Quoting Russell King's explaination of original bug I filed. ARM folk have recently pointed out a problem with sys_times(). When the kernel (2.6) boots, we set jiffies to -5 minutes. This causes sys_times() to return a negative number, which increments through zero. However, some negative numbers are used to return error codes. Hence, there's a period of time when sys_times() returns values which are indistinguishable from error codes shortly after boot. This probably only affects 32-bit architectures. However, one wonders whether sys_times() needs force_successful_syscall_return(). Also, it appears that glibc does indeed interpret the return value from sys_times in the way I describe above on at least ARM and x86. Other architectures may be similarly affected.
This is a kernel problem. The kernel should not return negative values unless it's due to true overflow. I'd say the kernel ought to subtract INITIAL_JIFFIES when converting to clock_t for sys_times. But at any rate, it is not something glibc can do anything about.