I've run the little test script and it seemed to be fixed after I had
patched my server to set VA_UTIMES_NULL when setting to server time(s).

So, I think that your diagnosis was correct, but your patch didn't quite
fix it. Although I haven't tested it, here is the equivalent patch to
what I did, for OpenBSD3.8. Maybe you can try this?

Good luck with it, rick
ps: I tested using a Linux2.6 Fedora core 4 client.
*** nfsm_subs.h.orig Tue Mar 28 11:16:50 2006
--- nfsm_subs.h Tue Mar 28 11:18:31 2006
***************
*** 472,477 ****
--- 472,478 ----
break; \
case NFSV3SATTRTIME_TOSERVER: \
getnanotime(&(a)->va_atime); \
+ (a)->va_vaflags |= VA_UTIMES_NULL; \
break; \
}; \
nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED); \
***************
*** 482,487 ****
--- 483,489 ----
break; \
case NFSV3SATTRTIME_TOSERVER: \
getnanotime(&(a)->va_mtime); \
+ (a)->va_vaflags |= VA_UTIMES_NULL; \
break; \
}; }