[PATCH 08/18] nvme-tcp: do not set MSG_SENDPAGE_NOTLAST

[Hannes Reinecke]

On 4/17/23 22:16, Jakub Kicinski wrote:
> On Mon, 17 Apr 2023 18:28:12 +0300 Sagi Grimberg wrote:
>>> But then we're running afoul with tls_sw_do_sendpage(), which uses
>>> MSG_SENDPAGE_NOTLAST as a record indicator. Forwarding
>>> MSG_SENDPAGE_NOTLAST to tls results in incorrect TLS records on the wire
>>> and a transmission stall.
>>>
>>> Took me days to figure that out.
>>>
>>> If you have a better idea, I'm all ears.
>>
>> So sendfile is unsupported over kTLS? Or is there some special
>> case for it in the upper layer?
>>
>> Jakub?
> 
> First I hear of it. 10 sec look at the code doesn't indicate that
> the author missed this case completely either.
> 
> Hannes, any more details on what's going wrong?
> 
> What incorrect TLS record have you seen? What's the sequence of sends?
> Offload or no offload? Easy to repro or stars need to (mis)align?
> 
Ah-ha. Found it.

Turns out to be an arguably invalid control flow on our side.

The initial 'connect' command is a command PDU with inline data.
So in our (current) flow we do a 'sendpage()' with the 
MORE/SENDPAGE_NOTLAST flag set as we have to transfer more data.
Ok so far.
But for transmitting the actual data we evaluate 'sendpage_ok()', which 
returns _false_. And then we continue to call 'sendmsg()' for the inline 
data.
This confuses the TLS software stack as sendpage() and sendmsg() are two 
distinct code paths, and we end up with a TX stall.

Solution: check 'sendpage_ok()' when we have inline data and use 
'sendmsg' for the command pdu if required.

It always pays to dig deeper if one's not exactly sure what's happening ...

Will be updating the patchset.

Cheers,

Hannes