0,
239 "USCHED_DFLY(dfly_acquire_curproc after lwkt_switch: pid %d, "
240 "cpuid %d, curr_cpuid %d)",
241 pid_t pid, int cpuid, int curr);
243 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_release_curproc, 0,
244 "USCHED_DFLY(dfly_release_curproc before select: pid %d, "
245 "cpuid %d, curr_cpuid %d)",
246 pid_t pid, int cpuid, int curr);
248 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_select_curproc, 0,
249 "USCHED_DFLY(dfly_release_curproc before select: pid %d, "
250 "cpuid %d, old_pid %d, old_cpuid %d, curr_cpuid %d)",
251 pid_t pid, int cpuid, pid_t old_pid, int old_cpuid, int curr);
254 KTR_INFO(KTR_USCHED_DFLY, usched, batchy_test_false, 0,
255 "USCHED_DFLY(batchy_looser_pri_test false: pid %d, "
256 "cpuid %d, verify_mask %lu)",
257 pid_t pid, int cpuid, cpumask_t mask);
258 KTR_INFO(KTR_USCHED_DFLY, usched, batchy_test_true, 0,
259 "USCHED_DFLY(batchy_looser_pri_test true: pid %d, "
260 "cpuid %d, verify_mask %lu)",
261 pid_t pid, int cpuid, cpumask_t mask);
263 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_fc_smt, 0,
264 "USCHED_DFLY(dfly_setrunqueue free cpus smt: pid %d, cpuid %d, "
265 "mask %lu, curr_cpuid %d)",
266 pid_t pid, int cpuid, cpumask_t mask, int curr);
267 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_fc_non_smt, 0,
268 "USCHED_DFLY(dfly_setrunqueue free cpus check non_smt: pid %d, "
269 "cpuid %d, mask %lu, curr_cpuid %d)",
270 pid_t pid, int cpuid, cpumask_t mask, int curr);
271 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_rc, 0,
272 "USCHED_DFLY(dfly_setrunqueue running cpus check: pid %d, "
273 "cpuid %d, mask %lu, curr_cpuid %d)",
274 pid_t pid, int cpuid, cpumask_t mask, int curr);
275 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_found, 0,
276 "USCHED_DFLY(dfly_setrunqueue found cpu: pid %d, cpuid %d, "
277 "mask %lu, found_cpuid %d, curr_cpuid %d)",
278 pid_t pid, int cpuid, cpumask_t mask, int found_cpuid, int curr);
279 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_not_found, 0,
280 "USCHED_DFLY(dfly_setrunqueue not found cpu: pid %d, cpuid %d, "
281 "try_cpuid %d, curr_cpuid %d)",
282 pid_t pid, int cpuid, int try_cpuid, int curr);
283 KTR_INFO(KTR_USCHED_DFLY, usched, dfly_setrunqueue_found_best_cpuid, 0,
284 "USCHED_DFLY(dfly_setrunqueue found cpu: pid %d, cpuid %d, "
285 "mask %lu, found_cpuid %d, curr_cpuid %d)",
286 pid_t pid, int cpuid, cpumask_t mask, int found_cpuid, int curr);
290 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc, 0,
291 "USCHED_DFLY(chooseproc: pid %d, old_cpuid %d, curr_cpuid %d)",
292 pid_t pid, int old_cpuid, int curr);
295 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc_cc, 0,
296 "USCHED_DFLY(chooseproc_cc: pid %d, old_cpuid %d, curr_cpuid %d)",
297 pid_t pid, int old_cpuid, int curr);
298 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc_cc_not_good, 0,
299 "USCHED_DFLY(chooseproc_cc not good: pid %d, old_cpumask %lu, "
300 "sibling_mask %lu, curr_cpumask %lu)",
301 pid_t pid, cpumask_t old_cpumask, cpumask_t sibling_mask, cpumask_t curr);
302 KTR_INFO(KTR_USCHED_DFLY, usched, chooseproc_cc_elected, 0,
303 "USCHED_DFLY(chooseproc_cc elected: pid %d, old_cpumask %lu, "
304 "sibling_mask %lu, curr_cpumask: %lu)",
305 pid_t pid, cpumask_t old_cpumask, cpumask_t sibling_mask, cpumask_t curr);
308 KTR_INFO(KTR_USCHED_DFLY, usched, sched_thread_no_process, 0,
309 "USCHED_DFLY(sched_thread %d no process scheduled: pid %d, old_cpuid %d)",
310 int id, pid_t pid, int cpuid);
311 KTR_INFO(KTR_USCHED_DFLY, usched, sched_thread_process, 0,
312 "USCHED_DFLY(sched_thread %d process scheduled: pid %d, old_cpuid %d)",
313 int id, pid_t pid, int cpuid);
315 KTR_INFO(KTR_USCHED_DFLY, usched, sched_thread_no_process_found, 0,
316 "USCHED_DFLY(sched_thread %d no process found; tmpmask %lu)",
317 int id, cpumask_t tmpmask);
322 * DFLY_ACQUIRE_CURPROC
324 * This function is called when the kernel intends to return to userland.
325 * It is responsible for making the thread the current designated userland
326 * thread for this cpu, blocking if necessary.
328 * The kernel has already depressed our LWKT priority so we must not switch
329 * until we have either assigned or disposed of the thread.
331 * WARNING! THIS FUNCTION IS ALLOWED TO CAUSE THE CURRENT THREAD TO MIGRATE
332 * TO ANOTHER CPU! Because most of the kernel assumes that no migration will
333 * occur, this function is called only under very controlled circumstances.
336 dfly_acquire_curproc(struct lwp *lp)
343 * Make sure we aren't sitting on a tsleep queue.
346 crit_enter_quick(td);
347 if (td->td_flags & TDF_TSLEEPQ)
349 dfly_recalculate_estcpu(lp);
352 * If a reschedule was requested give another thread the
355 if (user_resched_wanted()) {
356 clear_user_resched();
357 dfly_release_curproc(lp);
361 * Loop until we are the current user thread
364 dd = &dfly_pcpu[gd->gd_cpuid];
368 * Process any pending events and higher priority threads.
373 * Become the currently scheduled user thread for this cpu
374 * if we can do so trivially.
376 * We can steal another thread's current thread designation
377 * on this cpu since if we are running that other thread
378 * must not be, so we can safely deschedule it.
380 if (dd->uschedcp == lp) {
382 * We are already the current lwp (hot path).
