Re: [PATCH 1/2] blk-throttle: avoid ilog2(0) in calculate_bytes_allowed()

On Tue, 14 Jul 2026 18:30:27 +0800
Tao Cui <[email protected]> wrote:

> From: Tao Cui <[email protected]>
> 
> __tg_update_carryover() can call calculate_bytes_allowed() with a zero
> jiffy_elapsed right after a slice starts. The overflow guard
> 
>     if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
> 
> relies on ilog2(0) == -1 (fls64(0) - 1) to stay below the threshold so
> that the subsequent mul_u64_u64_div_u64(bps, 0, HZ) == 0 is reached.
> That works, but the ilog2(0) dependency is non-obvious.
> 
> Add an explicit early return for jiffy_elapsed == 0, which is equivalent
> (mul_u64_u64_div_u64(bps_limit, 0, HZ) == 0) and removes the reliance on
> ilog2(0).
> 
> No behavior change.

There is a pending patch to make the x86-64 mul_u64_u64_div_u64()
return ~0ULL on overflow to match the generic code.
That would completely remove the requirement for the overflow guard.

I can't remember why it got lost (again).
There have been arguments about what should happen for divide by zero.
Personally I'd return ~0ULL and let the caller decide what that means,
but the generic mul_u64_u64_div_u64() goes to lengths to generate any
trap that a normal divide by zero would generate.

	David

> 
> Signed-off-by: Tao Cui <[email protected]>
> ---
>  block/blk-throttle.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/block/blk-throttle.c b/block/blk-throttle.c
> index ffc3b70065d4..f37911abefdd 100644
> --- a/block/blk-throttle.c
> +++ b/block/blk-throttle.c
> @@ -603,6 +603,10 @@ static unsigned int calculate_io_allowed(u32 iops_limit,
>  
>  static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
>  {
> +	/* 0 elapsed => 0 bytes allowed; also avoids ilog2(0) below. */
> +	if (!jiffy_elapsed)
> +		return 0;
> +
>  	/*
>  	 * Can result be wider than 64 bits?
>  	 * We check against 62, not 64, due to ilog2 truncation.