RE: [PATCH v5 3/5] bnx2x: Eliminate duplicate barriers on weakly-ordered archs

From: Chopra, Manish
Date: Sat Mar 24 2018 - 10:30:32 EST


> -----Original Message-----
> From: Sinan Kaya [mailto:okaya@xxxxxxxxxxxxxx]
> Sent: Friday, March 23, 2018 10:44 PM
> To: David Miller <davem@xxxxxxxxxxxxx>
> Cc: netdev@xxxxxxxxxxxxxxx; timur@xxxxxxxxxxxxxx; sulrich@xxxxxxxxxxxxxx;
> linux-arm-msm@xxxxxxxxxxxxxxx; linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; Elior,
> Ariel <Ariel.Elior@xxxxxxxxxx>; Dept-Eng Everest Linux L2 <Dept-
> EngEverestLinuxL2@xxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v5 3/5] bnx2x: Eliminate duplicate barriers on weakly-
> ordered archs
>
> On 3/23/2018 1:04 PM, David Miller wrote:
> > From: Sinan Kaya <okaya@xxxxxxxxxxxxxx>
> > Date: Fri, 23 Mar 2018 12:51:47 -0400
> >
> >> It could if txdata->tx_db was not a union. There is a data dependency
> >> between txdata->tx_db.data.prod and txdata->tx_db.raw.
> >>
> >> So, no reordering.
> >
> > I don't see it that way, the code requires that:
> >
> > txdata->tx_db.data.prod += nbd;
> >
> > is visible before the doorbell update.>
> > barrier() doesn't provide that.
> >
> > Neither does writel_relaxed(). However plain writel() does.
>
> Correct for some architectures including ARM but not correct universally.
>
> writel() just guarantees register read/writes before and after to be ordered
> when HW observes it.
>
> writel() doesn't guarantee that the memory update is visible to the HW on all
> architectures.
>
> If you need memory update visibility, that barrier() should have been a
> wmb()
>
> A correct multi-arch pattern is
>
> wmb()
> writel_relaxed()
> mmiowb()
>

Sinan, Since you have mentioned the use of mmiowb() here after writel_relaxed().
I believe this is not always correct for all types of IO mapped memory [Specially if IO memory is mapped using write combined (for ex. Ioremap_wc())].
We have a current issue on our NIC (qede) driver on x86 for which the patch is already been sent more than a week ago [Still awaiting to hear from David on that].
where mmiowb() seems to be useless since we use write combined mapped doorbell and mmiowb() just seems to be a compiler barrier() there.
So in order to flush write combined buffer we really need writel_relaxed() followed by a wmb() to synchronize writes among CPU cores.
I think the correct pattern in such cases (for write combined IO) would have been like below -

wmb();
writel_relaxed();
wmb(); -> To flush the writes actually.

Thanks.