2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2010-2013 Alexander Motin <mav@FreeBSD.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer,
12 * without modification, immediately at the beginning of the file.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
33 * Common routines to manage event timers hardware.
36 #include "opt_device_polling.h"
38 #include <sys/param.h>
39 #include <sys/systm.h>
41 #include <sys/limits.h>
45 #include <sys/mutex.h>
47 #include <sys/kernel.h>
48 #include <sys/sched.h>
50 #include <sys/sysctl.h>
51 #include <sys/timeet.h>
52 #include <sys/timetc.h>
54 #include <machine/atomic.h>
55 #include <machine/clock.h>
56 #include <machine/cpu.h>
57 #include <machine/smp.h>
59 int cpu_disable_c2_sleep = 0; /* Timer dies in C2. */
60 int cpu_disable_c3_sleep = 0; /* Timer dies in C3. */
62 static void setuptimer(void);
63 static void loadtimer(sbintime_t now, int first);
64 static int doconfigtimer(void);
65 static void configtimer(int start);
66 static int round_freq(struct eventtimer *et, int freq);
68 static sbintime_t getnextcpuevent(int idle);
69 static sbintime_t getnextevent(void);
70 static int handleevents(sbintime_t now, int fake);
72 static struct mtx et_hw_mtx;
74 #define ET_HW_LOCK(state) \
76 if (timer->et_flags & ET_FLAGS_PERCPU) \
77 mtx_lock_spin(&(state)->et_hw_mtx); \
79 mtx_lock_spin(&et_hw_mtx); \
82 #define ET_HW_UNLOCK(state) \
84 if (timer->et_flags & ET_FLAGS_PERCPU) \
85 mtx_unlock_spin(&(state)->et_hw_mtx); \
87 mtx_unlock_spin(&et_hw_mtx); \
90 static struct eventtimer *timer = NULL;
91 static sbintime_t timerperiod; /* Timer period for periodic mode. */
92 static sbintime_t statperiod; /* statclock() events period. */
93 static sbintime_t profperiod; /* profclock() events period. */
94 static sbintime_t nexttick; /* Next global timer tick time. */
95 static u_int busy = 1; /* Reconfiguration is in progress. */
96 static int profiling; /* Profiling events enabled. */
98 static char timername[32]; /* Wanted timer. */
99 TUNABLE_STR("kern.eventtimer.timer", timername, sizeof(timername));
101 static int singlemul; /* Multiplier for periodic mode. */
102 SYSCTL_INT(_kern_eventtimer, OID_AUTO, singlemul, CTLFLAG_RWTUN, &singlemul,
103 0, "Multiplier for periodic mode");
105 static u_int idletick; /* Run periodic events when idle. */
106 SYSCTL_UINT(_kern_eventtimer, OID_AUTO, idletick, CTLFLAG_RWTUN, &idletick,
107 0, "Run periodic events when idle");
109 static int periodic; /* Periodic or one-shot mode. */
110 static int want_periodic; /* What mode to prefer. */
111 TUNABLE_INT("kern.eventtimer.periodic", &want_periodic);
114 struct mtx et_hw_mtx; /* Per-CPU timer mutex. */
115 u_int action; /* Reconfiguration requests. */
116 u_int handle; /* Immediate handle resuests. */
117 sbintime_t now; /* Last tick time. */
118 sbintime_t nextevent; /* Next scheduled event on this CPU. */
119 sbintime_t nexttick; /* Next timer tick time. */
120 sbintime_t nexthard; /* Next hardclock() event. */
121 sbintime_t nextstat; /* Next statclock() event. */
122 sbintime_t nextprof; /* Next profclock() event. */
123 sbintime_t nextcall; /* Next callout event. */
124 sbintime_t nextcallopt; /* Next optional callout event. */
125 int ipi; /* This CPU needs IPI. */
126 int idle; /* This CPU is in idle mode. */
129 DPCPU_DEFINE_STATIC(struct pcpu_state, timerstate);
130 DPCPU_DEFINE(sbintime_t, hardclocktime);
133 * Timer broadcast IPI handler.
