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 static DPCPU_DEFINE(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_cnt(runs, usermode);
191 while (now >= state->nextstat) {
192 state->nextstat += statperiod;
195 if (runs && fake < 2) {
196 statclock_cnt(runs, usermode);
201 while (now >= state->nextprof) {
202 state->nextprof += profperiod;
206 profclock_cnt(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 hardfreq = (u_int)hz / 2;
243 if (tc_min_ticktock_freq > 2
245 && curcpu == CPU_FIRST()
248 hardfreq = hz / tc_min_ticktock_freq;
250 event += tick_sbt * (hardfreq - 1);
252 /* Handle callout events. */
253 if (event > state->nextcall)
254 event = state->nextcall;
255 if (!idle) { /* If CPU is active - handle other types of events. */
256 if (event > state->nextstat)
257 event = state->nextstat;
258 if (profiling && event > state->nextprof)
259 event = state->nextprof;
265 * Schedule binuptime of the next event on all CPUs.
270 struct pcpu_state *state;
277 state = DPCPU_PTR(timerstate);
278 event = state->nextevent;
281 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
283 state = DPCPU_ID_PTR(cpu, timerstate);
284 if (event > state->nextevent) {
285 event = state->nextevent;
291 CTR4(KTR_SPARE2, "next at %d: next %d.%08x by %d",
292 curcpu, (int)(event >> 32), (u_int)(event & 0xffffffff), c);
296 /* Hardware timer callback function. */
298 timercb(struct eventtimer *et, void *arg)
302 struct pcpu_state *state;
307 /* Do not touch anything if somebody reconfiguring timers. */
310 /* Update present and next tick times. */
311 state = DPCPU_PTR(timerstate);
312 if (et->et_flags & ET_FLAGS_PERCPU) {
313 next = &state->nexttick;
318 *next = now + timerperiod;
320 *next = -1; /* Next tick is not scheduled yet. */
322 CTR3(KTR_SPARE2, "intr at %d: now %d.%08x",
323 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
326 #ifdef EARLY_AP_STARTUP
327 MPASS(mp_ncpus == 1 || smp_started);
329 /* Prepare broadcasting to other CPUs for non-per-CPU timers. */
331 #ifdef EARLY_AP_STARTUP
332 if ((et->et_flags & ET_FLAGS_PERCPU) == 0) {
334 if ((et->et_flags & ET_FLAGS_PERCPU) == 0 && smp_started) {
337 state = DPCPU_ID_PTR(cpu, timerstate);
340 if (now >= state->nextevent) {
341 state->nextevent += SBT_1S;
352 /* Handle events for this time on this CPU. */
353 handleevents(now, 0);
356 /* Broadcast interrupt to other CPUs for non-per-CPU timers. */
361 state = DPCPU_ID_PTR(cpu, timerstate);
364 ipi_cpu(cpu, IPI_HARDCLOCK);
372 * Load new value into hardware timer.
375 loadtimer(sbintime_t now, int start)
377 struct pcpu_state *state;
383 if (timer->et_flags & ET_FLAGS_PERCPU) {
384 state = DPCPU_PTR(timerstate);
385 next = &state->nexttick;
391 * Try to start all periodic timers aligned
392 * to period to make events synchronous.
394 tmp = now % timerperiod;
395 new = timerperiod - tmp;
396 if (new < tmp) /* Left less then passed. */
398 CTR5(KTR_SPARE2, "load p at %d: now %d.%08x first in %d.%08x",
399 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff),
400 (int)(new >> 32), (u_int)(new & 0xffffffff));
402 et_start(timer, new, timerperiod);
405 new = getnextevent();
407 CTR4(KTR_SPARE2, "load at %d: next %d.%08x eq %d",
408 curcpu, (int)(new >> 32), (u_int)(new & 0xffffffff), eq);
411 et_start(timer, new - now, 0);
417 * Prepare event timer parameters after configuration changes.
