[PATCH] nvme-tcp: add basic support for the C2HTermReq PDU

[Sagi Grimberg]

On 24/02/2025 19:03, Hannes Reinecke wrote:
> On 2/24/25 17:48, Maurizio Lombardi wrote:
>> On Mon Feb 24, 2025 at 5:04 PM CET, Hannes Reinecke wrote:
>>> I really wouldn't be doing that, as we need to terminate the queue upon
>>> fatal error and don't really want to delay that (eg by waiting for the
>>> transfer to finish) and
>>
>> I did some tests and it seems to work with MSG_DONTWAIT as well.
>>
>> I am just wondering if calling kernel_sock_shutdown() could
>> in some cases race with MSG_DONTWAIT, preventing the packet from
>> being effectively sent.
>> Not a big problem anyway because, as you said, it's a kind
>> of best-effort thing.
>>
> 'xactly.
> If in doubt we should prefer latency over correctness; and especially
> when running over a more complicated network setup (or doing TLS) it's 
> anyone's guess if the entire stack is still operational by the time
> we're trying to send the c2hterm.
> Part of the reason why it wasn't implement initially I guess :-)

Agree, this is a best effort thing. No need to jump through hoops to get 
it...