Re: [PATCH for-7.2] rtl8139: honor large send MSS value

[Jason Wang <jasowang@redhat.com>]

On Wed, Nov 16, 2022 at 10:59 AM Tobias Fiebig <tobias@fiebig.nl> wrote:
>
> Heho,
> I just tested around with the patch;
> Good news: Certainly my builds are being executed. Also, if I patch the old code to have a MAX_MTU <= the max MTU on my path, throughput is ok.
>
> Bad news: Something is wrong with getting the MSS in the patch you shared. When enabling DPRINT, values are off (sent MSS vs. printed MSS):
> 600 2060
> 800 2308
> 1000 2316
> 1023 2307
> 1200 3076
> 1400 3340 (most likely clamped to 1320)
>
> Fiddling around a bit more, I found txdw0 printed earlier in the stack as hex (sent MSS, txdw0):
> 769 900502f5
> 1000 900503dc
> 1280 900504f4
> 1281 900504f5
> 1301 90050509
> 1317 90050519
> 1320 9005051c
>
> This maps rather well to:
> MSS = txdw0 - 2416246772
> MSS = txdw0 - 9004FFF4
>
> Sadly, my C is 'non-existent' and it is kind-of 4AM, so also not in the brainspace to fill those gaps. But if one of you could look at the patch again, that would be nice. Otherwise, I should have some brainspace for this tomorrow night (UTC) again.
>

Ok, I think I found at least one issue:

/* large send MSS mask, bits 16...25 */
#define CP_TC_LGSEN_MSS_MASK ((1 << 12) - 1)

First, MSS occupies 11 bits from 16 to 26
Second, the mask is wrong it should be ((1 << 11) - 1)

Thanks

> With best regards,
> Tobias
>