139 struct pcpu_state *state;
142 if (doconfigtimer() || busy)
143 return (FILTER_HANDLED);
144 state = DPCPU_PTR(timerstate);
146 CTR3(KTR_SPARE2, "ipi at %d: now %d.%08x",
147 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
148 done = handleevents(now, 0);
149 return (done ? FILTER_HANDLED : FILTER_STRAY);
153 * Handle all events for specified time on this CPU
156 handleevents(sbintime_t now, int fake)
159 struct trapframe *frame;
160 struct pcpu_state *state;
164 CTR3(KTR_SPARE2, "handle at %d: now %d.%08x",
165 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
171 frame = curthread->td_intr_frame;
172 usermode = TRAPF_USERMODE(frame);
175 state = DPCPU_PTR(timerstate);
178 while (now >= state->nexthard) {
179 state->nexthard += tick_sbt;
183 hct = DPCPU_PTR(hardclocktime);
184 *hct = state->nexthard - tick_sbt;
186 hardclock(runs, usermode);
191 while (now >= state->nextstat) {
192 state->nextstat += statperiod;
195 if (runs && fake < 2) {
196 statclock(runs, usermode);
201 while (now >= state->nextprof) {
202 state->nextprof += profperiod;
206 profclock(runs, usermode, TRAPF_PC(frame));
210 state->nextprof = state->nextstat;
211 if (now >= state->nextcallopt || now >= state->nextcall) {
212 state->nextcall = state->nextcallopt = SBT_MAX;
213 callout_process(now);
216 t = getnextcpuevent(0);
220 state->nextevent = t;
221 loadtimer(now, (fake == 2) &&
222 (timer->et_flags & ET_FLAGS_PERCPU));
229 * Schedule binuptime of the next event on current CPU.
232 getnextcpuevent(int idle)
235 struct pcpu_state *state;
238 state = DPCPU_PTR(timerstate);
239 /* Handle hardclock() events, skipping some if CPU is idle. */
240 event = state->nexthard;
242 if (tc_min_ticktock_freq > 1
244 && curcpu == CPU_FIRST()
247 hardfreq = hz / tc_min_ticktock_freq;
251 event += tick_sbt * (hardfreq - 1);
253 /* Handle callout events. */
254 if (event > state->nextcall)
255 event = state->nextcall;
256 if (!idle) { /* If CPU is active - handle other types of events. */
257 if (event > state->nextstat)
258 event = state->nextstat;
259 if (profiling && event > state->nextprof)
260 event = state->nextprof;
266 * Schedule binuptime of the next event on all CPUs.
271 struct pcpu_state *state;
281 state = DPCPU_PTR(timerstate);
282 event = state->nextevent;
284 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
286 state = DPCPU_ID_PTR(cpu, timerstate);
287 if (event > state->nextevent) {
288 event = state->nextevent;
296 CTR4(KTR_SPARE2, "next at %d: next %d.%08x by %d",
297 curcpu, (int)(event >> 32), (u_int)(event & 0xffffffff), c);
301 /* Hardware timer callback function. */
303 timercb(struct eventtimer *et, void *arg)
307 struct pcpu_state *state;
312 /* Do not touch anything if somebody reconfiguring timers. */
315 /* Update present and next tick times. */
316 state = DPCPU_PTR(timerstate);
317 if (et->et_flags & ET_FLAGS_PERCPU) {
318 next = &state->nexttick;
323 *next = now + timerperiod;
325 *next = -1; /* Next tick is not scheduled yet. */
327 CTR3(KTR_SPARE2, "intr at %d: now %d.%08x",
328 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
331 #ifdef EARLY_AP_STARTUP
332 MPASS(mp_ncpus == 1 || smp_started);
334 /* Prepare broadcasting to other CPUs for non-per-CPU timers. */
336 #ifdef EARLY_AP_STARTUP
337 if ((et->et_flags & ET_FLAGS_PERCPU) == 0) {
339 if ((et->et_flags & ET_FLAGS_PERCPU) == 0 && smp_started) {
342 state = DPCPU_ID_PTR(cpu, timerstate);
345 if (now >= state->nextevent) {
346 state->nextevent += SBT_1S;
357 /* Handle events for this time on this CPU. */
358 handleevents(now, 0);
361 /* Broadcast interrupt to other CPUs for non-per-CPU timers. */
366 state = DPCPU_ID_PTR(cpu, timerstate);
369 ipi_cpu(cpu, IPI_HARDCLOCK);
377 * Load new value into hardware timer.
