Re: [PATCH bpf-next 2/5] bpf: bpf_csum_diff: optimize and homogenize for all archs

[Daniel Borkmann]

On 10/21/24 2:21 PM, Puranjay Mohan wrote:
> 1. Optimization
>     ------------
>
> The current implementation copies the 'from' and 'to' buffers to a
> scratchpad and it takes the bitwise NOT of 'from' buffer while copying.
> In the next step csum_partial() is called with this scratchpad.
>
> so, mathematically, the current implementation is doing:
>
> 	result = csum(to - from)
>
> Here, 'to'  and '~ from' are copied in to the scratchpad buffer, we need
> it in the scratchpad buffer because csum_partial() takes a single
> contiguous buffer and not two disjoint buffers like 'to' and 'from'.
>
> We can re write this equation to:
>
> 	result = csum(to) - csum(from)
>
> using the distributive property of csum().
>
> this allows 'to' and 'from' to be at different locations and therefore
> this scratchpad and copying is not needed.
>
> This in C code will look like:
>
> result = csum_sub(csum_partial(to, to_size, seed),
>                    csum_partial(from, from_size, 0));
>
> 2. Homogenization
>     --------------
>
> The bpf_csum_diff() helper calls csum_partial() which is implemented by
> some architectures like arm and x86 but other architectures rely on the
> generic implementation in lib/checksum.c
>
> The generic implementation in lib/checksum.c returns a 16 bit value but
> the arch specific implementations can return more than 16 bits, this
> works out in most places because before the result is used, it is passed
> through csum_fold() that turns it into a 16-bit value.
>
> bpf_csum_diff() directly returns the value from csum_partial() and
> therefore the returned values could be different on different
> architectures. see discussion in [1]:
>
> for the int value 28 the calculated checksums are:
>
> x86                    :    -29 : 0xffffffe3
> generic (arm64, riscv) :  65507 : 0x0000ffe3
> arm                    : 131042 : 0x0001ffe2
>
> Pass the result of bpf_csum_diff() through from32to16() before returning
> to homogenize this result for all architectures.
>
> NOTE: from32to16() is used instead of csum_fold() because csum_fold()
> does from32to16() + bitwise NOT of the result, which is not what we want
> to do here.
>
> [1] https://lore.kernel.org/bpf/CAJ+HfNiQbOcqCLxFUP2FMm5QrLXUUaj852Fxe3hn_2JNiucn6g@xxxxxxxxxxxxxx/
>
> Signed-off-by: Puranjay Mohan <puranjay@xxxxxxxxxx>

Thanks for looking into this!

Acked-by: Daniel Borkmann <daniel@xxxxxxxxxxxxx>