Re: [PATCH for-next 2/2] RDMA/efa: Pass relaxed ordering flag to device

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Aug 06, 2026 at 07:57:58AM +0000, Yonatan Nachum wrote:

> @@ -1647,6 +1650,9 @@ static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags,
>  		(EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0) |
>  		(EFA_DEV_CAP(dev, RDMA_WRITE) ? IB_ACCESS_REMOTE_WRITE : 0);
>  
> +	if (EFA_DEV_CAP(dev, MR_RELAXED_ORDERING))
> +		supp_access_flags |= IB_ACCESS_RELAXED_ORDERING;
> +
>  	access_flags &= ~IB_ACCESS_OPTIONAL;

Since:

	IB_UVERBS_ACCESS_RELAXED_ORDERING = IB_UVERBS_ACCESS_OPTIONAL_FIRST,

This is confusing / dead code

The rest of the series looked Ok, but why are you reporting a new CAP
back to userspace? Did you have a use in mind?

As above existing kernels will accept IB_ACCESS_RELAXED_ORDERING just
fine, and it is really just a performance hint anyhow so it is fine if
old kernels ignore it.

Jason