Re: [PATCH rdma-next v2 0/5] RDMA/irdma: Adopt robust udata

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Jun 30, 2026 at 01:23:04PM -0400, Jacob Moroni wrote:
> Hi Jason,
> 
> I need to respin this yet again to fix a few more sashiko findings (sorry), but
> I was curious:
> 
> Looking at bnxt_re and irdma, in many cases where we use
> ib_respond_empty_udata, there is also a check for
> ib_is_udata_in_empty.
> 
> Should I make a combined helper? Something like "ib_check_no_udata_io"?

I imagined that the functions should all have a common pattern,
validate the input at the top and generate the response on the success
path.

> Also, for bnxt_re, I think there are places where the ib_respond_empty_udata
> call at the end of the method could be problematic, particularly for "destroy"
> ops since the driver doesn't check the return value. V1 of this series did the
> same for irdma and sashiko freaked out.

It should generally be called at the end of the method, and everything
should be careful to undo their stuff if it fails. Much of that is
bundled into the uobject abort mechanism so it happens transparently
in many cases.

> The issue is that ib_respond_empty_udata can return a failure but the driver
> doesn't check for it.

It does check, I guess this is problematic that the object is already
destroyed and cannot be undestroyed at this point, yet could be
destroyed again.

> Does this seem legitimate? Anyway, a unified helper that is called right
> at the beginning would eliminate the issue if so.

Yes, that does seem like a good reason to make that kind of change.

Jason