384 dd->upri = lp->lwp_priority;
385 } else if (dd->uschedcp == NULL) {
387 * We can trivially become the current lwp.
389 atomic_set_cpumask(&dfly_curprocmask, gd->gd_cpumask);
391 dd->upri = lp->lwp_priority;
392 KKASSERT(lp->lwp_qcpu == dd->cpuid);
393 } else if (dd->uschedcp && dd->upri > lp->lwp_priority) {
395 * We can steal the current cpu's lwp designation
396 * away simply by replacing it. The other thread
397 * will stall when it tries to return to userland,
398 * possibly rescheduling elsewhere when it calls
402 dd->upri = lp->lwp_priority;
403 KKASSERT(lp->lwp_qcpu == dd->cpuid);
406 * We cannot become the current lwp, place the lp
407 * on the run-queue of this or another cpu and
408 * deschedule ourselves.
410 * When we are reactivated we will have another
413 lwkt_deschedule(lp->lwp_thread);
414 dfly_setrunqueue(lp);
417 * Reload after a switch or setrunqueue/switch possibly
418 * moved us to another cpu.
422 dd = &dfly_pcpu[gd->gd_cpuid];
424 } while (dd->uschedcp != lp);
427 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
431 * DFLY_RELEASE_CURPROC
433 * This routine detaches the current thread from the userland scheduler,
434 * usually because the thread needs to run or block in the kernel (at
435 * kernel priority) for a while.
437 * This routine is also responsible for selecting a new thread to
438 * make the current thread.
440 * NOTE: This implementation differs from the dummy example in that
441 * dfly_select_curproc() is able to select the current process, whereas
442 * dummy_select_curproc() is not able to select the current process.
443 * This means we have to NULL out uschedcp.
445 * Additionally, note that we may already be on a run queue if releasing
446 * via the lwkt_switch() in dfly_setrunqueue().
450 dfly_release_curproc(struct lwp *lp)
452 globaldata_t gd = mycpu;
453 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
456 * Make sure td_wakefromcpu is defaulted. This will be overwritten
459 lp->lwp_thread->td_wakefromcpu = gd->gd_cpuid;
461 if (dd->uschedcp == lp) {
463 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
465 dd->uschedcp = NULL; /* don't let lp be selected */
466 dd->upri = PRIBASE_NULL;
467 atomic_clear_cpumask(&dfly_curprocmask, gd->gd_cpumask);
468 dfly_select_curproc(gd);
474 * DFLY_SELECT_CURPROC
476 * Select a new current process for this cpu and clear any pending user
477 * reschedule request. The cpu currently has no current process.
479 * This routine is also responsible for equal-priority round-robining,
480 * typically triggered from dfly_schedulerclock(). In our dummy example
481 * all the 'user' threads are LWKT scheduled all at once and we just
482 * call lwkt_switch().
484 * The calling process is not on the queue and cannot be selected.
488 dfly_select_curproc(globaldata_t gd)
490 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
492 int cpuid = gd->gd_cpuid;
496 /*spin_lock(&dfly_spin);*/
497 spin_lock(&dd->spin);
498 nlp = dfly_chooseproc_locked(dd, dd->uschedcp, 0);
501 atomic_set_cpumask(&dfly_curprocmask, CPUMASK(cpuid));
502 dd->upri = nlp->lwp_priority;
504 dd->rrcount = 0; /* reset round robin */
505 spin_unlock(&dd->spin);
506 /*spin_unlock(&dfly_spin);*/
508 lwkt_acquire(nlp->lwp_thread);
510 lwkt_schedule(nlp->lwp_thread);
512 spin_unlock(&dd->spin);
513 /*spin_unlock(&dfly_spin);*/
519 * Place the specified lwp on the user scheduler's run queue. This routine
520 * must be called with the thread descheduled. The lwp must be runnable.
521 * It must not be possible for anyone else to explicitly schedule this thread.
523 * The thread may be the current thread as a special case.
526 dfly_setrunqueue(struct lwp *lp)
532 * First validate the process LWKT state.
535 KASSERT(lp->lwp_stat == LSRUN, ("setrunqueue: lwp not LSRUN"));
536 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0,
537 ("lwp %d/%d already on runq! flag %08x/%08x", lp->lwp_proc->p_pid,
538 lp->lwp_tid, lp->lwp_proc->p_flags, lp->lwp_flags));
539 KKASSERT((lp->lwp_thread->td_flags & TDF_RUNQ) == 0);
542 * NOTE: rdd does not necessarily represent the current cpu.
543 * Instead it represents the cpu the thread was last
546 rdd = &dfly_pcpu[lp->lwp_qcpu];
549 * This process is not supposed to be scheduled anywhere or assigned
550 * as the current process anywhere. Assert the condition.
552 KKASSERT(rdd->uschedcp != lp);
556 * If we are not SMP we do not have a scheduler helper to kick
557 * and must directly activate the process if none are scheduled.
559 * This is really only an issue when bootstrapping init since
560 * the caller in all other cases will be a user process, and
561 * even if released (rdd->uschedcp == NULL), that process will
562 * kickstart the scheduler when it returns to user mode from
565 * NOTE: On SMP we can't just set some other cpu's uschedcp.
567 if (rdd->uschedcp == NULL) {
568 spin_lock(&rdd->spin);
569 if (rdd->uschedcp == NULL) {
570 atomic_set_cpumask(&dfly_curprocmask, 1);
572 rdd->upri = lp->lwp_priority;
573 spin_unlock(&rdd->spin);
574 lwkt_schedule(lp->lwp_thread);
578 spin_unlock(&rdd->spin);
584 * XXX fixme. Could be part of a remrunqueue/setrunqueue
585 * operation when the priority is recalculated, so TDF_MIGRATING
586 * may already be set.
588 if ((lp->lwp_thread->td_flags & TDF_MIGRATING) == 0)
589 lwkt_giveaway(lp->lwp_thread);
594 * Ok, we have to setrunqueue some target cpu and request a reschedule
597 * We have to choose the best target cpu. It might not be the current
598 * target even if the current cpu has no running user thread (for
599 * example, because the current cpu might be a hyperthread and its
600 * sibling has a thread assigned).
602 /*spin_lock(&dfly_spin);*/
603 rdd = dfly_choose_best_queue(lp);
604 rgd = globaldata_find(rdd->cpuid);
607 * We lose control of lp the moment we release the spinlock after
608 * having placed lp on the queue. i.e. another cpu could pick it
609 * up and it could exit, or its priority could be further adjusted,
610 * or something like that.