424 if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
426 else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
428 singlemul = MIN(MAX(singlemul, 1), 20);
429 freq = hz * singlemul;
430 while (freq < (profiling ? profhz : stathz))
432 freq = round_freq(timer, freq);
433 timerperiod = SBT_1S / freq;
437 * Reconfigure specified per-CPU timer on other CPU. Called from IPI handler.
443 struct pcpu_state *state;
445 state = DPCPU_PTR(timerstate);
446 switch (atomic_load_acq_int(&state->action)) {
453 atomic_store_rel_int(&state->action, 0);
460 atomic_store_rel_int(&state->action, 0);
463 if (atomic_readandclear_int(&state->handle) && !busy) {
465 handleevents(now, 0);
472 * Reconfigure specified timer.
473 * For per-CPU timers use IPI to make other CPUs to reconfigure.
476 configtimer(int start)
478 sbintime_t now, next;
479 struct pcpu_state *state;
488 ET_HW_LOCK(DPCPU_PTR(timerstate));
490 /* Initialize time machine parameters. */
491 next = now + timerperiod;
496 #ifdef EARLY_AP_STARTUP
497 MPASS(mp_ncpus == 1 || smp_started);
500 state = DPCPU_ID_PTR(cpu, timerstate);
502 #ifndef EARLY_AP_STARTUP
503 if (!smp_started && cpu != CPU_FIRST())
504 state->nextevent = SBT_MAX;
507 state->nextevent = next;
509 state->nexttick = next;
511 state->nexttick = -1;
512 state->nexthard = next;
513 state->nextstat = next;
514 state->nextprof = next;
515 state->nextcall = next;
516 state->nextcallopt = next;
520 /* Start global timer or per-CPU timer of this CPU. */
524 /* Stop global timer or per-CPU timer of this CPU. */
527 ET_HW_UNLOCK(DPCPU_PTR(timerstate));
529 #ifdef EARLY_AP_STARTUP
530 /* If timer is global we are done. */
531 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0) {
533 /* If timer is global or there is no other CPUs yet - we are done. */
534 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || !smp_started) {
539 /* Set reconfigure flags for other CPUs. */
541 state = DPCPU_ID_PTR(cpu, timerstate);
542 atomic_store_rel_int(&state->action,
543 (cpu == curcpu) ? 0 : ( start ? 1 : 2));
545 /* Broadcast reconfigure IPI. */
546 ipi_all_but_self(IPI_HARDCLOCK);
547 /* Wait for reconfiguration completed. */
553 state = DPCPU_ID_PTR(cpu, timerstate);
554 if (atomic_load_acq_int(&state->action))
562 * Calculate nearest frequency supported by hardware timer.
565 round_freq(struct eventtimer *et, int freq)
569 if (et->et_frequency != 0) {
570 div = lmax((et->et_frequency + freq / 2) / freq, 1);
571 if (et->et_flags & ET_FLAGS_POW2DIV)
572 div = 1 << (flsl(div + div / 2) - 1);
573 freq = (et->et_frequency + div / 2) / div;
575 if (et->et_min_period > SBT_1S)
576 panic("Event timer \"%s\" doesn't support sub-second periods!",
578 else if (et->et_min_period != 0)
579 freq = min(freq, SBT2FREQ(et->et_min_period));
580 if (et->et_max_period < SBT_1S && et->et_max_period != 0)
581 freq = max(freq, SBT2FREQ(et->et_max_period));
586 * Configure and start event timers (BSP part).
