This patch switches arm64 to clockintr(9). This has survived about a dozen parallel kernel builds, parallel release builds, and bsd.rd upgrades on my RPi4b.
kettenis@ said he tried it on one of his M1 machines, too. It could use more testing, particularly on arm64 machines that aren't a Raspberry Pi. Notes: - The spec is a little unclear about whether or not the values written to CNTV_TVAL_EL0 are interpretted as signed or unsigned. To be safe, I have capped the cycle count at INT32_MAX. Index: sys/arch/arm64/dev/agtimer.c =================================================================== RCS file: /cvs/src/sys/arch/arm64/dev/agtimer.c,v retrieving revision 1.19 diff -u -p -r1.19 agtimer.c --- sys/arch/arm64/dev/agtimer.c 24 Oct 2021 17:52:28 -0000 1.19 +++ sys/arch/arm64/dev/agtimer.c 6 Nov 2022 18:48:22 -0000 @@ -18,10 +18,12 @@ #include <sys/param.h> #include <sys/systm.h> +#include <sys/clockintr.h> #include <sys/queue.h> #include <sys/malloc.h> #include <sys/device.h> #include <sys/kernel.h> +#include <sys/stdint.h> #include <sys/timetc.h> #include <sys/evcount.h> @@ -53,28 +55,13 @@ static struct timecounter agtimer_timeco .tc_user = TC_AGTIMER, }; -struct agtimer_pcpu_softc { - uint64_t pc_nexttickevent; - uint64_t pc_nextstatevent; - u_int32_t pc_ticks_err_sum; -}; - struct agtimer_softc { struct device sc_dev; int sc_node; - struct agtimer_pcpu_softc sc_pstat[MAXCPUS]; - - u_int32_t sc_ticks_err_cnt; u_int32_t sc_ticks_per_second; - u_int32_t sc_ticks_per_intr; - u_int32_t sc_statvar; - u_int32_t sc_statmin; - -#ifdef AMPTIMER_DEBUG - struct evcount sc_clk_count; - struct evcount sc_stat_count; -#endif + uint64_t sc_nsec_cycle_ratio; + uint64_t sc_nsec_max; void *sc_ih; }; @@ -84,9 +71,11 @@ uint64_t agtimer_readcnt64(void); int agtimer_intr(void *); void agtimer_cpu_initclocks(void); void agtimer_delay(u_int); +void agtimer_rearm(void *, uint64_t); void agtimer_setstatclockrate(int stathz); void agtimer_set_clockrate(int32_t new_frequency); void agtimer_startclock(void); +void agtimer_trigger(void *); const struct cfattach agtimer_ca = { sizeof (struct agtimer_softc), agtimer_match, agtimer_attach @@ -96,6 +85,11 @@ struct cfdriver agtimer_cd = { NULL, "agtimer", DV_DULL }; +struct intrclock agtimer_intrclock = { + .ic_rearm = agtimer_rearm, + .ic_trigger = agtimer_trigger +}; + uint64_t agtimer_readcnt64(void) { @@ -173,13 +167,11 @@ agtimer_attach(struct device *parent, st agtimer_frequency = OF_getpropint(sc->sc_node, "clock-frequency", agtimer_frequency); sc->sc_ticks_per_second = agtimer_frequency; + sc->sc_nsec_cycle_ratio = + sc->sc_ticks_per_second * (1ULL << 32) / 1000000000; + sc->sc_nsec_max = UINT64_MAX / sc->sc_nsec_cycle_ratio; - printf(": %d kHz\n", sc->sc_ticks_per_second / 1000); - -#ifdef AMPTIMER_DEBUG - evcount_attach(&sc->sc_clk_count, "clock", NULL); - evcount_attach(&sc->sc_stat_count, "stat", NULL); -#endif + printf(": %u kHz\n", sc->sc_ticks_per_second / 1000); /* * private timer and interrupts not enabled until @@ -191,8 +183,9 @@ agtimer_attach(struct device *parent, st agtimer_timecounter.tc_frequency = sc->sc_ticks_per_second; agtimer_timecounter.tc_priv = sc; - tc_init(&agtimer_timecounter); + + agtimer_intrclock.ic_cookie = sc; } u_int @@ -209,72 +202,30 @@ agtimer_get_timecount(struct timecounter return (val & 0xffffffff); } -int -agtimer_intr(void *frame) +void +agtimer_rearm(void *cookie, uint64_t nsecs) { - struct agtimer_softc *sc = agtimer_cd.cd_devs[0]; - struct agtimer_pcpu_softc *pc = &sc->sc_pstat[CPU_INFO_UNIT(curcpu())]; - uint64_t now; - uint64_t nextevent; - uint32_t r; -#if defined(USE_GTIMER_CMP) - int skip = 1; -#else - int64_t delay; -#endif - int rc = 0; - - /* - * DSR - I know that the tick timer is 64 bits, but the following - * code deals with rollover, so there is no point in dealing - * with the 64 bit math, just let the 32 bit rollover - * do the right thing - */ - - now = agtimer_readcnt64(); + struct agtimer_softc *sc = cookie; + uint32_t cycles; - while (pc->pc_nexttickevent <= now) { - pc->pc_nexttickevent += sc->sc_ticks_per_intr; - pc->pc_ticks_err_sum += sc->sc_ticks_err_cnt; - - /* looping a few times is faster than divide */ - while (pc->pc_ticks_err_sum > hz) { - pc->pc_nexttickevent += 1; - pc->pc_ticks_err_sum -= hz; - } - -#ifdef AMPTIMER_DEBUG - sc->sc_clk_count.ec_count++; -#endif - rc = 1; - hardclock(frame); - } - while (pc->pc_nextstatevent <= now) { - do { - r = random() & (sc->sc_statvar -1); - } while (r == 0); /* random == 0 not allowed */ - pc->pc_nextstatevent += sc->sc_statmin + r; - - /* XXX - correct nextstatevent? */ -#ifdef AMPTIMER_DEBUG - sc->sc_stat_count.ec_count++; -#endif - rc = 1; - statclock(frame); - } - - if (pc->pc_nexttickevent < pc->pc_nextstatevent) - nextevent = pc->pc_nexttickevent; - else - nextevent = pc->pc_nextstatevent; - - delay = nextevent - now; - if (delay < 0) - delay = 1; + if (nsecs > sc->sc_nsec_max) + nsecs = sc->sc_nsec_max; + cycles = (nsecs * sc->sc_nsec_cycle_ratio) >> 32; + if (cycles > INT32_MAX) + cycles = INT32_MAX; + agtimer_set_tval(cycles); +} - agtimer_set_tval(delay); +void +agtimer_trigger(void *unused) +{ + agtimer_set_tval(0); +} - return (rc); +int +agtimer_intr(void *frame) +{ + return clockintr_dispatch(frame); } void @@ -288,8 +239,13 @@ agtimer_set_clockrate(int32_t new_freque return; sc->sc_ticks_per_second = agtimer_frequency; + sc->sc_nsec_cycle_ratio = + sc->sc_ticks_per_second * (1ULL << 32) / 1000000000; + sc->sc_nsec_max = UINT64_MAX / sc->sc_nsec_cycle_ratio; + agtimer_timecounter.tc_frequency = sc->sc_ticks_per_second; - printf("agtimer0: adjusting clock: new tick rate %d kHz\n", + + printf("agtimer0: adjusting clock: new tick rate %u kHz\n", sc->sc_ticks_per_second / 1000); } @@ -297,37 +253,31 @@ void agtimer_cpu_initclocks(void) { struct agtimer_softc *sc = agtimer_cd.cd_devs[0]; - struct agtimer_pcpu_softc *pc = &sc->sc_pstat[CPU_INFO_UNIT(curcpu())]; uint32_t reg; - uint64_t next; uint64_t kctl; stathz = hz; - profhz = hz * 10; + profhz = stathz * 10; + clockintr_init(CL_RNDSTAT); if (sc->sc_ticks_per_second != agtimer_frequency) { agtimer_set_clockrate(agtimer_frequency); } - agtimer_setstatclockrate(stathz); - - sc->sc_ticks_per_intr = sc->sc_ticks_per_second / hz; - sc->sc_ticks_err_cnt = sc->sc_ticks_per_second % hz; - pc->pc_ticks_err_sum = 0; - /* configure virtual timer interrupt */ sc->sc_ih = arm_intr_establish_fdt_idx(sc->sc_node, 2, IPL_CLOCK|IPL_MPSAFE, agtimer_intr, NULL, "tick"); - next = agtimer_readcnt64() + sc->sc_ticks_per_intr; - pc->pc_nexttickevent = pc->pc_nextstatevent = next; + clockintr_cpu_init(&agtimer_intrclock); reg = agtimer_get_ctrl(); reg &= ~GTIMER_CNTV_CTL_IMASK; reg |= GTIMER_CNTV_CTL_ENABLE; - agtimer_set_tval(sc->sc_ticks_per_second); + agtimer_set_tval(INT32_MAX); agtimer_set_ctrl(reg); + clockintr_trigger(); + /* enable userland access to virtual counter */ kctl = READ_SPECIALREG(CNTKCTL_EL1); WRITE_SPECIALREG(CNTKCTL_EL1, kctl | CNTKCTL_EL0VCTEN); @@ -367,48 +317,27 @@ agtimer_delay(u_int usecs) void agtimer_setstatclockrate(int newhz) { - struct agtimer_softc *sc = agtimer_cd.cd_devs[0]; - int minint, statint; - int s; - - s = splclock(); - - statint = sc->sc_ticks_per_second / newhz; - /* calculate largest 2^n which is smaller that just over half statint */ - sc->sc_statvar = 0x40000000; /* really big power of two */ - minint = statint / 2 + 100; - while (sc->sc_statvar > minint) - sc->sc_statvar >>= 1; - - sc->sc_statmin = statint - (sc->sc_statvar >> 1); - - splx(s); - - /* - * XXX this allows the next stat timer to occur then it switches - * to the new frequency. Rather than switching instantly. - */ + clockintr_setstatclockrate(newhz); } void agtimer_startclock(void) { struct agtimer_softc *sc = agtimer_cd.cd_devs[0]; - struct agtimer_pcpu_softc *pc = &sc->sc_pstat[CPU_INFO_UNIT(curcpu())]; - uint64_t nextevent; uint64_t kctl; uint32_t reg; - nextevent = agtimer_readcnt64() + sc->sc_ticks_per_intr; - pc->pc_nexttickevent = pc->pc_nextstatevent = nextevent; - arm_intr_route(sc->sc_ih, 1, curcpu()); + clockintr_cpu_init(&agtimer_intrclock); + reg = agtimer_get_ctrl(); reg &= ~GTIMER_CNTV_CTL_IMASK; reg |= GTIMER_CNTV_CTL_ENABLE; - agtimer_set_tval(sc->sc_ticks_per_second); + agtimer_set_tval(INT32_MAX); agtimer_set_ctrl(reg); + + clockintr_trigger(); /* enable userland access to virtual counter */ kctl = READ_SPECIALREG(CNTKCTL_EL1); Index: sys/arch/arm64/include/cpu.h =================================================================== RCS file: /cvs/src/sys/arch/arm64/include/cpu.h,v retrieving revision 1.28 diff -u -p -r1.28 cpu.h --- sys/arch/arm64/include/cpu.h 29 Aug 2022 02:01:18 -0000 1.28 +++ sys/arch/arm64/include/cpu.h 6 Nov 2022 18:48:22 -0000 @@ -81,6 +81,7 @@ void arm32_vector_init(vaddr_t, int); * Per-CPU information. For now we assume one CPU. */ +#include <sys/clockintr.h> #include <sys/device.h> #include <sys/sched.h> #include <sys/srp.h> @@ -142,7 +143,7 @@ struct cpu_info { #ifdef GPROF struct gmonparam *ci_gmon; #endif - + struct clockintr_queue ci_queue; char ci_panicbuf[512]; }; Index: sys/arch/arm64/include/_types.h =================================================================== RCS file: /cvs/src/sys/arch/arm64/include/_types.h,v retrieving revision 1.4 diff -u -p -r1.4 _types.h --- sys/arch/arm64/include/_types.h 5 Mar 2018 01:15:25 -0000 1.4 +++ sys/arch/arm64/include/_types.h 6 Nov 2022 18:48:22 -0000 @@ -34,6 +34,8 @@ #ifndef _MACHINE__TYPES_H_ #define _MACHINE__TYPES_H_ +#define __HAVE_CLOCKINTR + #if defined(_KERNEL) typedef struct label_t { long val[13];