612 * WARNING! dd can point to a foreign cpu!
614 spin_lock(&rdd->spin);
615 dfly_setrunqueue_locked(rdd, lp);
616 /*spin_unlock(&dfly_spin);*/
619 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
620 spin_unlock(&rdd->spin);
621 if (rdd->uschedcp == NULL) {
622 wakeup_mycpu(&rdd->helper_thread); /* XXX */
628 spin_unlock(&rdd->spin);
631 atomic_clear_cpumask(&dfly_rdyprocmask, rgd->gd_cpumask);
632 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
633 spin_unlock(&rdd->spin);
634 lwkt_send_ipiq(rgd, dfly_need_user_resched_remote,
637 spin_unlock(&rdd->spin);
638 wakeup(&rdd->helper_thread);
643 * Request a reschedule if appropriate.
645 spin_lock(&rdd->spin);
646 dfly_setrunqueue_locked(rdd, lp);
647 spin_unlock(&rdd->spin);
648 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK)) {
656 * This routine is called from a systimer IPI. It MUST be MP-safe and
657 * the BGL IS NOT HELD ON ENTRY. This routine is called at ESTCPUFREQ on
662 dfly_schedulerclock(struct lwp *lp, sysclock_t period, sysclock_t cpstamp)
664 globaldata_t gd = mycpu;
665 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
668 * Do we need to round-robin? We round-robin 10 times a second.
669 * This should only occur for cpu-bound batch processes.
671 if (++dd->rrcount >= usched_dfly_rrinterval) {
677 * Adjust estcpu upward using a real time equivalent calculation.
679 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUMAX / ESTCPUFREQ + 1);
682 * Spinlocks also hold a critical section so there should not be
685 KKASSERT(gd->gd_spinlocks_wr == 0);
687 dfly_resetpriority(lp);
691 * Called from acquire and from kern_synch's one-second timer (one of the
692 * callout helper threads) with a critical section held.
694 * Decay p_estcpu based on the number of ticks we haven't been running
695 * and our p_nice. As the load increases each process observes a larger
696 * number of idle ticks (because other processes are running in them).
697 * This observation leads to a larger correction which tends to make the
698 * system more 'batchy'.
700 * Note that no recalculation occurs for a process which sleeps and wakes
701 * up in the same tick. That is, a system doing thousands of context
702 * switches per second will still only do serious estcpu calculations
703 * ESTCPUFREQ times per second.
707 dfly_recalculate_estcpu(struct lwp *lp)
709 globaldata_t gd = mycpu;
710 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
717 * We have to subtract periodic to get the last schedclock
718 * timeout time, otherwise we would get the upcoming timeout.
719 * Keep in mind that a process can migrate between cpus and
720 * while the scheduler clock should be very close, boundary
721 * conditions could lead to a small negative delta.
723 cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
725 if (lp->lwp_slptime > 1) {
727 * Too much time has passed, do a coarse correction.
729 lp->lwp_estcpu = lp->lwp_estcpu >> 1;
730 dfly_resetpriority(lp);
731 lp->lwp_cpbase = cpbase;
733 lp->lwp_batch -= ESTCPUFREQ;
734 if (lp->lwp_batch < 0)
736 } else if (lp->lwp_cpbase != cpbase) {
738 * Adjust estcpu if we are in a different tick. Don't waste
739 * time if we are in the same tick.
741 * First calculate the number of ticks in the measurement
742 * interval. The ttlticks calculation can wind up 0 due to
743 * a bug in the handling of lwp_slptime (as yet not found),
744 * so make sure we do not get a divide by 0 panic.
746 ttlticks = (cpbase - lp->lwp_cpbase) /
747 gd->gd_schedclock.periodic;
750 lp->lwp_cpbase = cpbase;
754 updatepcpu(lp, lp->lwp_cpticks, ttlticks);
757 * Calculate the percentage of one cpu used factoring in ncpus
758 * and the load and adjust estcpu. Handle degenerate cases
759 * by adding 1 to runqcount.
761 * estcpu is scaled by ESTCPUMAX.
763 * runqcount is the excess number of user processes
764 * that cannot be immediately scheduled to cpus. We want
765 * to count these as running to avoid range compression
766 * in the base calculation (which is the actual percentage
769 estcpu = (lp->lwp_cpticks * ESTCPUMAX) *
770 (dd->runqcount + ncpus) / (ncpus * ttlticks);
773 * If estcpu is > 50% we become more batch-like
774 * If estcpu is <= 50% we become less batch-like
776 * It takes 30 cpu seconds to traverse the entire range.
778 if (estcpu > ESTCPUMAX / 2) {
779 lp->lwp_batch += ttlticks;
780 if (lp->lwp_batch > BATCHMAX)
781 lp->lwp_batch = BATCHMAX;
783 lp->lwp_batch -= ttlticks;
784 if (lp->lwp_batch < 0)
788 if (usched_dfly_debug == lp->lwp_proc->p_pid) {
789 kprintf("pid %d lwp %p estcpu %3d %3d bat %d cp %d/%d",
790 lp->lwp_proc->p_pid, lp,
791 estcpu, lp->lwp_estcpu,
793 lp->lwp_cpticks, ttlticks);
797 * Adjust lp->lwp_esetcpu. The decay factor determines how
798 * quickly lwp_estcpu collapses to its realtime calculation.
799 * A slower collapse gives us a more accurate number but
800 * can cause a cpu hog to eat too much cpu before the
801 * scheduler decides to downgrade it.
803 * NOTE: p_nice is accounted for in dfly_resetpriority(),
804 * and not here, but we must still ensure that a
805 * cpu-bound nice -20 process does not completely
806 * override a cpu-bound nice +20 process.
808 * NOTE: We must use ESTCPULIM() here to deal with any
811 decay_factor = usched_dfly_decay;
812 if (decay_factor < 1)
814 if (decay_factor > 1024)
817 lp->lwp_estcpu = ESTCPULIM(
818 (lp->lwp_estcpu * decay_factor + estcpu) /
821 if (usched_dfly_debug == lp->lwp_proc->p_pid)
822 kprintf(" finalestcpu %d\n", lp->lwp_estcpu);
823 dfly_resetpriority(lp);
824 lp->lwp_cpbase += ttlticks * gd->gd_schedclock.periodic;
830 * Compute the priority of a process when running in user mode.
