Re: [Qemu-devel] [PATCH v4 14/30] stellaris_enet: avoid buffer overrun on incoming migration (part 2)

[Peter Maydell]

On 1 April 2014 11:05, Peter Maydell <address@hidden> wrote:
> But note that there seems to be a bug or two in
> the DATA read logic: our cutoff for tx frame too
> long is tx_frame_len > 2032, but for the limit
> case of 2032, if we add 14 for the ethernet header
> and 4 for explicit CRC then we get 2050, which is
> slightly larger than the tx_fifo buffer... I think
> we either need to wind that 2032 value down or
> enlarge the tx_fifo buffer, or possibly just tweak
> the logic not to try to actually write the CRC bytes
> into the FIFO since we promptly ignore them -- need
> to check the h/w datasheet to find out which.

The datasheet isn't entirely clear on this point.
It is fairly definite that the TX overrun check is
done for tx_frame_len written as >2032, but it
doesn't really say what happens if there's not
enough space for the CRC. I rather suspect everybody
is using this device with auto-CRC enabled anyway,
and I have neither the hardware nor the time to
experiment with it. I'll put together a patch that
does something plausible and doesn't allow data
buffer overruns; roughly, we separate the check for
"can we put this word into the FIFO?" from the
one for "should we transmit the packet now?". This
is actually closer to the h/w behaviour, because the
h/w doesn't auto-transmit the packet on a full FIFO
in all cases -- unless you set the transmit-start
threshold you have to explicitly write to a different
register to say "now start transmitting".

thanks
-- PMM