On Sun, 24 Jan 2021 22:38:02 -0600 Lijun Pan wrote:
> On Sat, Jan 23, 2021 at 11:11 PM Jakub Kicinski <k...@kernel.org> wrote:
> > On Thu, 21 Jan 2021 00:17:09 -0600 Lijun Pan wrote:
> > > rmb() was introduced to load rx_scrq->msgs after calling
> > > pending_scrq(). Now since pending_scrq() itself already
> > > has dma_rmb() at the end of the function, rmb() is
> > > duplicated and can be removed.
> > >
> > > Fixes: ec20f36bb41a ("ibmvnic: Correctly re-enable interrupts in NAPI
> > > polling routine")
> > > Signed-off-by: Lijun Pan <l...@linux.ibm.com>
> >
> > rmb() is a stronger barrier than dma_rmb()
>
> Yes. I think the weaker dma_rmb() here is enough.
> And I let it reuse the dma_rmb() in the pending_scrq().
>
> >
> > also again, I don't see how this fixes any bugs
>
> I will send to net-next if you are ok with it.
If there is consensus at IBM that the first 2 changes are an
improvement you can drop the Fixes tags and resubmit to net-next.
In patch 3 it looks like the dma_rmb() may indeed be missing so that
one could go to net, but I don't think the dma_wmb() is needed,
especially not where you put it. I think dma_wmb() is only needed
before the device is notified that new buffer was posted, not on
completion.
