hwclock causes __might_sleep dump

Hi

at system shutdown when hwclock --systohc is called there's a might_sleep error dump from the kernel in do_page_fault. I found that the problem is when hwclock uses direct cmos access when there's no RTC support in kernel and only when hwclock is compiled with -O3 hwclock does an atomic access achieved with __asm__ cli and sti. Is this a hwclock bug? I suppose the kernel is not responsible for userland disabling interrupts ... putting delays and printfs in hwclock either makes the dump show a different backtrace (although still with 2 do_page_faults) or results in no dump at all.

2.6.0-test1 and test2