Disable lwIP checksum-on-copy

[kjbracey]

Current version of lwIP has a bug in its checksum-on-copy code - see

#4140 and https://savannah.nongnu.org/bugs/?50914

Pending a fix from lwIP, set LWIP_CHECKSUM_ON_COPY to 0 to work around.
Will impact performance.

@mikaleppanen , @geky - if there's not a patch for the bug in the next few days, I think we should do this.

[geky]

Makes sense to me 👍

This seems to just be an optimization, so we can drop it until a fix is available.

[0xc0170]

/morph test-nightly

[mbed-bot]

Result: SUCCESS

Your command has finished executing! Here's what you wrote!

/morph test-nightly

Output

mbed Build Number: 222

All builds and test passed!

[sg-]

@kjbracey-arm given we merge this, will you be tracking an upstream fix that we will later pickup?

[kjbracey]

There's a link to the lwIP bug tracker left in the code - we're not planning to investigate ourselves, and I'm not sure it's a significant enough issue to warrant cherry-picking a fix but we'll try to remember to recheck whether it's fixed next time we go for a general update.

I think it's arguable we should have the option disabled anyway, at least by default. There's a moderate amount of additional code and small RAM cost involved in doing the checksum-on-copy, and more mbed platforms are worried about space than CPU time spent on IP.

It's not that much extra memory, but every little helps. Does that makes sense?

[kjbracey]

This one's probably backportable.

Some people will never see it, but some TCP connections will just die - the stack will persistently resend a packet with an incorrect checksum, as in the original report.

Looks like it may be the usual zero checksum confusion, which suggests 1 in 65536 packets may suffer from it. If you're lucky, there will be more data and hence different checksum in the retransmit. If not, connection over.