589 cpu_initclocks_bsp(void)
591 struct pcpu_state *state;
594 mtx_init(&et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
596 state = DPCPU_ID_PTR(cpu, timerstate);
597 mtx_init(&state->et_hw_mtx, "et_hw_mtx", NULL, MTX_SPIN);
598 state->nextcall = SBT_MAX;
599 state->nextcallopt = SBT_MAX;
601 periodic = want_periodic;
602 /* Grab requested timer or the best of present. */
604 timer = et_find(timername, 0, 0);
605 if (timer == NULL && periodic) {
606 timer = et_find(NULL,
607 ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
610 timer = et_find(NULL,
611 ET_FLAGS_ONESHOT, ET_FLAGS_ONESHOT);
613 if (timer == NULL && !periodic) {
614 timer = et_find(NULL,
615 ET_FLAGS_PERIODIC, ET_FLAGS_PERIODIC);
618 panic("No usable event timer found!");
619 et_init(timer, timercb, NULL, NULL);
621 /* Adapt to timer capabilities. */
622 if (periodic && (timer->et_flags & ET_FLAGS_PERIODIC) == 0)
624 else if (!periodic && (timer->et_flags & ET_FLAGS_ONESHOT) == 0)
626 if (timer->et_flags & ET_FLAGS_C3STOP)
627 cpu_disable_c3_sleep++;
630 * We honor the requested 'hz' value.
631 * We want to run stathz in the neighborhood of 128hz.
632 * We would like profhz to run as often as possible.
634 if (singlemul <= 0 || singlemul > 20) {
635 if (hz >= 1500 || (hz % 128) == 0)
643 base = round_freq(timer, hz * singlemul);
644 singlemul = max((base + hz / 2) / hz, 1);
645 hz = (base + singlemul / 2) / singlemul;
650 if (div >= singlemul && (div % singlemul) == 0)
655 while ((profhz + stathz) <= 128 * 64)
657 profhz = round_freq(timer, profhz);
659 hz = round_freq(timer, hz);
660 stathz = round_freq(timer, 127);
661 profhz = round_freq(timer, stathz * 64);
664 tick_sbt = SBT_1S / hz;
665 tick_bt = sbttobt(tick_sbt);
666 statperiod = SBT_1S / stathz;
667 profperiod = SBT_1S / profhz;
674 * Start per-CPU event timers on APs.
677 cpu_initclocks_ap(void)
680 struct pcpu_state *state;
683 state = DPCPU_PTR(timerstate);
687 hardclock_sync(curcpu);
691 td->td_intr_nesting_level++;
692 handleevents(state->now, 2);
693 td->td_intr_nesting_level--;
698 * Switch to profiling clock rates.
701 cpu_startprofclock(void)
705 if (profiling == 0) {
718 * Switch to regular clock rates.
721 cpu_stopprofclock(void)
725 if (profiling == 1) {
738 * Switch to idle mode (all ticks handled).
744 struct pcpu_state *state;
746 if (idletick || busy ||
747 (periodic && (timer->et_flags & ET_FLAGS_PERCPU))
748 #ifdef DEVICE_POLLING
749 || curcpu == CPU_FIRST()
753 state = DPCPU_PTR(timerstate);
758 CTR3(KTR_SPARE2, "idle at %d: now %d.%08x",
759 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
760 t = getnextcpuevent(1);
763 state->nextevent = t;
767 return (MAX(t - now, 0));
771 * Switch to active mode (skip empty ticks).
774 cpu_activeclock(void)
777 struct pcpu_state *state;
780 state = DPCPU_PTR(timerstate);
781 if (state->idle == 0 || busy)
787 CTR3(KTR_SPARE2, "active at %d: now %d.%08x",
788 curcpu, (int)(now >> 32), (u_int)(now & 0xffffffff));
791 td->td_intr_nesting_level++;
792 handleevents(now, 1);
793 td->td_intr_nesting_level--;
798 * Change the frequency of the given timer. This changes et->et_frequency and
799 * if et is the active timer it reconfigures the timer on all CPUs. This is
800 * intended to be a private interface for the use of et_change_frequency() only.
803 cpu_et_frequency(struct eventtimer *et, uint64_t newfreq)
809 et->et_frequency = newfreq;
812 et->et_frequency = newfreq;
817 cpu_new_callout(int cpu, sbintime_t bt, sbintime_t bt_opt)
819 struct pcpu_state *state;
821 /* Do not touch anything if somebody reconfiguring timers. */
824 CTR6(KTR_SPARE2, "new co at %d: on %d at %d.%08x - %d.%08x",
825 curcpu, cpu, (int)(bt_opt >> 32), (u_int)(bt_opt & 0xffffffff),
826 (int)(bt >> 32), (u_int)(bt & 0xffffffff));
828 KASSERT(!CPU_ABSENT(cpu), ("Absent CPU %d", cpu));
829 state = DPCPU_ID_PTR(cpu, timerstate);
833 * If there is callout time already set earlier -- do nothing.
