Re: [PATCH net-next] tcp: devmem: only pre-allocate tokens the receiver can consume

[Bobby Eshleman]

On Fri, Sep 04, 2026 at 03:11:46PM -0700, Jakub Kicinski wrote:
> On Wed, 02 Sep 2026 15:24:19 -0700 Bobby Eshleman wrote:
> > Testing on a CX7 w/ GRO and a steady sendmsg() flow of 1MB per send, we
> > see a typical RX-side skb touch upwards of ~32KB. With a 4KB recvmsg
> > size, probing shows that ~75% of the allocated tokens are not used.
> 
> 4kB seems unreasonably small. Is there a real life reason to use such
> a small buffer / read length?

I have workloads now that deal w/ varying sized tensors (<32K up to
100MB), and data send/recv is done per-tensor (with the size/metadata
sent over a control path via normal TCP). Right now the application just
safe guards against the small tensor case by receiving anything under
32K via regular TCP, and anything over via TCP dm. Figured might as well
try to lift the restriction. To your point though, I don't expect a
consistent flow of <32K to be the common case. Could be something we
shelve until it really becomes painful.

Best,
Bobby