Current implementation of __delay() function uses so-called zero-delay loops. And the only condition to exit that loop is LP_COUNT (loop count register) = 1 (but not 0 as it might be easily imagined). So if our calculation of "loops" gives 0 (and that is pretty possible given result of multiplication being >> 32) then zero-delay loop mechanism starts with LP_COUNT=0 and it ends up decrementing LP_COUNT while staying in the loop effectively producing close to infinite delay instead of very short one. I bumped into it with AXS101 + external DDR controller and caches disabled. In that case I've got very small loops_per_jiffy=0xf00: ------------------------>8-------------------- Calibrating delay loop... 0.77 BogoMIPS (lpj=3862) ------------------------>8-------------------- And on console output delays were way too long. Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com> Cc: Vineet Gupta <vgupta at synopsys.com> Please enter the commit message for your changes. Lines starting --- arch/arc/include/asm/delay.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/arch/arc/include/asm/delay.h b/arch/arc/include/asm/delay.h index 08e7e2a..1a7a1dc 100644 --- a/arch/arc/include/asm/delay.h +++ b/arch/arc/include/asm/delay.h @@ -57,7 +57,8 @@ static inline void __udelay(unsigned long usecs) */ loops = ((u64) usecs * 4295 * HZ * loops_per_jiffy) >> 32; - __delay(loops); + if (loops) + __delay(loops); } #define udelay(n) (__builtin_constant_p(n) ? ((n) > 20000 ? __bad_udelay() \ -- 2.4.3
