Re: af_alg: Fail in continuable manner on Linux/powerpc64le

[Pádraig Brady]

On 24/06/18 19:32, Bruno Haible wrote:
> This fixes the error message reported by Assaf Gordon on non-seekable
> input.
> 
> Four approaches come to mind for fixing this:
>  1) Change the afalg_stream signature, adding an argument where the function
>     can store read but not yet processed input.
>  2) Change the afalg_stream signature, passing it the fallback computation
>     function as argument. So that afalg_stream can invoke this function on
>     read input that it could not process itself.
>  3) Try the first send() with 1 byte only, because upon failure it can be
>     pushed back onto the stream.
>  4) Try the first send() with the number of bytes returned by freadptr,
>     instead of fread().
> 
> The first two approaches would be a major code change, for just a little
> particular case. I prefer the third approach.

For reference a quick search suggests this bug may have been:
https://patchwork.kernel.org/patch/8124631/
and was noticed by Assaf on a 3.10 kernel.

3) Seems like the best approach.
I'd have a slight reservation on perf impact of I/O
aligned 1 byte over natural boundaries.

Then implementation looks sound.

thanks,
Pádraig