Martijn van Oosterhout <kleptog@svana.org> writes:
> The interesting thing about this is that they obviously are gearing
> gettimeofday() to be accurate, rather than just considering it a
> counter that is somewhat close to real time. At the expense of speed.
Not sure that that's an accurate description. What I think the kernel
fuss is about is that they have to read the counter value as several
separate "byte read" operations, and if the hardware doesn't latch the
whole counter value when the first byte is pulled then they'll get bytes
from several distinct states of the counter, leading to something that's
not consistent or even monotonic. On non-latching hardware there's
really not a lot of choice what to do. The patch is about not using
that same very-slow code path on hardware that does latch.
regards, tom lane