[PATCH] nvmet-tcp: bound sgl->length check in nvmet_tcp_map_data()

[Maurizio Lombardi]

On Tue Apr 7, 2026 at 8:30 AM CEST, Christoph Hellwig wrote:
> On Sun, Apr 05, 2026 at 03:46:12PM -0400, Shivam Kumar wrote:
>> Hi Christoph, Maurizio
>> 
>> Just following up on this thread.
>> Would you like me to send a v2 patch that sets a sane default MDTS for
>> the TCP transport?
>
> Yes, that would be great!
>
>> If so, is there a preferred value I should use, or
>> should I follow what the other transports do?
>
> I'm honestly not sure.  Maurizio/Sagi: any good idea what a TCP default
> max MDTS might be?

I'm not sure there is a "correct" value here. It seems to mostly come
down to balancing performance vs memory usage, rather than anything
TCP-specific.

For reference, pci-epf allows a max MDTS of 2 MiB[1], while RDMA uses 1 MiB.

Personally, I would pick a default of 2 MiB, it should be large
enough to avoid excessive splitting while still being reasonable from a
memory perspective.

Sagi? Any thought on that?

I have seen a patch on this mailing list that makes mdts configurable
via configfs, so in case of need a user could still change it
to his preferred value.

Maurizio

[1] https://docs.kernel.org/nvme/nvme-pci-endpoint-target.html