834 * This check may appear redundant because we check already in
835 * callout_process() but this double check guarantees we're safe
836 * with respect to race conditions between interrupts execution
839 state->nextcallopt = bt_opt;
840 if (bt >= state->nextcall)
842 state->nextcall = bt;
843 /* If there is some other event set earlier -- do nothing. */
844 if (bt >= state->nextevent)
846 state->nextevent = bt;
847 /* If timer is periodic -- there is nothing to reprogram. */
850 /* If timer is global or of the current CPU -- reprogram it. */
851 if ((timer->et_flags & ET_FLAGS_PERCPU) == 0 || cpu == curcpu) {
852 loadtimer(sbinuptime(), 0);
857 /* Otherwise make other CPU to reprogram it. */
861 ipi_cpu(cpu, IPI_HARDCLOCK);
866 * Report or change the active event timers hardware.
869 sysctl_kern_eventtimer_timer(SYSCTL_HANDLER_ARGS)
872 struct eventtimer *et;
877 snprintf(buf, sizeof(buf), "%s", et->et_name);
879 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
882 if (error != 0 || req->newptr == NULL ||
883 strcasecmp(buf, et->et_name) == 0) {
887 et = et_find(buf, 0, 0);
894 if (et->et_flags & ET_FLAGS_C3STOP)
895 cpu_disable_c3_sleep++;
896 if (timer->et_flags & ET_FLAGS_C3STOP)
897 cpu_disable_c3_sleep--;
898 periodic = want_periodic;
900 et_init(timer, timercb, NULL, NULL);
905 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, timer,
906 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
907 0, 0, sysctl_kern_eventtimer_timer, "A", "Chosen event timer");
910 * Report or change the active event timer periodicity.
913 sysctl_kern_eventtimer_periodic(SYSCTL_HANDLER_ARGS)
918 error = sysctl_handle_int(oidp, &val, 0, req);
919 if (error != 0 || req->newptr == NULL)
923 periodic = want_periodic = val;
928 SYSCTL_PROC(_kern_eventtimer, OID_AUTO, periodic,
929 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
930 0, 0, sysctl_kern_eventtimer_periodic, "I", "Enable event timer periodic mode");
937 DB_SHOW_COMMAND(clocksource, db_show_clocksource)
939 struct pcpu_state *st;
943 st = DPCPU_ID_PTR(c, timerstate);
945 "CPU %2d: action %d handle %d ipi %d idle %d\n"
946 " now %#jx nevent %#jx (%jd)\n"
947 " ntick %#jx (%jd) nhard %#jx (%jd)\n"
948 " nstat %#jx (%jd) nprof %#jx (%jd)\n"
949 " ncall %#jx (%jd) ncallopt %#jx (%jd)\n",
950 c, st->action, st->handle, st->ipi, st->idle,
952 (uintmax_t)st->nextevent,
953 (uintmax_t)(st->nextevent - st->now) / tick_sbt,
954 (uintmax_t)st->nexttick,
955 (uintmax_t)(st->nexttick - st->now) / tick_sbt,
956 (uintmax_t)st->nexthard,
957 (uintmax_t)(st->nexthard - st->now) / tick_sbt,
958 (uintmax_t)st->nextstat,
959 (uintmax_t)(st->nextstat - st->now) / tick_sbt,
960 (uintmax_t)st->nextprof,
961 (uintmax_t)(st->nextprof - st->now) / tick_sbt,
962 (uintmax_t)st->nextcall,
963 (uintmax_t)(st->nextcall - st->now) / tick_sbt,
964 (uintmax_t)st->nextcallopt,
965 (uintmax_t)(st->nextcallopt - st->now) / tick_sbt);