831 * Arrange to reschedule if the resulting priority is better
832 * than that of the current process.
834 * This routine may be called with any process.
836 * This routine is called by fork1() for initial setup with the process
837 * of the run queue, and also may be called normally with the process on or
841 dfly_resetpriority(struct lwp *lp)
853 * Lock the scheduler (lp) belongs to. This can be on a different
854 * cpu. Handle races. This loop breaks out with the appropriate
859 rdd = &dfly_pcpu[rcpu];
860 spin_lock(&rdd->spin);
861 if (rcpu == lp->lwp_qcpu)
863 spin_unlock(&rdd->spin);
867 * Calculate the new priority and queue type
869 newrqtype = lp->lwp_rtprio.type;
872 case RTP_PRIO_REALTIME:
874 newpriority = PRIBASE_REALTIME +
875 (lp->lwp_rtprio.prio & PRIMASK);
877 case RTP_PRIO_NORMAL:
879 * Detune estcpu based on batchiness. lwp_batch ranges
880 * from 0 to BATCHMAX. Limit estcpu for the sake of
881 * the priority calculation to between 50% and 100%.
883 estcpu = lp->lwp_estcpu * (lp->lwp_batch + BATCHMAX) /
887 * p_nice piece Adds (0-40) * 2 0-80
888 * estcpu Adds 16384 * 4 / 512 0-128
890 newpriority = (lp->lwp_proc->p_nice - PRIO_MIN) * PPQ / NICEPPQ;
891 newpriority += estcpu * PPQ / ESTCPUPPQ;
892 newpriority = newpriority * MAXPRI / (PRIO_RANGE * PPQ /
893 NICEPPQ + ESTCPUMAX * PPQ / ESTCPUPPQ);
894 newpriority = PRIBASE_NORMAL + (newpriority & PRIMASK);
897 newpriority = PRIBASE_IDLE + (lp->lwp_rtprio.prio & PRIMASK);
899 case RTP_PRIO_THREAD:
900 newpriority = PRIBASE_THREAD + (lp->lwp_rtprio.prio & PRIMASK);
903 panic("Bad RTP_PRIO %d", newrqtype);
908 * The newpriority incorporates the queue type so do a simple masked
909 * check to determine if the process has moved to another queue. If
910 * it has, and it is currently on a run queue, then move it.
912 * Since uload is ~PPQMASK masked, no modifications are necessary if
913 * we end up in the same run queue.
915 if ((lp->lwp_priority ^ newpriority) & ~PPQMASK) {
919 * uload can change, calculate the adjustment to reduce
920 * edge cases since choosers scan the cpu topology without
923 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
925 -((lp->lwp_priority & ~PPQMASK) & PRIMASK) +
926 ((newpriority & ~PPQMASK) & PRIMASK);
927 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload,
930 if (lp->lwp_mpflags & LWP_MP_ONRUNQ) {
931 dfly_remrunqueue_locked(rdd, lp);
932 lp->lwp_priority = newpriority;
933 lp->lwp_rqtype = newrqtype;
934 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
935 dfly_setrunqueue_locked(rdd, lp);
938 lp->lwp_priority = newpriority;
939 lp->lwp_rqtype = newrqtype;
940 lp->lwp_rqindex = (newpriority & PRIMASK) / PPQ;
945 * In the same PPQ, uload cannot change.
947 lp->lwp_priority = newpriority;
953 * Determine if we need to reschedule the target cpu. This only
954 * occurs if the LWP is already on a scheduler queue, which means
955 * that idle cpu notification has already occured. At most we
956 * need only issue a need_user_resched() on the appropriate cpu.
958 * The LWP may be owned by a CPU different from the current one,
959 * in which case dd->uschedcp may be modified without an MP lock
960 * or a spinlock held. The worst that happens is that the code
961 * below causes a spurious need_user_resched() on the target CPU
962 * and dd->pri to be wrong for a short period of time, both of
963 * which are harmless.
965 * If checkpri is 0 we are adjusting the priority of the current
966 * process, possibly higher (less desireable), so ignore the upri
967 * check which will fail in that case.
970 if ((dfly_rdyprocmask & CPUMASK(rcpu)) &&
972 (rdd->upri & ~PRIMASK) > (lp->lwp_priority & ~PRIMASK))) {
974 if (rcpu == mycpu->gd_cpuid) {
975 spin_unlock(&rdd->spin);
978 atomic_clear_cpumask(&dfly_rdyprocmask,
980 spin_unlock(&rdd->spin);
981 lwkt_send_ipiq(globaldata_find(rcpu),
982 dfly_need_user_resched_remote,
986 spin_unlock(&rdd->spin);
990 spin_unlock(&rdd->spin);
993 spin_unlock(&rdd->spin);
1000 dfly_yield(struct lwp *lp)
1003 /* FUTURE (or something similar) */
1004 switch(lp->lwp_rqtype) {
1005 case RTP_PRIO_NORMAL:
1006 lp->lwp_estcpu = ESTCPULIM(lp->lwp_estcpu + ESTCPUINCR);
1012 need_user_resched();
1016 * Called from fork1() when a new child process is being created.
1018 * Give the child process an initial estcpu that is more batch then
1019 * its parent and dock the parent for the fork (but do not
1020 * reschedule the parent). This comprises the main part of our batch
1021 * detection heuristic for both parallel forking and sequential execs.
1023 * XXX lwp should be "spawning" instead of "forking"
1026 dfly_forking(struct lwp *plp, struct lwp *lp)
1029 * Put the child 4 queue slots (out of 32) higher than the parent
1030 * (less desireable than the parent).
1032 lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
1035 * The batch status of children always starts out centerline
1036 * and will inch-up or inch-down as appropriate. It takes roughly
1037 * ~15 seconds of >50% cpu to hit the limit.
1039 lp->lwp_batch = BATCHMAX / 2;
1042 * Dock the parent a cost for the fork, protecting us from fork
1043 * bombs. If the parent is forking quickly make the child more
1046 plp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ / 16);
1050 * Called when a lwp is being removed from this scheduler, typically
1051 * during lwp_exit().