380 loadtimer(sbintime_t now, int start)
382 struct pcpu_state *state;
388 if (timer->et_flags & ET_FLAGS_PERCPU) {
389 state = DPCPU_PTR(timerstate);
390 next = &state->nexttick;
396 * Try to start all periodic timers aligned
397 * to period to make events synchronous.
399 tmp = now % timerperiod;
400 new = timerperiod - tmp;
401 if (new < tmp) /* Left less then passed. */
403 CTR5(KTR_SPARE2, "load p at %d: now %d.%08x first in %d.%08x",
404 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff),
405 (int)(new >> 32), (u_int)(new & 0xffffffff));
407 et_start(timer, new, timerperiod);
410 new = getnextevent();
412 CTR4(KTR_SPARE2, "load at %d: next %d.%08x eq %d",
413 curcpu, (int)(new >> 32), (u_int)(new & 0xffffffff), eq);
416 et_start(timer, new - now, 0);
422 * Prepare event timer parameters after configuration changes.
429 if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
431 else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
433 singlemul = MIN(MAX(singlemul, 1), 20);
434 freq = hz * singlemul;
435 while (freq < (profiling ? profhz : stathz))
437 freq = round_freq(timer, freq);
438 timerperiod = SBT_1S / freq;
442 * Reconfigure specified per-CPU timer on other CPU. Called from IPI handler.
448 struct pcpu_state *state;
450 state = DPCPU_PTR(timerstate);
451 switch (atomic_load_acq_int(&state->action)) {
458 atomic_store_rel_int(&state->action, 0);
465 atomic_store_rel_int(&state->action, 0);
468 if (atomic_readandclear_int(&state->handle) && !busy) {
470 handleevents(now, 0);
477 * Reconfigure specified timer.
478 * For per-CPU timers use IPI to make other CPUs to reconfigure.
481 configtimer(int start)
483 sbintime_t now, next;
484 struct pcpu_state *state;
493 ET_HW_LOCK(DPCPU_PTR(timerstate));
495 /* Initialize time machine parameters. */
496 next = now + timerperiod;
501 #ifdef EARLY_AP_STARTUP
502 MPASS(mp_ncpus == 1 || smp_started);
505 state = DPCPU_ID_PTR(cpu, timerstate);
507 #ifndef EARLY_AP_STARTUP
508 if (!smp_started && cpu != CPU_FIRST())
509 state->nextevent = SBT_MAX;
512 state->nextevent = next;
514 state->nexttick = next;
516 state->nexttick = -1;
517 state->nexthard = next;
518 state->nextstat = next;
519 state->nextprof = next;
520 state->nextcall = next;
521 state->nextcallopt = next;
525 /* Start global timer or per-CPU timer of this CPU. */
529 /* Stop global timer or per-CPU timer of this CPU. */
532 ET_HW_UNLOCK(DPCPU_PTR(timerstate));
534 #ifdef EARLY_AP_STARTUP
535 /* If timer is global we are done. */
536 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
538 /* If timer is global or there is no other CPUs yet - we are done. */
539 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || !smp_started) {
544 /* Set reconfigure flags for other CPUs. */
546 state = DPCPU_ID_PTR(cpu, timerstate);
547 atomic_store_rel_int(&state->action,
548 (cpu == curcpu) ? 0 : ( start ? 1 : 2));
550 /* Broadcast reconfigure IPI. */
551 ipi_all_but_self(IPI_HARDCLOCK);
552 /* Wait for reconfiguration completed. */
558 state = DPCPU_ID_PTR(cpu, timerstate);
559 if (atomic_load_acq_int(&state->action))
567 * Calculate nearest frequency supported by hardware timer.
570 round_freq(struct eventtimer *et, int freq)
574 if (et->et_frequency != 0) {
575 div = lmax((et->et_frequency + freq / 2) / freq, 1);
576 if (et->et_flags & ET_FLAGS_POW2DIV)
577 div = 1 << (flsl(div + div / 2) - 1);
578 freq = (et->et_frequency + div / 2) / div;
580 if (et->et_min_period > SBT_1S)
581 panic("Event timer \"%s\" doesn't support sub-second periods!",
583 else if (et->et_min_period != 0)
584 freq = min(freq, SBT2FREQ(et->et_min_period));
585 if (et->et_max_period < SBT_1S && et->et_max_period != 0)
586 freq = max(freq, SBT2FREQ(et->et_max_period));
591 * Configure and start event timers (BSP part).
