On 2/25/19 7:10 PM, Stanislav Fomichev wrote:
> On 02/25, Alexei Starovoitov wrote:
>> On 2/25/19 3:07 PM, Stanislav Fomichev wrote:
>>>> +#define BPF_PROG_RUN(prog, ctx) ({ \
>>>> + u32 ret; \
>>>> + cant_sleep(); \
>>>> + if (static_branch_unlikely(&bpf_stats_enabled_key)) { \
>>>> + struct bpf_prog_stats *stats; \
>>>> + u64 start = sched_clock(); \
>>> QQ: why sched_clock() and not, for example, ktime_get_ns() which we do
>>> in the bpf_test_run()? Or even why not local_clock?
>>> I'm just wondering what king of trade off we are doing here
>>> regarding precision vs run time cost.
>>
>>
>> I'm making this decision based on documentation:
>> Documentation/timers/timekeeping.txt
>> "Compared to clock sources, sched_clock() has to be very fast: it is
>> called much more often, especially by the scheduler. If you have to do
>> trade-offs between accuracy compared to the clock source, you may
>> sacrifice accuracy for speed in sched_clock()."
> So sched_clock is fast, but imprecise; and ktime_get_ns (and
> lock_clock?) are slow(er), but more precise?
>
> If that's the case, would it make sense to use a more precise
> measurement? I suppose the BPF program execution time is on the order of
> nanoseconds and if sched_close has msec or usec resolution, all we get is
> essentially noise?
>
> I understand that you want this feature to have almost no overhead, but
> since it's gated by the static key, should we aim for a higher precision?
>
Considering everything I believe sched_clock() strikes the best trade off.