1054 dfly_exiting(struct lwp *lp, struct proc *child_proc)
1056 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1058 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1059 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1060 atomic_add_int(&dd->uload,
1061 -((lp->lwp_priority & ~PPQMASK) & PRIMASK));
1066 dfly_uload_update(struct lwp *lp)
1068 dfly_pcpu_t dd = &dfly_pcpu[lp->lwp_qcpu];
1070 if (lp->lwp_thread->td_flags & TDF_RUNQ) {
1071 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1072 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1073 atomic_add_int(&dd->uload,
1074 ((lp->lwp_priority & ~PPQMASK) & PRIMASK));
1077 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1078 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1079 atomic_add_int(&dd->uload,
1080 -((lp->lwp_priority & ~PPQMASK) & PRIMASK));
1086 * chooseproc() is called when a cpu needs a user process to LWKT schedule,
1087 * it selects a user process and returns it. If chklp is non-NULL and chklp
1088 * has a better or equal priority then the process that would otherwise be
1089 * chosen, NULL is returned.
1091 * Until we fix the RUNQ code the chklp test has to be strict or we may
1092 * bounce between processes trying to acquire the current process designation.
1094 * Must be called with dfly_spin exclusive held. The spinlock is
1095 * left intact through the entire routine.
1097 * if chklp is NULL this function will dive other cpu's queues looking
1098 * for work if the current queue is empty.
1102 dfly_chooseproc_locked(dfly_pcpu_t dd, struct lwp *chklp, int isremote)
1109 u_int32_t *which, *which2;
1115 rtqbits = dd->rtqueuebits;
1116 tsqbits = dd->queuebits;
1117 idqbits = dd->idqueuebits;
1120 pri = bsfl(rtqbits);
1121 q = &dd->rtqueues[pri];
1122 which = &dd->rtqueuebits;
1124 } else if (tsqbits) {
1125 pri = bsfl(tsqbits);
1126 q = &dd->queues[pri];
1127 which = &dd->queuebits;
1129 } else if (idqbits) {
1130 pri = bsfl(idqbits);
1131 q = &dd->idqueues[pri];
1132 which = &dd->idqueuebits;
1138 * Disallow remote->remote recursion
1143 * Pull a runnable thread from a remote run queue. We have
1144 * to adjust qcpu and uload manually because the lp we return
1145 * might be assigned directly to uschedcp (setrunqueue might
1148 xdd = dfly_choose_worst_queue(dd);
1149 if (xdd && xdd != dd && spin_trylock(&xdd->spin)) {
1150 lp = dfly_chooseproc_locked(xdd, NULL, 1);
1152 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1153 atomic_add_int(&xdd->uload,
1154 -((lp->lwp_priority & ~PPQMASK) &
1157 lp->lwp_qcpu = dd->cpuid;
1158 atomic_add_int(&dd->uload,
1159 ((lp->lwp_priority & ~PPQMASK) & PRIMASK));
1160 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1162 spin_unlock(&xdd->spin);
1173 lp = TAILQ_FIRST(q);
1174 KASSERT(lp, ("chooseproc: no lwp on busy queue"));
1177 * If the passed lwp <chklp> is reasonably close to the selected
1178 * lwp <lp>, return NULL (indicating that <chklp> should be kept).
1180 * Note that we must error on the side of <chklp> to avoid bouncing
1181 * between threads in the acquire code.
1184 if (chklp->lwp_priority < lp->lwp_priority + PPQ)
1188 KTR_COND_LOG(usched_chooseproc,
1189 lp->lwp_proc->p_pid == usched_dfly_pid_debug,
1190 lp->lwp_proc->p_pid,
1191 lp->lwp_thread->td_gd->gd_cpuid,
1194 TAILQ_REMOVE(q, lp, lwp_procq);
1197 *which &= ~(1 << pri);
1198 KASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) != 0, ("not on runq6!"));
1199 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1207 * USED TO PUSH RUNNABLE LWPS TO THE LEAST LOADED CPU.
1209 * Choose a cpu node to schedule lp on, hopefully nearby its current
1210 * node. We give the current node a modest advantage for obvious reasons.
1212 * We also give the node the thread was woken up FROM a slight advantage
1213 * in order to try to schedule paired threads which synchronize/block waiting
1214 * for each other fairly close to each other. Similarly in a network setting
1215 * this feature will also attempt to place a user process near the kernel
1216 * protocol thread that is feeding it data. THIS IS A CRITICAL PART of the
1217 * algorithm as it heuristically groups synchronizing processes for locality
1218 * of reference in multi-socket systems.
1220 * The caller will normally dfly_setrunqueue() lp on the returned queue.
1222 * When the topology is known choose a cpu whos group has, in aggregate,
1223 * has the lowest weighted load.
1227 dfly_choose_best_queue(struct lwp *lp)
1233 dfly_pcpu_t dd1 = &dfly_pcpu[lp->lwp_qcpu];
1234 dfly_pcpu_t dd2 = &dfly_pcpu[lp->lwp_thread->td_wakefromcpu];
1243 * When the topology is unknown choose a random cpu that is hopefully
1246 if (dd1->cpunode == NULL)
1247 return (dfly_choose_queue_simple(dd1, lp));
1250 * When the topology is known choose a cpu whos group has, in
1251 * aggregate, has the lowest weighted load.
1253 cpup = root_cpu_node;
1255 level = cpu_topology_levels_number;
1259 * Degenerate case super-root
1261 if (cpup->child_node && cpup->child_no == 1) {
1262 cpup = cpup->child_node;
1270 if (cpup->child_node == NULL) {
1271 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1276 lowest_load = 0x7FFFFFFF;
1278 for (n = 0; n < cpup->child_no; ++n) {
1280 * Accumulate load information for all cpus
1281 * which are members of this node.
1283 cpun = &cpup->child_node[n];
1284 mask = cpun->members & usched_global_cpumask &
1285 smp_active_mask & lp->lwp_cpumask;
1290 cpuid = BSFCPUMASK(mask);
1291 load += dfly_pcpu[cpuid].uload;
1292 mask &= ~CPUMASK(cpuid);
1296 * Give a slight advantage to nearby cpus.
1298 if (cpun->members & dd1->cpumask)
1299 load -= PPQ * level * usched_dfly_weight1 / 10;
1300 else if (cpun->members & dd2->cpumask)
1301 load -= PPQ * level * usched_dfly_weight2 / 10;
1304 * Calculate the best load
1306 if (cpub == NULL || lowest_load > load ||
1307 (lowest_load == load &&
1308 (cpun->members & dd1->cpumask))
1317 if (usched_dfly_chooser)
1318 kprintf("lp %02d->%02d %s\n",
1319 lp->lwp_qcpu, rdd->cpuid, lp->lwp_proc->p_comm);
1324 * USED TO PULL RUNNABLE LWPS FROM THE MOST LOADED CPU.