594 cpu_initclocks_bsp(void)
596 struct pcpu_state *state;
599 mtx_init(&et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
601 state = DPCPU_ID_PTR(cpu, timerstate);
602 mtx_init(&state->et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
603 state->nextcall = SBT_MAX;
604 state->nextcallopt = SBT_MAX;
606 periodic = want_periodic;
607 /* Grab requested timer or the best of present. */
609 timer = et_find(timername, 0, 0);
610 if (timer == NULL && periodic) {
611 timer = et_find(NULL,
612 ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
615 timer = et_find(NULL,
616 ET_FLAGS_ONESHOT, ET_FLAGS_ONESHOT);
618 if (timer == NULL && !periodic) {
619 timer = et_find(NULL,
620 ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
623 panic("No usable event timer found!");
624 et_init(timer, timercb, NULL, NULL);
626 /* Adapt to timer capabilities. */
627 if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
629 else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
631 if (timer->et_flags & ET_FLAGS_C3STOP)
632 cpu_disable_c3_sleep++;
635 * We honor the requested 'hz' value.
636 * We want to run stathz in the neighborhood of 128hz.
637 * We would like profhz to run as often as possible.
639 if (singlemul <= 0 || singlemul > 20) {
640 if (hz >= 1500 || (hz % 128) == 0)
648 base = round_freq(timer, hz * singlemul);
649 singlemul = max((base + hz / 2) / hz, 1);
650 hz = (base + singlemul / 2) / singlemul;
655 if (div >= singlemul && (div % singlemul) == 0)
660 while ((profhz + stathz) <= 128 * 64)
662 profhz = round_freq(timer, profhz);
664 hz = round_freq(timer, hz);
665 stathz = round_freq(timer, 127);
666 profhz = round_freq(timer, stathz * 64);
669 tick_sbt = SBT_1S / hz;
670 tick_bt = sbttobt(tick_sbt);
671 statperiod = SBT_1S / stathz;
672 profperiod = SBT_1S / profhz;
679 * Start per-CPU event timers on APs.
682 cpu_initclocks_ap(void)
685 struct pcpu_state *state;
688 state = DPCPU_PTR(timerstate);
692 hardclock_sync(curcpu);
696 td->td_intr_nesting_level++;
697 handleevents(state->now, 2);
698 td->td_intr_nesting_level--;
719 * Switch to profiling clock rates.
722 cpu_startprofclock(void)
726 if (profiling == 0) {
739 * Switch to regular clock rates.
742 cpu_stopprofclock(void)
746 if (profiling == 1) {
759 * Switch to idle mode (all ticks handled).
765 struct pcpu_state *state;
767 if (idletick || busy ||
768 (periodic && (timer->et_flags & ET_FLAGS_PERCPU))
769 #ifdef DEVICE_POLLING
770 || curcpu == CPU_FIRST()
774 state = DPCPU_PTR(timerstate);
779 CTR3(KTR_SPARE2, "idle at %d: now %d.%08x",
780 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
781 t = getnextcpuevent(1);
784 state->nextevent = t;
788 return (MAX(t - now, 0));
792 * Switch to active mode (skip empty ticks).
795 cpu_activeclock(void)
798 struct pcpu_state *state;
801 state = DPCPU_PTR(timerstate);
802 if (state->idle == 0 || busy)
808 CTR3(KTR_SPARE2, "active at %d: now %d.%08x",
809 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
812 td->td_intr_nesting_level++;
813 handleevents(now, 1);
814 td->td_intr_nesting_level--;
819 * Change the frequency of the given timer. This changes et->et_frequency and
820 * if et is the active timer it reconfigures the timer on all CPUs. This is
821 * intended to be a private interface for the use of et_change_frequency() only.
