Re: [PATCH RFC 3/3] RDMA/rtrs: Post client read requests inline

[Andrei-Alexandru Bleortu]

On Tue, Oct 20, 2026, Jinpu Wang wrote:
> The change looks fine to me, and I did some testing on it. However,
> the data doesn't seem to show any clear performance gain from posting
> inline requests.

Thanks for testing it.

That matches what I measured. The gain only showed with one I/O in
flight against a memory-backed device (4 KB random reads to null_blk,
ConnectX-6 Dx over RoCE): fio's mean completion latency went from 20.1
to 19.7 us, under 2 %, and it took 15 paired rounds to resolve. With
more I/O in flight or a real drive it would be lost in the noise.

As I wrote to Leon, I'll leave the series unless someone finds that
latency matters to them.