cvs commit: src/sys/sys time.h src/sys/kern kern_time.c

[Bruce Evans]

On Sun, 27 Nov 2005, Robert Watson wrote:

> ...  Attached is a small LD_PRELOAD library to 
> allow applications to be switched to the _FAST variants for gettimeofday().

% /*
%  * Small wrapper library to substitute implementations of gettimeofday(2) and
%  * time(3) with lower resolution variations.  time(3) is unconditionally
%  * degraded, since it will return a truncated time anyway.  gettimeofday(3)
%  * checks the TIMEWRAPPER environmental variable, which can be set to either
%  * "PRECISE" or "FAST".
%  */

time(3) should use the environment variable too, since the fast version
gives a value that is both imprecise and wrong.  It inherits bugs from
the kernel's time_second variable.  time_second is not the current
time truncated, but is the (current time less up to about tc_tick/HZ)
truncated.  It lags the current time by more than 1 second for up to
about tc_tick/HZ seconds before every rollover of the correct truncated
time.

When HZ was 100 it was very easy for applications to see time_second
being wring since the average size of the window of incoherency was
1/200 second.  time_second is visible in file times by default for ffs
file times.  The bug can be fixed for ffs too globally using the
vfs.timestamp_precision sysctl.  I only use this sysctl for running
POSIX tests for this bug, but would be surprised if always using it
caused noticeable pessimization.  Updates of file timestamps are cached
in non-broken file systems, so it is hard for them to occur as often
as ones for context switches -- they will usually be limited by physical
disk write bandwidth.

Bruce