Hi Wang,
On Sun, Feb 24, 2019 at 05:36:35PM +0800, kernel@xxxxxxxxxx wrote:
> From: Wang Xuerui <wangxuerui@xxxxxxxxx>
>
> The clocksource bits are ready, just wire things up in VDSO for a
> significant user-space timekeeping performance gain. There are several
> ABI problems uncovered by vdsotest though, but daily usage of the test
> system didn't expose any of them. These inconsistencies will be fixed
> in a later commit (presently TODO).
I'm afraid that's not how this works... :)
We should make sure this works properly before merging it - could you
provide details of the problems vdsotest found?
> According to vdsotest (formatting is manually added, only the affected
> figures are shown):
>
> category before after
> -------- ------ -----
> clock-gettime-monotonic: syscall: 409 401 nsec/call
> clock-gettime-monotonic: libc: 476 141 nsec/call
> clock-gettime-monotonic: vdso: 462 123 nsec/call
>
> clock-gettime-realtime: syscall: 405 407 nsec/call
> clock-gettime-realtime: libc: 474 142 nsec/call
> clock-gettime-realtime: vdso: 457 125 nsec/call
>
> gettimeofday: syscall: 406 407 nsec/call
> gettimeofday: libc: 455 121 nsec/call
> gettimeofday: vdso: 440 102 nsec/call
I'd rather be sure the kernel gives the right answer than have it give
the wrong answer 4x faster, so this looks promising but let's fix the
test failures before moving forwards.
Thanks,
Paul
