From: Haavard Skinnemoen During testing it was discovered that mdelay() didn't provide as long delay as it should. The reason is that __const_udelay() should have stored the result of (loops_per_jiffy * HZ * xloops) in a 64-bit register pair but didn't. Fix the problem by doing a 32 x 32 => 64 bit multiplication in inline assembly. This could probably have been solved by some casting, but IMO the inline asm makes the intention more clear. As an added bonus, the new code looks more like the i386 code. Signed-off-by: Haavard Skinnemoen Signed-off-by: Andrew Morton --- arch/avr32/lib/delay.c | 4 +++- 1 files changed, 3 insertions(+), 1 deletion(-) diff -puN arch/avr32/lib/delay.c~avr32-fix-__const_udelay-overflow-bug arch/avr32/lib/delay.c --- a/arch/avr32/lib/delay.c~avr32-fix-__const_udelay-overflow-bug +++ a/arch/avr32/lib/delay.c @@ -38,7 +38,9 @@ inline void __const_udelay(unsigned long { unsigned long long loops; - loops = (current_cpu_data.loops_per_jiffy * HZ) * xloops; + asm("mulu.d %0, %1, %2" + : "=r"(loops) + : "r"(current_cpu_data.loops_per_jiffy * HZ), "r"(xloops)); __delay(loops >> 32); } _