Sven Schnelle <sv...@linux.ibm.com> writes:
> - CPUCLOCK_VIRT is now handled with a syscall fallback, which might
> be slower/less accurate than the old implementation.
I can understand the slower, but why does it become less accurate?
> Performance number from my system do 100 mio gettimeofday() calls:
>
> Plain syscall: 8.6s
> Generic VDSO: 1.3s
> old ASM VDSO: 1s
>
> So it's a bit slower but still much faster than syscalls.
Where is the overhead coming from?
> +static inline u64 __arch_get_hw_counter(s32 clock_mode)
> +{
> + const struct vdso_data *vdso = __arch_get_vdso_data();
> + u64 adj, now;
> + int cnt;
> +
> + do {
> + do {
> + cnt = READ_ONCE(vdso->arch.tb_update_cnt);
> + } while (cnt & 1);
smp_rmb() ?
> + now = get_tod_clock();
> + adj = vdso->arch.tod_steering_end - now;
> + if (unlikely((s64) adj > 0))
> + now += (vdso->arch.tod_steering_delta < 0) ? (adj >>
> 15) : -(adj >> 15);
smp_rmb() ?
> + } while (cnt != READ_ONCE(vdso->arch.tb_update_cnt));
> + return now;
> if (ptff_query(PTFF_QTO) && ptff(&qto, sizeof(qto), PTFF_QTO) == 0)
> lpar_offset = qto.tod_epoch_difference;
> @@ -599,6 +550,13 @@ static int stp_sync_clock(void *data)
> if (stp_info.todoff[0] || stp_info.todoff[1] ||
> stp_info.todoff[2] || stp_info.todoff[3] ||
> stp_info.tmd != 2) {
> + vdso_data->arch.tb_update_cnt++;
> + /*
> + * This barrier isn't really needed as we're called
> + * from stop_machine_cpuslocked(). However it doesn't
> + * hurt in case the code gets changed.
> + */
> + smp_wmb();
WMB without a corresponding RMB and an explanation what's ordered
against what is voodoo at best.
> rc = chsc_sstpc(stp_page, STP_OP_SYNC, 0,
> &clock_delta);
> if (rc == 0) {
> @@ -609,6 +567,8 @@ static int stp_sync_clock(void *data)
> if (rc == 0 && stp_info.tmd != 2)
> rc = -EAGAIN;
> }
> + smp_wmb(); /* see comment above */
See my comments above :)
Thanks,
tglx
