lkml.org 
[lkml]   [2014]   [Sep]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] tile: add clock_gettime support to vDSO
On Tue, 30 Sep 2014, Chris Metcalf wrote:

> This change adds support for clock_gettime with CLOCK_REALTIME
> and CLOCK_MONOTONIC using vDSO. In addition, with this change
> we switch to use seqlocks instead of integer counters.

I'd rather split that into two patches. One changing the code to use
the seqlock and the other to add the clock_gettime() stuff.

> void update_vsyscall(struct timekeeper *tk)
> @@ -263,20 +260,30 @@ void update_vsyscall(struct timekeeper *tk)
> struct timespec wall_time = tk_xtime(tk);
> struct timespec *wtm = &tk->wall_to_monotonic;
> struct clocksource *clock = tk->clock;
> + struct timespec ts;
>
> if (clock != &cycle_counter_cs)
> return;
>
> - /* Userspace gettimeofday will spin while this value is odd. */
> - ++vdso_data->tb_update_count;
> - smp_wmb();
> + write_seqcount_begin(&vdso_data->tb_seq);
> +
> vdso_data->xtime_tod_stamp = clock->cycle_last;
> vdso_data->xtime_clock_sec = wall_time.tv_sec;
> vdso_data->xtime_clock_nsec = wall_time.tv_nsec;
> - vdso_data->wtom_clock_sec = wtm->tv_sec;
> - vdso_data->wtom_clock_nsec = wtm->tv_nsec;
> +
> + ts = timespec_add(wall_time, *wtm);
> + vdso_data->wtom_clock_sec = ts.tv_sec;
> + vdso_data->wtom_clock_nsec = ts.tv_nsec;
> +
> vdso_data->mult = clock->mult;
> vdso_data->shift = clock->shift;
> - smp_wmb();
> - ++vdso_data->tb_update_count;
> +
> + ts = __current_kernel_time();
> + vdso_data->xtime_clock_coarse_sec = ts.tv_sec;
> + vdso_data->xtime_clock_coarse_nsec = ts.tv_nsec;
> + ts = timespec_add(ts, *wtm);
> + vdso_data->wtom_clock_coarse_sec = ts.tv_sec;
> + vdso_data->wtom_clock_coarse_nsec = ts.tv_nsec;

I'm fine with the code, but you might think about doing the math and
preparation stuff outside of the seqcount protected region to a shadow
struct and then do a simple memcpy inside of the seqcount protected
region. Nothing significant, but it might be worthwhile as a follow
up or preparatory change.

> +static inline int do_monotonic(struct vdso_data *vdso, struct timespec *ts)
> {
> + int count;
> cycles_t cycles;
> - unsigned long count, sec, ns;
> - volatile struct vdso_data *vdso_data;
> + unsigned long ns;
> +
> + do {
> + count = read_seqcount_begin(&vdso->tb_seq);
> + cycles = get_cycles() - vdso->xtime_tod_stamp;
> + ns = (cycles * vdso->mult) >> vdso->shift;
> + ts->tv_sec = vdso->wtom_clock_sec;
> + ts->tv_nsec = vdso->wtom_clock_nsec;
> + } while (unlikely(read_seqcount_retry(&vdso->tb_seq, count)));

I doubt that the unlikely makes any difference.

Otherwise this looks good.

Thanks,

tglx


\
 
 \ /
  Last update: 2014-10-01 01:01    [W:0.482 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site