824 cpu_et_frequency(struct eventtimer *et, uint64_t newfreq)
830 et->et_frequency = newfreq;
833 et->et_frequency = newfreq;
838 cpu_new_callout(int cpu, sbintime_t bt, sbintime_t bt_opt)
840 struct pcpu_state *state;
842 /* Do not touch anything if somebody reconfiguring timers. */
845 CTR6(KTR_SPARE2, "new co at %d: on %d at %d.%08x - %d.%08x",
846 curcpu, cpu, (int)(bt_opt >> 32), (u_int)(bt_opt & 0xffffffff),
847 (int)(bt >> 32), (u_int)(bt & 0xffffffff));
849 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
850 state = DPCPU_ID_PTR(cpu, timerstate);
854 * If there is callout time already set earlier -- do nothing.
855 * This check may appear redundant because we check already in
856 * callout_process() but this double check guarantees we're safe
857 * with respect to race conditions between interrupts execution
860 state->nextcallopt = bt_opt;
861 if (bt >= state->nextcall)
863 state->nextcall = bt;
864 /* If there is some other event set earlier -- do nothing. */
865 if (bt >= state->nextevent)
867 state->nextevent = bt;
868 /* If timer is periodic -- there is nothing to reprogram. */
871 /* If timer is global or of the current CPU -- reprogram it. */
872 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || cpu == curcpu) {
873 loadtimer(sbinuptime(), 0);
878 /* Otherwise make other CPU to reprogram it. */
882 ipi_cpu(cpu, IPI_HARDCLOCK);
887 * Report or change the active event timers hardware.
890 sysctl_kern_eventtimer_timer(SYSCTL_HANDLER_ARGS)
893 struct eventtimer *et;
898 snprintf(buf, sizeof(buf), "%s", et->et_name);
900 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
903 if (error != 0 || req->newptr == NULL ||
904 strcasecmp(buf, et->et_name) == 0) {
908 et = et_find(buf, 0, 0);
915 if (et->et_flags & ET_FLAGS_C3STOP)
916 cpu_disable_c3_sleep++;
917 if (timer->et_flags & ET_FLAGS_C3STOP)
918 cpu_disable_c3_sleep--;
919 periodic = want_periodic;
921 et_init(timer, timercb, NULL, NULL);
926 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, timer,
927 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
928 0, 0, sysctl_kern_eventtimer_timer, "A", "Chosen event timer");
931 * Report or change the active event timer periodicity.
934 sysctl_kern_eventtimer_periodic(SYSCTL_HANDLER_ARGS)
939 error = sysctl_handle_int(oidp, &val, 0, req);
940 if (error != 0 || req->newptr == NULL)
944 periodic = want_periodic = val;
949 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, periodic,
950 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
951 0, 0, sysctl_kern_eventtimer_periodic, "I", "Enable event timer periodic mode");
958 DB_SHOW_COMMAND(clocksource, db_show_clocksource)
960 struct pcpu_state *st;
964 st = DPCPU_ID_PTR(c, timerstate);
966 "CPU %2d: action %d handle %d ipi %d idle %d\n"
967 " now %#jx nevent %#jx (%jd)\n"
968 " ntick %#jx (%jd) nhard %#jx (%jd)\n"
969 " nstat %#jx (%jd) nprof %#jx (%jd)\n"
970 " ncall %#jx (%jd) ncallopt %#jx (%jd)\n",
971 c, st->action, st->handle, st->ipi, st->idle,
973 (uintmax_t)st->nextevent,
974 (uintmax_t)(st->nextevent - st->now) / tick_sbt,
975 (uintmax_t)st->nexttick,
976 (uintmax_t)(st->nexttick - st->now) / tick_sbt,
977 (uintmax_t)st->nexthard,
978 (uintmax_t)(st->nexthard - st->now) / tick_sbt,
979 (uintmax_t)st->nextstat,
980 (uintmax_t)(st->nextstat - st->now) / tick_sbt,
981 (uintmax_t)st->nextprof,
982 (uintmax_t)(st->nextprof - st->now) / tick_sbt,
983 (uintmax_t)st->nextcall,
984 (uintmax_t)(st->nextcall - st->now) / tick_sbt,
985 (uintmax_t)st->nextcallopt,
986 (uintmax_t)(st->nextcallopt - st->now) / tick_sbt);