Re: [PATCH v6 2/2] lib: checksum: Use aligned accesses for ip_fast_csum and csum_ipv6_magic tests

[Charlie Jenkins <charlie@rivosinc.com>]

On Mon, Feb 12, 2024 at 09:18:14AM -0800, Guenter Roeck wrote:
> On Mon, Feb 12, 2024 at 06:26:14AM +0000, Al Viro wrote:
> > On Wed, Feb 07, 2024 at 04:22:51PM -0800, Charlie Jenkins wrote:
> > > +	struct csum_ipv6_magic_data {
> > > +		const struct in6_addr saddr;
> > > +		const struct in6_addr daddr;
> > > +		unsigned int len;
> > > +		__wsum csum;
> > > +		unsigned char proto;
> > > +	} data, *data_ptr;
> > 
> > Huh?
> > 
> > > +	int num_tests = MAX_LEN / WORD_ALIGNMENT - sizeof(struct csum_ipv6_magic_data);
> > > +
> > > +	for (int i = 0; i < num_tests; i++) {
> > > +		data_ptr = (struct csum_ipv6_magic_data *)(random_buf + (i * WORD_ALIGNMENT));
> > > +
> > > +		cpu_to_be32_array((__be32 *)&data.saddr, (const u32 *)&data_ptr->saddr,
> > > +				  sizeof_field(struct csum_ipv6_magic_data, saddr) / 4);
> > > +		cpu_to_be32_array((__be32 *)&data.daddr, (const u32 *)&data_ptr->daddr,
> > > +				  sizeof_field(struct csum_ipv6_magic_data, daddr) / 4);
> > > +		data.len = data_ptr->len;
> > > +		data.csum = (__force __wsum)htonl((__force u32)data_ptr->csum);
> > 
> > What are those cpu_to_be32() about?  Checksum calculations *DO* *NOT* involve
> > any endianness conversions.  At any point.
> > 
> > Replace those assignments with memcpy() and be done with that - that will take
> > care of unaligned accesses.
> > 
> > Result will have host-independent memory representation.  The only place where you
> > might want to play with byteswaps (only 16-bit ones) is if you initialized the
> > array of expected results with u16 constants.  That will have opposite memory
> > representations on l-e and b-e, so you'll need to byteswap to compare with
> > what you get from function.  Alternatively, make it an array of bytes and
> > do
> > 	sum16 = csum_ipv6_magic(saddr, daddr, len, proto, csum);
> > 	if (memcmp(sum16, expected_csum_ipv6_magic + i * 2, 2))
> > 		complain
> > 
> > That's it.
> 
> Almost. Turns out the csum parameter of csum_ipv6_magic() needs to be in
> network byte order, and the length parameter needs to be in host byte order.
> So instead of
> 	data.len = data_ptr->len;
> 	data.csum = (__force __wsum)htonl((__force u32)data_ptr->csum);
> it needs to be something like
> 	data.len = ntohl(data_ptr->len);
> 	data.csum = data_ptr->csum;
> 
> Also, as you mentioned, either the returned checksum or the expected
> checksum needs to be converted for the comparison because one is in
> network byte order and the other in host byte order.
> 
> Address conversions are indeed not needed.
> 
> Thanks,
> Guenter

Aww that makes sense thank you. I was reversing everything except len
and the solution was to only reverse len... Thank you for figuring that
out for me.

I will send out another version with your change. Should I add a
signed-off-by with your tag for this patch?

- Charlie