1326 * Choose the worst queue close to dd's cpu node with a non-empty runq.
1328 * This is used by the thread chooser when the current cpu's queues are
1329 * empty to steal a thread from another cpu's queue. We want to offload
1330 * the most heavily-loaded queue.
1334 dfly_choose_worst_queue(dfly_pcpu_t dd)
1349 * When the topology is unknown choose a random cpu that is hopefully
1352 if (dd->cpunode == NULL) {
1357 * When the topology is known choose a cpu whos group has, in
1358 * aggregate, has the lowest weighted load.
1360 cpup = root_cpu_node;
1362 level = cpu_topology_levels_number;
1365 * Degenerate case super-root
1367 if (cpup->child_node && cpup->child_no == 1) {
1368 cpup = cpup->child_node;
1376 if (cpup->child_node == NULL) {
1377 rdd = &dfly_pcpu[BSFCPUMASK(cpup->members)];
1384 for (n = 0; n < cpup->child_no; ++n) {
1386 * Accumulate load information for all cpus
1387 * which are members of this node.
1389 cpun = &cpup->child_node[n];
1390 mask = cpun->members & usched_global_cpumask &
1397 cpuid = BSFCPUMASK(mask);
1398 load += dfly_pcpu[cpuid].uload;
1399 if (dfly_pcpu[cpuid].uload)
1401 mask &= ~CPUMASK(cpuid);
1405 * Give a slight advantage to nearby cpus.
1407 if (cpun->members & dd->cpumask)
1408 load += PPQ * level;
1411 * The best candidate is the one with the worst
1412 * (highest) load. Prefer candiates that are
1413 * closer to our cpu.
1416 (cpub == NULL || highest_load < load ||
1417 (highest_load == load &&
1418 (cpun->members & dd->cpumask)))
1420 highest_load = load;
1432 dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
1440 * Fallback to the original heuristic, select random cpu,
1441 * first checking cpus not currently running a user thread.
1443 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1444 mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
1445 smp_active_mask & usched_global_cpumask;
1448 tmpmask = ~(CPUMASK(cpuid) - 1);
1450 cpuid = BSFCPUMASK(mask & tmpmask);
1452 cpuid = BSFCPUMASK(mask);
1453 rdd = &dfly_pcpu[cpuid];
1455 if ((rdd->upri & ~PPQMASK) >= (lp->lwp_priority & ~PPQMASK))
1457 mask &= ~CPUMASK(cpuid);
1461 * Then cpus which might have a currently running lp
1463 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1464 mask = dfly_curprocmask & dfly_rdyprocmask &
1465 lp->lwp_cpumask & smp_active_mask & usched_global_cpumask;
1468 tmpmask = ~(CPUMASK(cpuid) - 1);
1470 cpuid = BSFCPUMASK(mask & tmpmask);
1472 cpuid = BSFCPUMASK(mask);
1473 rdd = &dfly_pcpu[cpuid];
1475 if ((rdd->upri & ~PPQMASK) > (lp->lwp_priority & ~PPQMASK))
1477 mask &= ~CPUMASK(cpuid);
1481 * If we cannot find a suitable cpu we reload from dfly_scancpu
1482 * and round-robin. Other cpus will pickup as they release their
1483 * current lwps or become ready.
1485 * Avoid a degenerate system lockup case if usched_global_cpumask
1486 * is set to 0 or otherwise does not cover lwp_cpumask.
1488 * We only kick the target helper thread in this case, we do not
1489 * set the user resched flag because
1491 cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
1492 if ((CPUMASK(cpuid) & usched_global_cpumask) == 0)
1494 rdd = &dfly_pcpu[cpuid];
1501 dfly_need_user_resched_remote(void *dummy)
1503 globaldata_t gd = mycpu;
1504 dfly_pcpu_t dd = &dfly_pcpu[gd->gd_cpuid];
1506 need_user_resched();
1508 /* Call wakeup_mycpu to avoid sending IPIs to other CPUs */
1509 wakeup_mycpu(&dd->helper_thread);
1515 * dfly_remrunqueue_locked() removes a given process from the run queue
1516 * that it is on, clearing the queue busy bit if it becomes empty.
1518 * Note that user process scheduler is different from the LWKT schedule.
1519 * The user process scheduler only manages user processes but it uses LWKT
1520 * underneath, and a user process operating in the kernel will often be
1521 * 'released' from our management.
1523 * uload is NOT adjusted here. It is only adjusted if the lwkt_thread goes
1524 * to sleep or the lwp is moved to a different runq.
1527 dfly_remrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1533 KKASSERT(lp->lwp_mpflags & LWP_MP_ONRUNQ);
1534 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1536 /*rdd->uload -= (lp->lwp_priority & ~PPQMASK) & PRIMASK;*/
1537 KKASSERT(rdd->runqcount >= 0);
1539 pri = lp->lwp_rqindex;
1540 switch(lp->lwp_rqtype) {
1541 case RTP_PRIO_NORMAL:
1542 q = &rdd->queues[pri];
1543 which = &rdd->queuebits;
1545 case RTP_PRIO_REALTIME:
1547 q = &rdd->rtqueues[pri];
1548 which = &rdd->rtqueuebits;
1551 q = &rdd->idqueues[pri];
1552 which = &rdd->idqueuebits;
1555 panic("remrunqueue: invalid rtprio type");
1558 TAILQ_REMOVE(q, lp, lwp_procq);
1559 if (TAILQ_EMPTY(q)) {
1560 KASSERT((*which & (1 << pri)) != 0,
1561 ("remrunqueue: remove from empty queue"));
1562 *which &= ~(1 << pri);
1567 * dfly_setrunqueue_locked()
1569 * Add a process whos rqtype and rqindex had previously been calculated
1570 * onto the appropriate run queue. Determine if the addition requires
1571 * a reschedule on a cpu and return the cpuid or -1.
1573 * NOTE: Lower priorities are better priorities.
1575 * NOTE ON ULOAD: This variable specifies the aggregate load on a cpu, the
1576 * sum of the rough lwp_priority for all running and runnable
1577 * processes. Lower priority processes (higher lwp_priority
1578 * values) actually DO count as more load, not less, because
1579 * these are the programs which require the most care with
1580 * regards to cpu selection.
