Re: [lwip-users] TCP fragmentation over PPP - fragments are lost in LwIP if data is queued for sending

[Simon Goldschmidt]

Marco Jakobs wrote:
> We're not using nonblocking connections as this never did work properly … in 
> my today's
> debug session I've found why. In api_lib.c, line 610 this code just prevents 
> using
> NETCONN_WRITE for nonblocking connections:
> 
>  dontblock = netconn_is_nonblocking(conn) || (apiflags & NETCONN_DONTBLOCK);
>  if (dontblock && !bytes_written) {
>    /* This implies netconn_write() cannot be used for non-blocking send, since
>       it has no way to return the number of bytes written. */
>    return ERR_VAL;
>  }

This code is correct. You're just using the wrong function. For nonblocking, 
you need to
use 'netconn_write_partly()' instead of 'netconn_write()'. This ensures you get 
the number
of actually written bytes - as you saw, ERR_OK can be returned without all 
bytes written.

Returning ERR_WOULDBLOCK is not a solution as some bytes are already written 
(in that case at least).

The above code is simply missing a check for send-timeout. Using send-timeout 
without passing
a 'bytes_written' pointer (or using 'netconn_write()' instead of 
'netconn_write_partly()'),
does not work (for the reason described above). So the bug you found is that 
this check is missing.

> [..]
> Giving back an ERR_OK where data is definitely lost is clearly a bug,

Yes. I'll add the check that nonblocking write needs a 'bytes_left' pointer, 
too. Thanks for
tracking that down.

> so as a solution we've changed the SNDTIMEO handling in do_writemore in this 
> way:

Congrats, with this change, you effectively turned send-timeout into 
nonblocking ;-)


Simon