1583 dfly_setrunqueue_locked(dfly_pcpu_t rdd, struct lwp *lp)
1589 if (lp->lwp_qcpu != rdd->cpuid) {
1590 if (lp->lwp_mpflags & LWP_MP_ULOAD) {
1591 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1592 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload,
1593 -((lp->lwp_priority & ~PPQMASK) & PRIMASK));
1595 lp->lwp_qcpu = rdd->cpuid;
1598 KKASSERT((lp->lwp_mpflags & LWP_MP_ONRUNQ) == 0);
1599 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ONRUNQ);
1601 if ((lp->lwp_mpflags & LWP_MP_ULOAD) == 0) {
1602 atomic_set_int(&lp->lwp_mpflags, LWP_MP_ULOAD);
1603 atomic_add_int(&dfly_pcpu[lp->lwp_qcpu].uload,
1604 (lp->lwp_priority & ~PPQMASK) & PRIMASK);
1607 pri = lp->lwp_rqindex;
1609 switch(lp->lwp_rqtype) {
1610 case RTP_PRIO_NORMAL:
1611 q = &rdd->queues[pri];
1612 which = &rdd->queuebits;
1614 case RTP_PRIO_REALTIME:
1616 q = &rdd->rtqueues[pri];
1617 which = &rdd->rtqueuebits;
1620 q = &rdd->idqueues[pri];
1621 which = &rdd->idqueuebits;
1624 panic("remrunqueue: invalid rtprio type");
1629 * Add to the correct queue and set the appropriate bit. If no
1630 * lower priority (i.e. better) processes are in the queue then
1631 * we want a reschedule, calculate the best cpu for the job.
1633 * Always run reschedules on the LWPs original cpu.
1635 TAILQ_INSERT_TAIL(q, lp, lwp_procq);
1642 * For SMP systems a user scheduler helper thread is created for each
1643 * cpu and is used to allow one cpu to wakeup another for the purposes of
1644 * scheduling userland threads from setrunqueue().
1646 * UP systems do not need the helper since there is only one cpu.
1648 * We can't use the idle thread for this because we might block.
1649 * Additionally, doing things this way allows us to HLT idle cpus
1653 dfly_helper_thread(void *dummy)
1662 cpuid = gd->gd_cpuid; /* doesn't change */
1663 mask = gd->gd_cpumask; /* doesn't change */
1664 dd = &dfly_pcpu[cpuid];
1667 * Since we only want to be woken up only when no user processes
1668 * are scheduled on a cpu, run at an ultra low priority.
1670 lwkt_setpri_self(TDPRI_USER_SCHEDULER);
1672 tsleep(&dd->helper_thread, 0, "schslp", 0);
1676 * We use the LWKT deschedule-interlock trick to avoid racing
1677 * dfly_rdyprocmask. This means we cannot block through to the
1678 * manual lwkt_switch() call we make below.
1681 tsleep_interlock(&dd->helper_thread, 0);
1683 /*spin_lock(&dfly_spin);*/
1684 spin_lock(&dd->spin);
1686 atomic_set_cpumask(&dfly_rdyprocmask, mask);
1687 clear_user_resched(); /* This satisfied the reschedule request */
1688 dd->rrcount = 0; /* Reset the round-robin counter */
1690 if ((dfly_curprocmask & mask) == 0) {
1692 * No thread is currently scheduled.
1694 KKASSERT(dd->uschedcp == NULL);
1695 if ((nlp = dfly_chooseproc_locked(dd, NULL, 0)) != NULL) {
1696 KTR_COND_LOG(usched_sched_thread_no_process,
1697 nlp->lwp_proc->p_pid == usched_dfly_pid_debug,
1699 nlp->lwp_proc->p_pid,
1700 nlp->lwp_thread->td_gd->gd_cpuid);
1702 atomic_set_cpumask(&dfly_curprocmask, mask);
1703 dd->upri = nlp->lwp_priority;
1705 dd->rrcount = 0; /* reset round robin */
1706 spin_unlock(&dd->spin);
1707 /*spin_unlock(&dfly_spin);*/
1708 lwkt_acquire(nlp->lwp_thread);
1709 lwkt_schedule(nlp->lwp_thread);
1711 spin_unlock(&dd->spin);
1712 /*spin_unlock(&dfly_spin);*/
1714 } else if (dd->runqcount) {
1716 * Possibly find a better process to schedule.
1718 nlp = dfly_chooseproc_locked(dd, dd->uschedcp, 0);
1720 KTR_COND_LOG(usched_sched_thread_process,
1721 nlp->lwp_proc->p_pid == usched_dfly_pid_debug,
1723 nlp->lwp_proc->p_pid,
1724 nlp->lwp_thread->td_gd->gd_cpuid);
1726 dd->upri = nlp->lwp_priority;
1728 dd->rrcount = 0; /* reset round robin */
1729 spin_unlock(&dd->spin);
1730 /*spin_unlock(&dfly_spin);*/
1731 lwkt_acquire(nlp->lwp_thread);
1732 lwkt_schedule(nlp->lwp_thread);
1735 * Leave the thread on our run queue. Another
1736 * scheduler will try to pull it later.
1738 spin_unlock(&dd->spin);
1739 /*spin_unlock(&dfly_spin);*/
1743 * The runq is empty.
1745 spin_unlock(&dd->spin);
1746 /*spin_unlock(&dfly_spin);*/
1750 * We're descheduled unless someone scheduled us. Switch away.
1751 * Exiting the critical section will cause splz() to be called
1752 * for us if interrupts and such are pending.
1755 tsleep(&dd->helper_thread, PINTERLOCKED, "schslp", 0);
1759 /* sysctl stick_to_level parameter */
1761 sysctl_usched_dfly_stick_to_level(SYSCTL_HANDLER_ARGS)
1765 new_val = usched_dfly_stick_to_level;
1767 error = sysctl_handle_int(oidp, &new_val, 0, req);
1768 if (error != 0 || req->newptr == NULL)
1770 if (new_val > cpu_topology_levels_number - 1 || new_val < 0)
1772 usched_dfly_stick_to_level = new_val;
1777 * Setup our scheduler helpers. Note that curprocmask bit 0 has already
1778 * been cleared by rqinit() and we should not mess with it further.
1781 dfly_helper_thread_cpu_init(void)
1786 int smt_not_supported = 0;
1787 int cache_coherent_not_supported = 0;
1790 kprintf("Start scheduler helpers on cpus:\n");
1792 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
1793 usched_dfly_sysctl_tree =
1794 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
1795 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
1796 "usched_dfly", CTLFLAG_RD, 0, "");
1798 for (i = 0; i < ncpus; ++i) {
1799 dfly_pcpu_t dd = &dfly_pcpu[i];
1800 cpumask_t mask = CPUMASK(i);
1802 if ((mask & smp_active_mask) == 0)
1805 spin_init(&dd->spin);
1806 dd->cpunode = get_cpu_node_by_cpuid(i);
1808 dd->cpumask = CPUMASK(i);
1809 for (j = 0; j < NQS; j++) {
1810 TAILQ_INIT(&dd->queues[j]);
1811 TAILQ_INIT(&dd->rtqueues[j]);
1812 TAILQ_INIT(&dd->idqueues[j]);
1814 atomic_clear_cpumask(&dfly_curprocmask, 1);
1816 if (dd->cpunode == NULL) {
1817 smt_not_supported = 1;
1818 cache_coherent_not_supported = 1;
1820 kprintf ("\tcpu%d - WARNING: No CPU NODE "
1821 "found for cpu\n", i);
1823 switch (dd->cpunode->type) {
1826 kprintf ("\tcpu%d - HyperThreading "
1827 "available. Core siblings: ",
1831 smt_not_supported = 1;
1834 kprintf ("\tcpu%d - No HT available, "
1835 "multi-core/physical "
1836 "cpu. Physical siblings: ",
1840 smt_not_supported = 1;
1843 kprintf ("\tcpu%d - No HT available, "
1844 "single-core/physical cpu. "
1845 "Package Siblings: ",
1849 /* Let's go for safe defaults here */
1850 smt_not_supported = 1;
1851 cache_coherent_not_supported = 1;
1853 kprintf ("\tcpu%d - Unknown cpunode->"
1854 "type=%u. Siblings: ",
1856 (u_int)dd->cpunode->type);
1861 if (dd->cpunode->parent_node != NULL) {
1862 CPUSET_FOREACH(cpuid, dd->cpunode->parent_node->members)
1863 kprintf("cpu%d ", cpuid);
1866 kprintf(" no siblings\n");
1871 lwkt_create(dfly_helper_thread, NULL, NULL, &dd->helper_thread,
1872 0, i, "usched %d", i);
1875 * Allow user scheduling on the target cpu. cpu #0 has already
1876 * been enabled in rqinit().
1879 atomic_clear_cpumask(&dfly_curprocmask, mask);
1880 atomic_set_cpumask(&dfly_rdyprocmask, mask);
1881 dd->upri = PRIBASE_NULL;
1885 /* usched_dfly sysctl configurable parameters */
1887 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1888 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1889 OID_AUTO, "rrinterval", CTLFLAG_RW,
1890 &usched_dfly_rrinterval, 0, "");
1891 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1892 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1893 OID_AUTO, "decay", CTLFLAG_RW,
1894 &usched_dfly_decay, 0, "Extra decay when not running");
1895 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1896 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1897 OID_AUTO, "batch_time", CTLFLAG_RW,
1898 &usched_dfly_batch_time, 0, "Min batch counter value");
1900 /* Add enable/disable option for SMT scheduling if supported */
1901 if (smt_not_supported) {
1902 usched_dfly_smt = 0;
1903 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
1904 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1905 OID_AUTO, "smt", CTLFLAG_RD,
1906 "NOT SUPPORTED", 0, "SMT NOT SUPPORTED");
1908 usched_dfly_smt = 1;
1909 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1910 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1911 OID_AUTO, "smt", CTLFLAG_RW,
1912 &usched_dfly_smt, 0, "Enable SMT scheduling");
1916 * Add enable/disable option for cache coherent scheduling
1919 if (cache_coherent_not_supported) {
1920 usched_dfly_cache_coherent = 0;
1921 SYSCTL_ADD_STRING(&usched_dfly_sysctl_ctx,
1922 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1923 OID_AUTO, "cache_coherent", CTLFLAG_RD,
1925 "Cache coherence NOT SUPPORTED");
1927 usched_dfly_cache_coherent = 1;
1928 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1929 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1930 OID_AUTO, "cache_coherent", CTLFLAG_RW,
1931 &usched_dfly_cache_coherent, 0,
1932 "Enable/Disable cache coherent scheduling");
1934 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1935 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1936 OID_AUTO, "weight1", CTLFLAG_RW,
1937 &usched_dfly_weight1, 10,
1938 "Weight selection for current cpu");
1940 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1941 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1942 OID_AUTO, "weight2", CTLFLAG_RW,
1943 &usched_dfly_weight2, 5,
1944 "Weight selection for wakefrom cpu");
1946 SYSCTL_ADD_PROC(&usched_dfly_sysctl_ctx,
1947 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1948 OID_AUTO, "stick_to_level",
1949 CTLTYPE_INT | CTLFLAG_RW,
1950 NULL, sizeof usched_dfly_stick_to_level,
1951 sysctl_usched_dfly_stick_to_level, "I",
1952 "Stick a process to this level. See sysctl"
1953 "paremter hw.cpu_topology.level_description");
1956 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1957 dfly_helper_thread_cpu_init, NULL)
1959 #else /* No SMP options - just add the configurable parameters to sysctl */
1962 sched_sysctl_tree_init(void)
1964 sysctl_ctx_init(&usched_dfly_sysctl_ctx);
1965 usched_dfly_sysctl_tree =
1966 SYSCTL_ADD_NODE(&usched_dfly_sysctl_ctx,
1967 SYSCTL_STATIC_CHILDREN(_kern), OID_AUTO,
1968 "usched_dfly", CTLFLAG_RD, 0, "");
1970 /* usched_dfly sysctl configurable parameters */
1971 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1972 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1973 OID_AUTO, "rrinterval", CTLFLAG_RW,
1974 &usched_dfly_rrinterval, 0, "");
1975 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1976 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1977 OID_AUTO, "decay", CTLFLAG_RW,
1978 &usched_dfly_decay, 0, "Extra decay when not running");
1979 SYSCTL_ADD_INT(&usched_dfly_sysctl_ctx,
1980 SYSCTL_CHILDREN(usched_dfly_sysctl_tree),
1981 OID_AUTO, "batch_time", CTLFLAG_RW,
1982 &usched_dfly_batch_time, 0, "Min batch counter value");
1984 SYSINIT(uschedtd, SI_BOOT2_USCHED, SI_ORDER_SECOND,
1985 sched_sysctl_tree_init, NULL)